Conference Program

IEEE 2nd International Conference on AI and Emerging Technology For Sustainable Future

Department of Mathematics and Computer Science, University of Catania
July 24-25, 2026
Program Summary
Time Zone: Central European Time - Rome (GMT+02:00)
Venue: Department of Mathematics and Computer Science, University of Catania
Friday, July 24, 2026 Day - 1
Time: 09:00 - 09:15
Opening Ceremony Location: Room - 127
Dr. Alaa Ali Hameed
Istinye University, Turkey
Associate Professor
Dr. Orazio Muscato
University of Catania, Italy
Professor
Time: 09:15 - 09:45
Keynote Session Location: Room - 127
Dr. Laura R. M. Scrimali
University of Catnia
Time: 09:45 - 10:15
Coffee Break Location: SALA CONSIGLIO (1st Floor)
Time: 10:15 - 12:30
Technical Session - 1
Location: Room - 126
Session Chairs: Alessia Rondinella
Technical Session - 2
Location: Room - 127
Session Chairs: Lorenzo Catania
Time: 12:30 - 14:00
Lunch Break Location: SALA CONSIGLIO (1st Floor)
Time: 14:00 - 15:30
Technical Session - 3
Location: Room - 126
Session Chairs: Dr. Georgia Fargetta
Technical Session - 4
Location: Room - 127
Session Chairs: Francesco Guarnera
Time: 15:30 - 17:00
Panel Discussion Location: Room - 127
Women Panel - TechDrink
Moderator:
• Dr. Georgia Fargetta, University of Catania
Panel Members:
• Dr. Ivana Guarneri, STMicroelectronics
• Dr. Silvia Cariolo, CEO & Co-Founder of aiVAGO, Chief Product and Technology Officer (CPTO) at 4Gift
Time: 17:00 - 17:30
Coffee Break Location: SALA CONSIGLIO (1st Floor)
Time: 17:30 - 18:30
Technical Session 4-A
Location: Room - 126
Session Chairs: Massimo Orazio Spata
Technical Session 4-B
Location: Room - 127
Session Chairs: Dr. Alessandro Ortis
Time: 18:30 - 19:00
Closing Remarks Location: Room - 127
Time: 20:00 - 22:00
ICAISF 2026 Gala Dinner Location: Trattoria da Peppino
Saturday, July 25, 2026 Day - 2
Time: 09:00 - 09:30
Keynote Session Location: Virtual Room -1
Dr. Luca Guarnera
University of Catania, Italy
Time: 09:30 - 11:30
Technical Session - 5
Location: Virtual Room -1
Session Chairs: Dr. Georgia Fargetta
Technical Session - 6
Location: Virtual Room -2
Session Chairs: Mirko Casu
Technical Session - 7
Location: Virtual Room - 3
Session Chairs: Jose Estupinan
Time: 11:30 - 13:45
Technical Session - 8
Location: Virtual Room -1
Session Chairs: Alessia Rondinella
Technical Session - 9
Location: Virtual Room -2
Session Chairs: Francesco spina
Technical Session - 10
Location: Virtual Room - 3
Session Chairs: Riccardo Raciti
Time: 13:45 - 16:00
Technical Session - 11
Location: Virtual Room -1
Session Chairs: Dario Samuele PISHVAI
Technical Session - 12
Location: Virtual Room -2
Session Chairs: Daniele Cocuzza
Technical Session - 13
Location: Virtual Room - 3
Session Chairs: Rosa Zuccará
Time: 16:00 - 19:25
Technical Session - 14
Location: Virtual Room -1
Session Chairs: Muhammad Atif Saeed
Technical Session - 15
Location: Virtual Room -2
Session Chairs: Lorenzo Catania
Technical Session - 16
Location: Virtual Room - 3
Session Chairs: Rosario Licciardello
Full Program
Time Zone: Central European Time - Rome (GMT+02:00)
Venue: Department of Mathematics and Computer Science, University of Catania
Friday, July 24, 2026 Day - 1
Opening Ceremony In Person Time: 09:00 - 09:15. Location: Room - 127
Dr. Alaa Ali Hameed
Dr. Alaa Ali Hameed
Istinye University, Turkey
Dr. Orazio Muscato
Dr. Orazio Muscato
University of Catania, Italy
Keynote Session In Person Time: 09:15 - 09:45. Location: Room - 127
Dr. Laura R. M. Scrimali
Dr. Laura R. M. Scrimali
University of Catnia
Talk: Bridging Equilibrium Theory and Multi-Agent Reinforcement Learning for Trade Networks under Crisis
Abstract:Global disruptions, from climate extremes to geopolitical crises, increasingly threaten agricultural trade and food security, calling for quantitative tools capable of capturing both the rigor of equilibrium theory and the adaptivity of modern AI methods. This talk presents a hybrid framework that unifies classical variational inequality (VI) models of multi-commodity trade network equilibrium with Multi-Agent Reinforcement Learning (MARL), bridging two traditionally separate domains: rigorous mathematical equilibrium theory and adaptive, data-driven learning. Supply and demand agents interact within a shared market environment, learning strategies that converge toward theoretically grounded equilibrium states even under dynamic price shifts, route disruptions, and capacity constraints. Experimental results show that the proposed approach achieves faster and more robust convergence than standard MARL baselines, highlighting the potential of combining equilibrium theory with emerging AI techniques to support resilient, sustainable decision-making in complex socio-economic networks.
Biography:

Laura R. M. Scrimali is Full Professor of Operations Research at the University of Catania, Italy. Her research focuses on variational and quasi-variational inequalities, Nash and generalized Nash equilibrium problems, evolutionary game theory, and network equilibrium models, with applications to supply chains, transportation, healthcare, and digital platforms. Her recent work explores the intersection of AI and strategic decision-making, including multi-agent reinforcement learning and evolutionary game models for AI-driven systems, as well as sustainable and closed-loop supply chain network equilibria. She has participated in several research projects and is the author of numerous publications in leading journals in the field. She serves as co-editor of several Special Issues on Variational Inequalities and Nash Equilibrium Problems, and has been a member of the Scientific Committee of the International Conference on Operations Research and Enterprise Systems (ICORES) since 2017. Earlier in her career, she was a postdoctoral fellow at the Center for Operations Research and Econometrics (CORE), Université Catholique de Louvain, Belgium.

Coffee Break In Person Time: 09:45 - 10:15. Location: SALA CONSIGLIO (1st Floor)
Technical Session - 1 In Person Time: 10:15 - 12:30. Location: Room - 126
Session Chairs: Alessia Rondinella
10:15
10:30
Paper ID: 232
Process-Aware Counterfactuals for Emergency Department Long-Stay Prediction
Lerina Aversano; Felice Franchini; Debora Montano; Chiara Verdone
Presentation: In Person
Abstract: In emergency departments, some visits quietly become bottlenecks. When a patient remains longer than expected, beds, clinicians, and downstream resources stay occupied, reducing the system’s ability to serve incoming patients. Early identification of visits at risk of becoming prolonged can support timely intervention, but prediction alone is insufficient; staff also need explanations that translate risk into feasible, clinically meaningful actions. This paper presents PROCAX (PROcess-aware Constrained counterfActual eXplanations), a lightweight, process-aware framework for early long-stay prediction and constrained counterfactual explanation in emergency departments. Applied to 425,022 ED visits from MIMIC-IV-ED, the method extracts features from only the first 60 minutes of each stay and trains a LightGBM classifier. The combined clinical-plus-process model achieves macro-F1 0.607 and ROC-AUC 0.715, improving macro-F1 by 0.031 over a clinical-only baseline. On the explanation side, generic counterfactual methods flip every prediction (validity 1.000) but rarely produce usable guidance: only 46.3% of their suggested changes involve actionable features and just 13.1% respect temporal validity. PROCAX addresses this by enforcing four constraint families (immutability, clinical plausibility, temporal/process validity, and data-manifold bounds), guaranteeing that every returned explanation is actionable, temporally coherent, and clinically plausible by construction. Validity reaches 0.396 at standard search budget and 0.498 at expanded budget. The results show that sustainable AI deployment in healthcare demands more than accurate predictions. Explanations must respect the logic of the clinical process to be trustworthy and actionable in real-world settings.
10:30
10:45
Paper ID: 115
Forecast‑Driven Reinforcement Learning for Multi‑Asset Trading
MARCO PALOMINO; Federica Monachello; Luciana Dalla Valle
Presentation: In Person
Abstract: Integrating time‑series forecasting into deep reinforcement learning (DRL) offers a promising avenue for improving multi‑asset trading strategies. In this study, we evaluate the combination of PatchTST---a Transformer‑based forecasting model---with Proximal Policy Optimisation (PPO)---a widely used actor‑critic algorithm designed to stabilise policy updates through a clipped objective. Our aim is to determine whether predictive signals generated by a forecasting model can meaningfully enhance an agent's decision‑making in a dynamic financial environment and improve the quality of the portfolio strategies it learns. We test this framework on daily trading decisions across eight U.S. assets from 2000 to 2024, incorporating transaction costs and minimum holding constraints to reflect realistic market conditions. The agent equipped with PatchTST predictions achieves substantially higher cumulative returns and Sharpe ratios, albeit with increased turnover and fewer average long positions. Our analysis identifies when predictive information strengthens trading performance and when it may introduce instability, offering practical insights into the value and limits of combining forecasting models with DRL‑based decision systems.
10:45
11:00
Paper ID: 136
The effects of prompting and fine-tuning in medical text simplification: Evaluation of various models
Gulsum Yigit; Fatma Gümüş; Melike Nur Yegin
Presentation: In Person
Abstract: Medical text simplification is important for enhancing patients' understanding of complex healthcare information. This study evaluates the effectiveness of Qwen2-7B, Llama3.1-8B, and Mistral-7B models across two medical benchmark datasets. We compare multiple strategies, including zero-shot, few-shot, and fine-tuning. We also investigate the integration of rule-based guidelines into the prompting process. We evaluate the models using six standard metrics: BLEU, ROUGE, BERTScore, SARI, FKGL, and Compression Ratio. The results show that fine-tuning generally improves reference-overlap metrics. Qwen2-7B achieved the highest ROUGE-L score, increasing from 0.2993 in its best prompting setting to 0.4738 after fine-tuning. Llama3.1-8B achieved the highest BLEU score of 0.2622, and Mistral-7B achieved the highest SARI score of 42.41 after fine-tuning. However, prompting remained competitive for semantic similarity. For Mistral-7B, rule-guided zero-shot prompting achieved a BERTScore of 0.8951, slightly higher than its fine-tuned score of 0.8937. Overall, fine-tuning improves lexical similarity to reference, while carefully designed prompts remain a practical alternative. Few-shot prompting was sensitive to demonstration selection and led to a different interpretation in each model.
11:00
11:15
Paper ID: 138
Modelling Open Science Behaviour Using Synthetic Data with a Monte Carlo Approach
JUAN-JOSE BOTÉ-VERICAD
Presentation: In Person
Abstract: Open Science has gained increasing relevance in recent years, promoting transparency, accessibility, and reproducibility in research. However, the gap between researchers’ attitudes towards open science and their actual practices remains insufficiently explored. This study analyses a dataset based on a survey of researchers from Spanish institutions, focusing on data sharing, reuse, and attitudes towards open science principles. The study operationalises data deposit behaviour as an ordinal variable capturing non-adoption, future intention, and current adoption. Additionally, a composite index is constructed to measure overall open science attitude based on transparency, perceived improvements in peer review, and conflict perception. To validate the robustness of the dataset, a synthetic dataset is generated using a bootstrap-based Monte Carlo approach. The results reveal a balanced distribution across non-adoption, transitional intention, and full adoption, with a significant proportion of researchers in a transitional stage. While attitudes towards open science are generally positive, actual data reuse remains limited, indicating a gap between perception and practice. The comparison between real and synthetic datasets shows minimal differences, confirming that the synthetic data preserve the statistical properties of the original dataset. These findings demonstrate the potential of synthetic data as a reliable tool for modelling research behaviour and provide insights into the current state of open science adoption.
11:15
11:30
Paper ID: 141
A Scalable Distributed Multi-Modal Framework for Smart-City Traffic Analysis and Violation Detection in Heterogeneous Environments
Arun Natarajan; Kamble Aditya Dattatray; Hafeez Muhammed; Munavar Fairooz Cheranchery; Shihabudheen K. V.
Presentation: In Person
Abstract: Traditional Intelligent Transportation Systems (ITS) in developing regions often encounter critical latency bottlenecks when processing heterogeneous traffic, severe occlusions, and nonstandard vehicle geometries that challenge smart-city mobility. To overcome these limitations, we propose a scalable, distributed edge-cloud framework that uses asynchronous, microservices-based AI edge node orchestration yielding higher throughput and lower median latency than a monolithic baseline. We deploy a multi-modal engine with custom-optimised YOLO architectures for parallel attribute extraction, including vehicle detection, colour profiling, logo identification, and two-wheeler violation detection (helmet noncompliance and triple riding). We also implement a Multi-View Consensus (MVCS) Optical Character Recognition (OCR) pipeline that achieves a 14 percent absolute increase in word accuracy over industry-standard engines under low-contrast conditions. Scalability stress testing, mathematically modelled using the Universal Scalability Law, indicates a theoretical saturation limit of ~184 concurrent edge nodes under our test conditions. Finally, to establish new benchmarks for heterogeneous traffic analytics in smart‑city deployments, we have open-sourced three annotated datasets: a Motorcycle Safety Violation dataset, a Heterogeneous Traffic ingress dataset, and a Hybrid Synthetic-Real Logo dataset.
11:30
11:45
Paper ID: 144
Explainable Machine Learning for Predicting Dental Treatment Complexity in Patients with Special Needs and Dentophobia
Anita Petreska; Julijana Nikolovska; Efka Zabokova Bilbilova; Blagoj Ristevski; Vlatko Kokolanski; Nikola Rendevski
Presentation: In Person
Abstract: Dental treatment in patients with special needs and severe dentophobia represents a significant clinical challenge and often requires a complex therapeutic approach, including treatment under general anesthesia. Early identification of factors associated with treatment complexity is important for improved treatment planning, optimization of clinical resources, and provision of individualized care. This study proposes an explainable machine-learning framework to predict dental treatment complexity using clinical, behavioral, and sociodemographic data from 303 patients treated at a specialized university dental institution. Treatment complexity was modeled as a binary outcome variable indicating patients who required a higher number of dental procedures. Logistic Regression, Random Forest, and XGBoost were evaluated using stratified 5-fold cross-validation. Performance was assessed using ROC-AUC, PR-AUC, and Brier score. To improve transparency and interpretability, Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP) were applied. Logistic Regression was selected as the final model, achieving a ROC-AUC of 0.8828 and a PR-AUC of 0.9014 while maintaining high interpretability. The most influential predictors were primary tooth extractions, permanent tooth extractions, restorative interventions, age, and neurodevelopmental disorders. Fairness analysis showed relatively stable performance across sex and financial groups, with greater variation across geographic regions. The results suggest that explainable machine learning can provide transparent support for early identification of complex dental cases in vulnerable populations.
11:45
12:00
Paper ID: 146
Citizen Attitudes toward AI-Enabled Public Services: Six Urban Samples and Implications for the GZM Metropolis
Łukasz Walusiak; Bartosz Ignac; Marcin Budziński; Joanna Kalisz; Yurii Vitkovskyi; Agnieszka Szostak; Marek Gachowski; Ewa Palarczyk
Presentation: In Person
Abstract: Artificial intelligence (AI) may improve the quality and efficiency of public services, but its socially acceptable implementation is associated with citizen trust, perceived risks, and expectations of human oversight. This study reports a crosssectional survey of 3,528 respondents from six urban samples: the GZM Metropolis, Istanbul, Montreal, New York, Beijing, and Prague. A ten-item questionnaire measured AI use, confidence, expected future need, attitudes, trust, risk concerns, development priorities, perceived barriers, and conditional support for AI- enabled public services. Overall, 71.7% of respondents supported such services under a scenario in which a human could take over, ranging from 58.7% in Prague and 59.2% in GZM to 92.0% in Beijing. AI-use frequency and confidence were positively associated with perceived future need, with Spearman correlations of 0.507 and 0.616 and adjusted odds ratios of 1.29 and 2.66. Positive attitudes and trust were likewise associated with publicservice support, with correlations of 0.591 and 0.605 and adjusted odds ratios of 1.84 and 1.81. Privacy and data concerns were the leading barrier at 20.1%. Health care and education were the most frequently selected development priorities, at 28.6% and 25.4%, respectively. Significant differences across survey sites support transparent, human-supervised, and locally adapted implementation.
12:00
12:15
Paper ID: 147
Threshold-Optimized Explainable Deep Learning for Pediatric Craniofacial Fracture Triage in Computed Tomography
Bartosz Ignac; Łukasz Walusiak; Natalia Sitek-Ignac; Katarzyna Tyburska; Tomasz Wach; Krzysztof Dowgierd; Marcin Kozakiewicz; Zygmunt Wróbel; Bogusława Orzechowska-Wylęgała
Presentation: In Person
Abstract: Pediatric craniofacial computed tomography is difficult to interpret in emergency settings because developing anatomy and subtle fracture patterns can obscure clinically important findings. This paper presents a compact explainable decision-support pipeline for study-level fracture triage. A retrospective cohort of 209 anonymized pediatric craniofacial computed tomography examinations was processed with DICOM-to-NIfTI conversion, isotropic resampling, bone-window normalization, and axial, coronal, and sagittal multiplanar reconstructions. The three views were combined into a 2.5D input for an ImageNet-initialized ResNet-18 classifier trained with weighted binary cross-entropy and transfer learning. On a stratified validation set of 63 examinations, the model achieved an area under the receiver operating characteristic curve of 0.934 and an area under the precision-recall curve of 0.938. Youden-index threshold selection identified a decision threshold of 0.46, yielding sensitivity of 0.931, specificity of 0.882, precision of 0.871, F1-score of 0.900, and only two false-negative examinations. Gradient-weighted class activation mapping produced anatomically plausible activation patterns in representative positive cases and helped characterize false-negative behavior. The results support further external validation of explainable artificial intelligence as a second-reader aid for pediatric craniofacial trauma imaging.
12:15
12:30
Paper ID: 163
VaccinIAmoci: A Domain-Constrained LLM Chatbot for Personalized Vaccination Guidance in Italy
Andrea Giuseppe Maugeri; Martina Barchitta; Marco Enea; Domenica Matranga; Antonella Agodi
Presentation: In Person
Abstract: This paper presents VaccinIAmoci, a web-based conversational AI system designed to support Italian citizens in navigating the national vaccination schedule through personalized guidance grounded in the Italian National Immunization Plan (PNPV) 2023–2025 and complementary Italian institutional guidance for pregnancy, special populations, chronic conditions, and travel-related vaccination. The system employs a domain-constrained prompting strategy, termed Knowledge-Embedded System Prompt with Domain Guardrails (KESPDG), combining role anchoring, institutional grounding, and rejection-based boundary control. The final KESPDG configuration was evaluated on a held-out benchmark of 180 queries spanning six use-case categories and independently assessed by two domain-expert reviewers. Under reviewer-consensus scoring, the final KESPDG configuration achieved an overall accuracy of 96.7% (95% CI: 92.9%–98.6%) and 100% correct rejection on the off-topic subset. A generic-prompt baseline evaluated on the same benchmark achieved 73.3% overall accuracy and 0% correct off-topic rejection. The contrast was largest in domain-boundary control, while a secondary gain was observed in the institutional specificity of in-scope responses. Residual errors were concentrated in semantically adjacent pregnancy and travel-health queries, suggesting that bounded health-information systems may require a three-way interaction policy consisting of answer, safe redirection, and hard reject. These findings support the practical value of domain-constrained prompting in narrow public-health information settings, while underscoring the need for privacy compliance, governance, and real-world validation prior to deployment.
Technical Session - 2 In Person Time: 10:15 - 12:30. Location: Room - 127
Session Chairs: Lorenzo Catania
10:15
10:30
Paper ID: 359
AI powered mobile system for assessment of ocular motility based on cardinal gaze positions
Gang Luo
Abstract: Assessment of ocular motility is an important eye exam in the context of strabismus. Photography of eyes in nine cardinal gaze positions, with the head in straight ahead position, is a widely used method of documenting ocular motility. Such photos are usually not suitable for automated quantitative assessment because reliable anatomical landmarks for gaze estimation are difficult to identify. We propose a novel mobile AI-based system that reverses the conventional approach: instead of having patients keep the head still, they fixate on the camera while rotating their head (or equivalently, the camera orbits around the patient fixating on the camera). This photography method produces corneal light reflection as a landmark, enabling robust Hirschberg-based alignment measurement for all gaze positions. Our mobile system integrates three key components: (1) a head pose estimation model for standardizing cardinal positions; (2) a single-eye detection model for eye region cropping in large-angle head turn images, (3) AI eye feature extraction and deviation calculation pipeline using Hirschberg method. We compared the single-eye detection model on 306 images, achieving 99.9% detection rate as compared to 46.2% with Yolov8-face model. The head pose estimation model achieved a mean absolute error of 2.2° across yaw and pitch rotations. According to a superior oblique palsy case report, we demonstrated the system's capability to quantitatively measure hypertropia that only manifested at certain gaze positions. The proposed system offers an accessible solution for quantitative ocular motility assessment with potential applications in clinical screening and tele-ophthalmology.
10:30
10:45
Paper ID: 280
ChainShield: AI-Driven Enumeration for Supply Chain Cyber Risk Governance
Awynash Sewnandan; Mounir El Jerrari; Yuri Bobbert
Presentation: In Person
Abstract: This paper examines how an AI-supported enumeration capability can improve third-party cyber risk governance and continuous compliance in complex EU-regulated supply chains. The study addresses a structural gap in current third-party risk management: organisations are increasingly required to maintain continuous visibility over ICT dependencies, yet dominant practices remain episodic, self-reported, and weak at revealing dependency risk. The research combines a systematic literature review with a Group Support System (GSS)-supported structured expert panel informed by Delphi principles. Fourteen cybersecurity and governance experts assessed data sources, technical indicators, implementation barriers, and adoption conditions for enumerating external Tier-1 suppliers using publicly accessible, licensed, or contractually disclosed data. The resulting artefact, ChainShield, is proposed as a data-feed-oriented minimum viable product that continuously collects, resolves, and correlates external technical, contextual, and compliance signals about suppliers. The findings indicate conditional feasibility at the Minimum Viable Product (MVP) level. Cyber-threat intelligence feeds, version control metadata, software bills of materials combined with vulnerability scoring, and cloud service health dashboards emerged as priority signals. A three-layer Data Source Taxonomy and the Technical, Regulatory, Organisational, and Economic (TROE) framework structure the contribution. The paper concludes that AI-driven enumeration can strengthen supply chain resilience and compliance evidence when used as explainable decision support rather than as an autonomous scoring substitute for human governance
10:45
11:00
Paper ID: 168
CNN-Based Safety Procedure Recognition for Electrical Infrastructure Maintenance: A Preliminary Framework
Ottaviano Emma; Cecilia Manduca; Francesco Castelli; Nicole Dalia Cilia
Presentation: In Person
Abstract: Electrical infrastructure maintenance activities involve complex and potentially hazardous operational procedures, where procedural errors may compromise both operator safety and system reliability. Within the Design for Safety (DfS) paradigm, this paper presents a preliminary convolutional neural network (CNN)-based framework for the visual recognition of safety procedures during electrical infrastructure maintenance, shifting the focus from reactive hazard detection toward the proactive validation of correctly executed procedures.

The framework distinguishes between two representative procedures, namely Utility Pole Climbing (UPC) and Earthing and Short-Circuiting (ESC), using a balanced dataset of 400 grayscale images acquired in controlled operational environments. Several CNN architectures were evaluated, and a lightweight model was selected and validated through stratified k-fold cross-validation.

The proposed model achieved an average classification accuracy above 97%, demonstrating stable performance across folds. A robustness analysis based on data augmentation techniques simulating realistic field conditions confirmed the model’s ability to generalize under non-ideal acquisition scenarios.

The results demonstrate the feasibility of lightweight CNN-based systems as visual decision-support tools for safety-oriented procedure recognition and provide a methodological foundation for future integration with wearable devices, augmented reality systems, and real-time operator-assistance platforms.
11:00
11:15
Paper ID: 169
BioAnnotator: Ensemble-Based Named-Entity Recognition and Linking to Biomedical Knowledge Graphs for Semantic Interoperability
Sergio Consoli; Vıctor Suarez-Paniagua; Fabiola Curion; Nicholas Spadaro; Lorenzo Bertolini; Mario Ceresa
Presentation: In Person
Abstract: Semantic interoperability is a cornerstone of digital transformation, enabling systems to exchange and interpret data across organizational and sectoral boundaries. In the biomedical domain, the automatic recognition and disambiguation of named entities from unstructured text remains a critical challenge for achieving interoperability across heterogeneous health information systems. We present the BioAnnotator, an interactive web-based tool combining an ensemble of Named- Entity Recognition (NER) models — including transformer-based (XLM-RoBERTa, WikiNEuRaL, Medical-NER) and span-based architectures (GLiNER, GLiNER2) — with a Named- Entity Linking (NEL) pipeline to standard biomedical knowledge graphs. A fusion strategy resolves overlapping annotations across models, while configurable filters allow users to select entity categories, confidence thresholds, and target knowledge graphs. Validated in WHO Disease Outbreak News reports, BioAnnotator demonstrates effectiveness in extracting and linking domain relevant entities, contributing to semantic interoperability in digital health ecosystems.
11:15
11:30
Paper ID: 170
PFM-Agent: Orchestrating a Recurrent Reinforcement-Learning Policy with Behavioral-Economics Critics via LangGraph for Adaptive Personal Finance Management
Albaraa Alruwaymi; Wojdan BinSaeedan
Presentation: In Person
Abstract: Personal financial wellness is shaped less by raw arithmetic than by behavioural biases: loss aversion, mental accounting, and status-quo bias systematically push individuals away from welfare-maximising allocations. Existing roboadvisory and reinforcement-learning (RL) approaches model the optimisation side well, but treat user behaviour as exogenous noise rather than as a structured failure mode that an agent can be designed to detect and counteract. We introduce PFM-Agent, a research scaffold and design that combines three components rarely seen together: (i) LangGraph state-graph orchestration with typed shared state and bounded cyclic edges; (ii) a recurrent RL policy (RecurrentPPO) that allocates a user’s budget across five canonical buckets while preserving long-range temporal structure of cash flows; and (iii) three behavioural-economics critic agents that each evaluate proposed allocations through a specific bias lens (loss aversion, mental accounting, status-quobias) and whose natural-language verdicts are aggregated into a consensus score that both gates revisions at inference time and shapes the reward at training time. The system operates over the public IBM TabFormer synthetic credit-card transaction dataset (24M records, 2k users). We motivate the design from four converging strands of recent literature, formalise the problem as a partially-observable MDP with five simplex actions, present the full graph as exported from the compiled LangGraph, and report preliminary scaffold-validation measurements from real environment and graph runs. We pre-register three hypotheses (H1: recurrent vs. feed-forward; H2: critic-shaped reward; H3: interpretability) for the empirical study that follows.
11:30
11:45
Paper ID: 173
Competitive features of Redox Flow Batteries in powering AI Data Center
Antonio Zingales; Nicola Poli; Giacomo Marini; Massimo Guarnieri
Presentation: In Person
Abstract: The huge growth of artificial intelligence (AI) in data centers is challenging electric power supply and transmission system in the integration of large inverter loads into already constrained networks. While traditional data center demand is driven by traffic geographically dispersed, statistically diverse, and often relatively smooth, AI training workloads, differently, synchronize thousands of GPUs through collective communication: GPU stop ang go often move together, therefore the resulting workload is not random, it is forced to power “swings” which are really challenging the powering grid: power systems designed for normal data center behavior are strained by AI demand. The electrical storage needed for smoothing and interfacing AI data centers goes far beyond conventional backup applications as their load fluctuations occur across multiple timescales—from milliseconds to several days. The paper explores the potential role in this application of Redox Flow Batteries capable of absorbing synchronized training surges, often occurring over multiple cycles per day and requiring substantial energy depth. In this context, the characteristics of flow batteries—particularly their long-duration capability, intrinsic non-flammability, and fast response—align closely with the operational profile of AI-driven data centers. More broadly, flow batteries can play a pivotal role, alongside other technologies such as lithium-ion batteries and iron-air batteries, in supporting a resilient and low-carbon energy.
11:45
12:00
Paper ID: 195
An Efficient Machine Learning Approach for Degradation Forecasting in AEM Water Electrolysis
Marco Veneriano; Ani Gjergji; Sebastiano Bellani; Andrea Riva; Vito Paolo Pastore; Matteo Santacesaria
Presentation: In Person
Abstract: This study provides a data-driven analysis of a novel dataset of single-cell Anion Exchange Membrane water electrolyzers (AEMWE), operated under constant current load across multiple heterogeneous experimental campaigns. We train and evaluate a range of machine learning models with different complexity, including linear baselines, LSTMs and CNNs, to perform medium-term forecasting of the cell voltage degradation curve. The models are assessed within a rigorous training and evaluation framework specifically designed for heterogeneous industrial data.
12:00
12:15
Paper ID: 196
Silent Lip Reading and Visual Speech Recognition: A Survey on Modern Architectures, English Benchmarks, and Italian Datasets
Rosario Licciardello; Giada Lopo; Georgia Fargetta; Alessandro Ortis; Sebastiano Battiato
Presentation: In Person
Abstract: Silent Lip Reading, or Visual Speech Recognition (VSR), is the process of decoding spoken language exclusively from visual articulatory movements. This survey reviews the evolution of state-of-the-art (SOTA) processing pipelines, tracing the shift from early controlled datasets and CNN-RNN architectures to modern “in the wild” benchmarks and Conformer models coupled with Large Language Model (LLM) decoders. Furthermore, we provide a dedicated analysis of emerging resources for the Italian language, surveying existing resources for low-resource VSR deployments. Finally, we discuss modern semantic evaluation metrics, robustness under acoustic degradation, and open research challenges, highlighting the shift towards zero-shot cross-lingual generalization.
12:15
12:30
Paper ID: 225
Roadmap for Automation Integration for the AI-Driven Sustainable Service Design
Steve Wang; Jing Shi; Elmer Ragus; Tina Camba
Presentation: In Person
Abstract: While traditional service design focuses on customer journeys and user experience, emerging AI, digital twins, and autonomous systems demand a shift toward intelligent operational execution and lifecycle management. This paper proposes a comprehensive roadmap for AI-driven, sustainable service design by systematically aligning service intent with operational efficiency, resilience, and organizational transformation. We introduce a framework that integrates core engineering lifecycle metrics—availability, reliability, maintainability, supportability, and efficiency—alongside the Accountability–Predictability–Balanceability–Policy (APBA) continuous improvement cycle for governance. Employing a multi-AI consulting methodology (encompassing ChatGPT, Claude, Gemini, Grok, and DeepSeek) to evaluate technology readiness, the roadmap guides organizations through progressive stages from transparency to balance capability. The findings demonstrate that service design is evolving into an integrated socio-technical discipline. This study provides researchers and practitioners with a structured, scalable approach to implement advanced automation while ensuring robust human oversight and long-term ecosystem sustainability.
Lunch Break In Person Time: 12:30 - 14:00. Location: SALA CONSIGLIO (1st Floor)
Technical Session - 3 In Person Time: 14:00 - 15:30. Location: Room - 126
Session Chairs: Dr. Georgia Fargetta
14:00
14:15
Paper ID: 247
AI for Media Education and Digital Citizenship: A Triadic Framework for Sustainable and Responsible Educational Transformation
Daisy De Gioannini
Presentation: In Person
Abstract: This paper contributes to Artificial Intelligence for Education (AI4Ed) by examining how generative AI, instructed conversational agents, and collaborative digital platforms can support media education and digital citizenship in secondary schooling. Drawing on the eTwinning project Safe & Smart Online: Skills4Life (Italy, Spain, Hungary, Serbia, Jordan; learners aged 16–18; LifeComp-anchored), we propose the Triadic Loop of AI-Mediated Digital Citizenship: a three-stage cycle of Surfacing, Reframing, and Producing through which AI tools restructure the relation between critical reception and responsible civic action. We complement the framework with a four-role taxonomy—AI-as-Mirror, AI-as-Megaphone, AI-as-Mediator, AI-as-Mentor—and a critical account of the limits of AI mediation, including hallucination, privacy risks, and the fragility of AI-supported civic learning. The case is treated as paradigmatic; the contribution is conceptual and directly relevant to responsible AI-driven educational transformation in sustainable, future-oriented learning ecosystems.
14:15
14:30
Paper ID: 256
Hallucination as Wrong-Basis Fabrication in LLMs: Contextuality, BMS, and Multi-Agent Implications
Diego Gosmar
Presentation: In Person
Abstract: The standard framing of hallucination in large language models treats fabrication as a retrieval failure, a calibration error, or stochastic noise — all assuming a context-independent right answer exists for every query. We argue this assumption fails for a systematic subclass of fabrications, and propose an operational alternative grounded in Contextuality-by-Default (CbD), a purely statistical theory requiring no Hilbert-space structure. We define a wrong-basis fabrication as a coherent projection of an internal representation onto a readout context that is incompatible with the one the user intended, make the construct measurable via a Basis Misalignment Score (BMS) and a pilot proxy (BMS-lite), and state a falsifiable structural claim: wrong-basis fabrications are internally consistent because they are coherent projections, not noise. A pilot experiment across three capability levels confirms: existence-conditioned prompting reduces fabrication to 0% on all models (from baselines of 67–93%), BMS-lite rates of 60–93% are consistent with cross-context incompatibility, and fabricated details cluster at 1.76–1.94 times above the Monte Carlo chance baseline. We discuss consequences for inter-agent protocols and identify the grounding-control confound as the primary open issue for full-scale replication.
14:30
14:45
Paper ID: 261
Clinical Pathway Optimization via Causal Event Graphs
Lerina Aversano; Felice Franchini; Debora Montano; Chiara Verdone
Presentation: In Person
Abstract: Hospitals worldwide face growing pressure to deliver high-quality care while controlling costs and using limited resources efficiently. Accurately predicting how long a patient's clinical pathway will last is central to this goal: it supports better bed and staff planning, reduces operational waste, and enables a more sustainable use of healthcare resources. In practice, however, predictive models for this task are often hard to trust, because they rely on ambiguous causal claims, are validated without respecting the temporal order of events, or leak future information into training. This paper proposes CEG (Causal Event-Graph-informed ensemble), which combines multi-scale activity sequence representations, directed event-graph features encoding observed pathway transitions, and clinical and temporal descriptors within a transparent weighted ensemble of tree and linear regressors. The event graph captures the order in which clinical activities occur and supplies interpretable pathway-level features; here, causal refers to this ordered-event structure, not to treatment-effect estimation. CEG is evaluated on the BPIC 2011 Dutch academic hospital log (1,143 patient episodes, 150,291 events) using a strict temporal protocol that prevents data leakage. Under rolling-origin cross-validation, CEG predicts pathway duration with a mean absolute error of 0.823 ± 0.089 days, outperforming the strongest baseline (ExtraTrees: 0.896 ± 0.103 days), and reaches R² = 0.945 on a final-period hold-out. The model is interpretable, computationally light, and paired with a calibration procedure that turns each prediction into a reliable time interval for deployment, making CEG a practical and sustainable tool for clinical pathway optimisation and resource planning.
14:45
15:00
Paper ID: 262
Proposed Blockchain Framework for Local Carbon Market
Malick Ndiaye
Presentation: In Person
Abstract: Carbon markets operate at local and global levels, each with specific roles and legal frameworks. Local markets are based on community-led emission-reduction projects, but global markets support large-scale carbon credit trading and diversified green projects. However, both markets suffer from issues related to transparency , efficiency and trust. Integrating blockchain technology into the carbon market proposes a viable route to net-zero emissions due to its capability to address current market issues by increasing transparency, reducing cost, and facilitating broader involvement. Further studies are vital to realize the full potential of blockchain in elevating the carbon offset mechanism. This work develops a blockchain framework specifically for local carbon markets, utilizing decentralized verification, smart contracts, and IoT-integrated real-time emissions tracking and reporting. Our design ensures accountability and improves efficiency by increasing transparency and automation to meet local compliance requirements. The framework supports community-driven green projects and local carbon markets to achieve socio-economic benefits, deeper community engagement, and localized environmental goals. On the other hand, global carbon markets, while essential for massive emission reductions, mostly have difficulties with local relevance, standardized protocols, and stakeholder inclusion. Understanding these distinctions between local and international carbon markets is critical for designing efficient offsetting systems at both levels. Keywords: Carbon offsetting market, transparency, cross boarded, community engagement, blockchain, trust.
15:00
15:15
Paper ID: 268
Offline and Online BSREM Reconstruction with Support Refinement for Limited-Trajectory Gamma Imaging
Jakub Gawron; Dominik Rzepka; Katarzyna Heryan; Michael Friebe
Presentation: In Person
Abstract: This work presents a simulation-based evaluation of gamma source reconstruction for a compact linear multi-channel detector intended for limited-trajectory acquisition in the ULTRACLEAR hybrid SPECT/ultrasound imaging concept. Synthetic measurements were generated from compact radioactive source distributions and reconstructed using iterative statistical methods. Several iterative reconstruction strategies were compared, including a baseline maximum-likelihood method and a regularized method that used prior information about the expected source distribution. An additional refinement step was evaluated, in which an initial reconstruction was used to identify the most probable active region, and a second reconstruction was performed on this reduced search space. A sequential reconstruction variant was also tested, where the activity map was updated as new measurement batches became available. The results show that regularized reconstruction improved the average agreement with the true activity distribution compared to the baseline method. Offline refinement further improved the reconstructed maps and the estimated total activity scale. In the sequential setting, refinement produced more compact activity maps, although it did not increase the average correlation metric. These results indicate that the proposed reconstruction pipeline can support gamma source localization in limited-trajectory acquisition and may provide an algorithmic basis for future real-time reconstruction in compact hybrid SPECT/ultrasound systems such as ULTRACLEAR.
15:15
15:30
Paper ID: 269
Quantifying the Performance Gap in Hand and Hand Keypoint Detection Between Regularly and Irregularly Shaped Hands
Danilo Valdes Ramirez; Ana Belén Gil-González; Gema Chamoso Rodríguez; Juan Manuel Corchado
Presentation: In Person
Abstract: Hand and hand keypoint detection is a fundamental technology for neurorehabilitation and human-computer interaction, yet current popular models often lack robustness when applied to populations with atypical hand anatomy, such as those with cerebral palsy or stroke-induced deformities. This study presents a performance evaluation of six widely deployed algorithms -You Only Look Once (YOLO) YOLOv8, YOLOv11, YOLOv12, YOLOv26, Real-Time Detection Transformer (RT-DETR), and MediaPipe- to assess their visual resilience across regularly and irregularly shaped hands. While the YOLO-family models and RT-DETR were fine-tuned on a unified hand-keypoint dataset, the template-dependent MediaPipe framework was evaluated in its standard configuration. Our empirical results demonstrate a significant performance gap between regular and irregular hand structures. In bounding-box hand detection, the YOLO family and RT-DETR remained relatively robust, maintaining a mean Average Precision (mAP50) above 98.2% on deformed hands. In contrast, MediaPipe’s precision collapsed from 99.19% on regular hands to approximately 64.47% on irregular ones. However, a critical bottleneck was identified in keypoint estimation, where all models experienced sharp declines. For instance, the pose mAP50 for YOLO11 and YOLO12 plummeted from nearly 99% to approximately 67.73% and 69.51%, respectively, when encountering irregular anatomy. The findings suggest that current architectures suffer from a normative bias, attempting to make standard global templates onto non-standard hand topologies. This level of performance degradation is deemed unacceptable for clinical precision in medical tracking.
Technical Session - 4 In Person Time: 14:00 - 15:30. Location: Room - 127
Session Chairs: Francesco Guarnera
14:00
14:15
Paper ID: 172
Bridging the AI Translation Gap in Pharmaceutical Innovation: The Generative AI Innovation Translation (GAIT) Model
Victoria Freund
Abstract: Generative Artificial Intelligence (GenAI) is increasingly regarded as a transformative technology in pharmaceutical research and development. Despite significant advances in AI-driven drug discovery and process optimization, many pharmaceutical organizations struggle to translate technological AI capabilities into sustainable innovation outcomes. Existing research primarily focuses either on technological potential or on organizational adoption barriers, while the processes through which AI capabilities are transformed into realized innovation remain insufficiently understood. This paper introduces the Generative AI Innovation Translation (GAIT) Model, a conceptual framework addressing this gap by integrating perspectives from AI capability research, organizational adoption literature, and innovation management theory. The framework proposes that successful AI-enabled transformation depends not only on technological readiness or barrier reduction, but also on organizational capabilities that support experimentation, learning, adaptability, and cross-functional knowledge integration. Building on the concept of an AI Translation Gap, the model shifts the focus from AI adoption toward AI innovation translation and provides a process-oriented perspective on how pharmaceutical organizations convert AI opportunities into realized innovation outcomes. In doing so, the paper contributes to a more holistic understanding of AI-enabled pharmaceutical transformation and offers a conceptual foundation for future empirical research.
14:15
14:30
Paper ID: 281
Vision Guided Robotic Sorting of Epoxy Composite Materials Using Deep Learning for Circular Manufacturing
Syed Ali Hassan; Michail Beliatis
Presentation: In Person
Abstract: The increasing use of composite materials in industries such as wind turbine manufacturing has resulted in significant material waste during production processes. Epoxy based composite residues are often discarded despite their potential for reuse or recycling. This work proposes a vision guided robotic system for the automated identification and sorting of epoxy containing objects to support circular manufacturing practices. The system integrates a Real Sense camera D435, a CNN based object detection model, and a Dobot M1 Pro scara type robotic manipulator equipped with a vacuum gripper. A custom dataset of grid matrix epoxy glass fiber patterns was created and labeled followed by data augmentation to improve model generalization. The CNN model was trained for 100 epochs and achieved a mean average precision mAP@50 of 99.5% with a recall of 100%. During experiment, the camera captures images of objects within the workspace, and the trained model detects epoxy glass fiber regions in real time. Based on the detected bounding box location, the robot performs a pick and place operation to transfer epoxy glass fiber containing object to a designated location while ignoring object without epoxy. The Real-time experiment shows that the proposed system can successfully perform autonomous detection and sorting tasks with slight variations in object placement. The proposed approach highlights the potential of combining computer vision and robotic manipulation to support scrap sorting for sustainable manufacturing and circular economy initiatives in the composite material industries.
14:30
14:45
Paper ID: 283
BLIND: Bottom-Layer INtrinsic Document fingerprinting with a Siamese Network
Daniele Cocuzza; Francesco Guarnera; Sebastiano Battiato
Presentation: In Person
Abstract: Ensuring the authenticity of printed material is a fundamental requirement in numerous contexts, such as valuable documents, banknotes, tickets, or rare collectible cards, where reliable and easily accessible solutions are needed to effectively prevent and counteract counterfeiting. This paper introduces a novel technique for extracting distinctive fingerprints from physical paper documents by analyzing their intrinsic microstructural patterns. The proposed approach captures non-reproducible texture variations inherent to the paper substrate representing them through discriminative features learned by a siamese network. This representation effectively encodes the unique material signature of each document, enabling robust intrinsic fingerprinting across different acquisition conditions. We evaluate our approach through document retrieval experiments conducted on multiple datasets, comparing its performance with state-of-the-art methods. The results confirm the validity of the adopted strategy and suggest practical applications in the forensic domain. Code and dataset will be available at publication time.
14:45
15:00
Paper ID: 290
APTIM OS: An Enterprise Learning Framework for AI Literacy and Workflow Transformation
Seohoo Yi
Presentation: In Person
Abstract: Organizations are adopting generative artificial intelligence (AI) rapidly, yet tool access alone rarely produces durable change in everyday work. Many enterprise AI programs stop at one-time tool instruction or prompt workshops, leaving individual learning disconnected from workplace task application, internal service creation, and organizational routine formation. This paper proposes and evaluates APTIM OS, an AI-driven enterprise learning framework that links AI literacy to workflow transformation through five dimensions—Assessment, Process, Training, Integration, and Value Realization—and an integrated workflow spanning learner diagnosis, role-based curricula, AI-tutor support, assignment-based tacit knowledge discovery, change-agent development, and internal AI service creation. This paper reports an approximately 18-month field application with 77 participants in a mid-sized construction-manufacturing group in South Korea, a site-oriented context typically under-served in AI literacy. Using a mixed-method, multi-level evaluation, the study finds significant pre-to-post gains across four learning constructs and an increase in the average number of AI tools used competently from 0.49 to 3.35; participants produced 78 workplace AI use cases, 63 internal AI services entered team testing, and a selected engineering design workflow showed a 75.4 percent task-time reduction. APTIM OS offers field evidence for moving enterprises beyond one-time AI training toward workflow-level value realization and more sustainable workforce capability.
15:00
15:15
Paper ID: 319
Turbidity-Aware VLC for Sustainable Underwater Environmental IoT Monitoring
Antonio Costanzo; Laura Galmard; Charles Cattarello; Nicolas Michel; Alexandra Carriere
Presentation: In Person
Abstract: Monitoring water quality in rivers, lakes, ports, and aquaculture sites is an urgent sustainability challenge directly linked to UN Sustainable Development Goals 6 (Clean Water) and 14 (Life Below Water). Underwater visible light commu nication (VLC) offers a low-cost, electromagnetic interference free sensing and communication medium, but link performance degrades rapidly under water turbidity. In this paper, we reframe turbidity from a pure impairment into actionable channel state information (CSI) that drives an AI-inspired adaptive decision loop. A multi-wavelength RGB LED architecture is governed by a Sensing–Decision–Adaptation (SDA) loop that measures per channel bit error rate (BER) via pilot signals, classifies the tur bidity regime, and selects the optimal (colour, modulation) pair at each cycle. We validate two decision engines: a measurement driven look-up table (LUT) built from offline calibration, and a k-Nearest Neighbours (kNN) classifier trained on an independent noisy dataset. Experiments confirm a Blue→Green→Red degra dation hierarchy; both adaptive strategies extend link operability well beyond a fixed-colour baseline (+12g/L sediment, +247%), while the kNN achieves 92.2% agreement with the oracle-optimal policy (5-fold CV accuracy: 0.953, k = 10), demonstrating that calibration-free ML adaptation is feasible on a sub-$50 Arduino/photodiode prototype.
15:15
15:30
Paper ID: 323
AI-Driven SAR Flood Mapping with Probabilistic Temporal Modeling for Sustainable Environmental Monitoring
Antonio Costanzo; Guerin Thomas; Adrien Re; Touaaveau Taata; Nicolas Michel; Alexandra Carriere; Pierre Borella
Presentation: In Person
Abstract: Accurate and timely flood extent mapping is crit ical for climate-resilient emergency response and adaptive risk management,two pillars of sustainable development in the face of intensifying hydrometeorological extremes. This paper presents a comparative evaluation of two flood detection pipelines applied to Sentinel-1 SAR imagery: a deterministic adaptive-thresholding baseline and a U-Net deep convolutional neural network. Both pipelines are benchmarked on nine Sentinel-1 acquisitions span ning the complete flood–recession cycle of the 2022 Pakistan monsoon disaster (July25– October9, 2022). The U-Net achieves a mean Jaccard index of 0.750 versus 0.462 for thresholding (+62.3%) and a mean F1-score of 0.850 versus 0.608 (+39.8%). As a novel contribution, U-Net spatial segmentation is cou pled with Gaussian Process (GP) regression using a Mat´ ern5/2 kernel to interpolate flooded-surface dynamics between satellite passes and to provide calibrated temporal uncertainty bounds. The GP reduces the 95% predictive confidence interval from ±34.4%(two observations) to ±5.4% (nine observations),a 51% reduction that directly quantifies the marginal value of each new satellite acquisition. The Mat´ ern5/2 kernel outperforms the standard RBF kernel by 37% in mean absolute error, demonstrating superior handling of asymmetric flood–recession profiles. These results establish a viable, IoT-free AI pipeline for adaptive updating of flood-risk reference maps, directly aligned with the sustainable-future mission of ICAISF.
Panel Discussion In Person Time: 15:30 - 17:00. Location: Room - 127
Title: Women Panel - TechDrink
Moderator
Dr. Georgia Fargetta
Dr. Georgia Fargetta
University of Catania
Panel Members
Dr. Ivana Guarneri
Dr. Ivana Guarneri
STMicroelectronics
Dr. Silvia Cariolo
Dr. Silvia Cariolo
CEO & Co-Founder of aiVAGO, Chief Product and Technology Officer (CPTO) at 4Gift
Coffee Break In Person Time: 17:00 - 17:30. Location: SALA CONSIGLIO (1st Floor)
Technical Session 4-A In Person Time: 17:30 - 18:30. Location: Room - 126
Session Chairs: Massimo Orazio Spata
17:30
17:45
Paper ID: 258
MedAgent: A Multi-Agent Large Language Model Framework for Explainable Clinical Decision Support in Emergency Triage
Maria Bratu
Abstract: Emergency triage demands rapid, high-stakes decisions in conditions where patient data is often incomplete and clinician cognitive load is at its peak. While AI-assisted triage tools have made measurable progress, the dominant paradigm remains a single model trained end-to-end to map raw clinical notes to an acuity label — an approach that conflates several distinct reasoning steps and yields outputs that are difficult to audit or explain. In this paper, we propose MedAgent, a multiagent large language model (LLM) framework that decomposes the triage decision into four specialized, communicating agents: a Symptom Extraction Agent, a Differential Diagnosis Agent, an Evidence Retrieval Agent, and an Explainability Agent, all coordinated by a lightweight safety-aware Coordinator module. Each agent is independently fine-tuned from BioMistral-7B on a task-specific clinical corpus. The framework addresses two persistent limitations of existing systems: the tendency of monolithic models to entangle distinct reasoning subtasks, and the absence of structured, auditable explanations required by the EU AI Act for high-risk medical AI. We describe the architecture in detail, discuss the design rationale for each component, and outline a planned empirical evaluation on the publicly available MIMICIV- ED dataset. A comparative analysis against representative existing approaches highlights the architectural advantages of the proposed design with respect to evidence grounding, explanation fidelity, and regulatory compliance.
17:45
18:00
Paper ID: 272
A Sustainable Approach to Ex Situ Seed Conservation through Citizen Participation and Distributed Storage
Andrea Vitaletti; Khalil Massri
Abstract: More than 95\% of the crop genetic erosion articles analyzed by Khoury et al reported changes in diversity, with nearly 80% providing evidence of loss. The lack of diversity presents a severe risk to the sustainability of global food systems. Without seed diversity, it is difficult for plants to adapt to pests, diseases, and changing climate conditions. Genebanks, such as the Svalbard Global Seed Vault, are valuable initiatives to preserve seed diversity in a single secure and safe place. However, according to our analysis of the data available in the Seed Portal, the redundancy for some species might be limited, posing a potential threat to their future availability. Interestingly, the conditions to properly store seeds in genebanks, are the ones available in the freezers of our homes. This paper lays out a vision for Collective Seed Storage relying on a peer-to-peer infrastructure of domestic freezers to increase the overall availability of seeds. We present a Proof-of-Concept focused on monitoring the proper seed storage conditions and incentivizing user participation through a Blockchain lottery. The PoC proves the feasibility of the proposed approach and demonstrates that our vision can contribute to several of the 17 United Nations Sustainable Development Goals.
18:00
18:15
Paper ID: 80
STURIA: An AI-Assisted Conversational and Cross-Platform Time-Use Tracking System for Students
Mustapha Rachdi; Alain Fernex; Lei Luo; Idir Ouassou; Lionel Filippi; Axel Mabrouk
Abstract: Understanding student time allocation is a fundamental challenge in educational research, particularly in the context of self-regulated learning and academic performance prediction. However, traditional measurement approaches—such as retrospective questionnaires and static time-use diaries—suffer from recall bias, low temporal resolution, and limited capacity to capture multitasking and fine-grained behavioral dynamics. This paper introduces STURIA (Student Time Use and Regulation with Intelligent Assistance), an AI-enhanced, conversational, and cross-platform system designed to capture student time-use data in real time. STURIA combines continuous activity tracking, natural language interaction, and voice-based input to reduce user burden while improving data accuracy. Built with Flutter, the system ensures deployment across Android, iOS, web, and desktop environments. The proposed architecture integrates a temporal event-processing engine and a machine learning-based conversational layer for activity recognition and structuring. STURIA aims to provide a scalable framework for high-resolution time-use analysis and personalized academic feedback generation.
Technical Session 4-B In Person Time: 17:30 - 18:30. Location: Room - 127
Session Chairs: Dr. Alessandro Ortis
17:30
17:45
Paper ID: 333
From Policy Ambivalence to Consensus: An AI-Assisted Framework for Participatory Governance
Miguel Puig Cabrera; Resurrección Rodríguez-Gil; Ginesa Martínez-del Vas
Abstract: Participatory governance has become an essential mechanism for improving the legitimacy, transparency and inclusiveness of public decision-making. However, many participatory processes are characterized by the coexistence of multiple legitimate yet competing stakeholder preferences, generating situations of policy ambivalence that hinder consensus-building and delay collective action. While recent advances in Artificial Intelligence (AI) have expanded opportunities for supporting policy analysis and stakeholder engagement, their potential contribution to consensus-building remains underexplored. This paper proposes an AI-assisted framework aimed at facilitating the transition from policy ambivalence to consensus within participatory governance processes. The framework positions AI as a deliberative support mechanism capable of synthesizing stakeholder perspectives, identifying areas of convergence, mapping policy conflicts and evaluating alternative policy pathways. Rather than replacing human judgement, AI is conceived as a tool for enhancing collective deliberation and supporting more transparent and evidence-informed discussions. An illustrative application is presented to demonstrate how AI can contribute to consensus-building while preserving human oversight and democratic legitimacy. The proposed framework contributes to emerging debates on AI-enabled governance and offers a conceptual foundation for integrating AI into participatory decision-making processes characterised by competing policy preferences.
17:45
18:00
Paper ID: 334
Lightweight Person Detection Implementations on Edge Platforms: Benchmarking for Latency and Power Consumption
Ousama Noureddine; amin Haj-Ali; Orazio Aiello; Ali Ibrahim; Roberto La rosa
Abstract: Person detection on edge devices has shown its importance in many application domains such as surveillance,

robotics, and smart cities. These applications necessitate performing the inference locally, constraining time latency, and power

consumption. This paper benchmarks three lightweight YOLO- based models, namely Tiny-YOLOv2, YOLOv5n, and YOLO11n,

for person detection on three sound edge platforms: Jetson Nano/TX1, Raspberry Pi 5, and STM32N6570-DK. The study considers both time latency and power consumption, highlighting the power-delay product (PDP). Measured results demonstrate that although optimized Jetson deployments achieve the lowest latency, the STM32N6570-DK provides the best overall PDP because of its low power consumption feature and competitive INT8 inference latency. In particular, the lowest PDP value is achieved for the YOLOv5n model on STM32N6570-DK. This optimized implementation on the STM32N6570-DK outperforms the Jetson Nano/TX1 and the Raspberry Pi 5 by providing a reduction of 49.72% and 63.78% in PDP, respectively.
Closing Remarks In Person Time: 18:30 - 19:00. Location: Room - 127
ICAISF 2026 Gala Dinner In Person Time: 20:00 - 22:00. Location: Trattoria da Peppino
Saturday, July 25, 2026 Day - 2
Keynote Session Online Time: 09:00 - 09:30. Location: Virtual Room -1
Join Session
Dr. Luca Guarnera
Dr. Luca Guarnera
University of Catania, Italy
Talk: Generative AI and Multimedia Forensics: Opportunities, Risks, and Detection
Abstract:Recent advances in Generative Artificial Intelligence have dramatically transformed the way digital content is created, enabling the generation of highly realistic images, videos, speech, and text. While these technologies are opening unprecedented opportunities across numerous application domains, they also introduce significant challenges related to misinformation, identity fraud, and digital trust. This talk provides an overview of modern generative models, including GANs and diffusion models, highlighting their capabilities and real-world applications. The presentation then discusses the main categories of synthetic media and deepfakes, the associated security and societal risks, and the current state of multimedia forensic techniques for detecting AI-generated content. Finally, recent research challenges and future directions toward robust and generalizable forensic solutions are presented.
Biography:

Luca Guarnera is a Research Fellow in Computer Science at the University of Catania, where he has worked since January 2022. He received his Ph.D. in Computer Science from the University of Catania in 2021 with a thesis entitled "Discovering Fingerprints for Deepfake Detection and Multimedia-Enhanced Forensic Investigations." During his doctoral studies, he also conducted research at the University of Hertfordshire, UK, under the supervision of Prof. Salvatore Livatino, focusing on virtual-reality-based tools for forensic ballistics analysis and firearm comparison. He earned his M.Sc. in Computer Science with honours from the University of Catania in 2017 and has been a member of the Image Processing Laboratory (IPLab) since 2015. He participated in the Mohamed Bin Zayed International Robotics Challenge in 2017 and 2019 and attended several international summer schools in computer vision, medical imaging, and signal processing. His main research interests include computer vision, machine learning, multimedia forensics, and deepfake detection.

Technical Session - 5 Online Time: 09:30 - 11:30. Location: Virtual Room -1
Join Session
Session Chairs: Dr. Georgia Fargetta
09:30
09:45
Paper ID: 167
Hybrid Dense and Sparse Passage Retrieval for Efficient Log Analysis in Distributed Systems
Lubna Mohammed Hasan Hadi; Talha Karadeniz
Presentation: Online
Abstract: Operational log data generated by modern distributed systems has grown enormously, and is essential for fault diagnosis, anomaly investigation, and system health monitoring. However, the structure of these logs is quite distinct: they frequently exhibit repetitive template patterns, opaque system identifiers, extreme event imbalance, and a vocabulary that differs substantially from natural language, making conventional retrieval approaches inadequate. While Best Matching 25 (BM25) performs well on exact template matching, it fails on semantic queries; and dense encoders trained on web corpora yield virtually no improvement on semantic queries without targeted domain adaptation. This paper presents a hybrid dense–sparse passage retrieval architecture that addresses both shortcomings by fine-tuning a BAAI General Embedding (BGE) encoder on the specific domain, applying adaptive score fusion per query type, and incorporating a Retrieval-Augmented Generation (RAG) layer for structured analytical responses. The fine-tuned hybrid system (BM25 + BGE-FT) is evaluated on 12,620 semantically coherent retrieval windows extracted from two million Hadoop Distributed File System (HDFS) log lines. On a held-out test set of 201 queries, it achieves a statistically significant improvement of +13.7% (p < 0.001, paired t-test) over the BM25 baseline, with Mean Average Precision at rank 100 (MAP@100) = 0.3359. The five-stage RAG evaluation with Qwen2.5-7B-Instruct yields ROUGE-L = 0.183, semantic faithfulness = 0.878, and section completeness = 0.931. Results are further confirmed by cross-dataset validation on the BlueGene/L supercomputer corpus (+16.4% MAP gain) and multi-seed analysis (mean MAP = 0.3357 ± 0.0020 over 5 seeds), demonstrating statistical stability. These results show that domain fine-tuning and adaptive fusion are necessary for high-quality log retrieval, and support the use of hybrid retrieval systems in operational log analysis pipelines.
09:45
10:00
Paper ID: 125
UNIVERSAL WAVELET UNITS IN 3D RETINAL LAYER SEGMENTATION
An Le; Hung Nguyen; Mohamed Morsy; Amr Ali; Shane Griffin; Shadi Alashwal; Nehal Mehta; Melanie Tran; Jesse Most; Dirk-Uwe Bartsch; William Freeman; Cheolhong An; Truong Nguyen; Shyamanga Borooah
Presentation: Online
Abstract: This paper presents the first application of tunable wavelet units (UwUs) to 3D retinal layer segmentation from Optical Coherence Tomography (OCT) volumes. To mitigate information loss introduced by conventional max-pooling, we replace standard downsampling in a motion-corrected MGU-Net with three wavelet-based modules, OrthLatt-UwU, BiorthLattUwU, and LS-Biorth-UwU, built on learnable wavelet filter banks that preserve both low-frequency contextual information and high-frequency structural details. Evaluated on the Jacobs Retina Center (JRC) and PPS OCT datasets, the proposed framework consistently improves Dice scores over the max-pooling baseline, demonstrating that tunable wavelet-based downsampling provides a principled and effective alternative for volumetric medical image segmentation.
10:00
10:15
Paper ID: 129
A COMPREHENSIVE REVIEW OF MACHINE LEARNING AND NLP TECHNIQUES FOR ONLINE RECRUITMENT FRAUD DETECTION
Duha Hmeyem
Presentation: Online
Abstract: Online recruitment fraud that targets job seekers with fake job listings posted on online platforms has become a significant cybersecurity threat. Traditional fraud detection approaches are based on rule-based frameworks and manual analysis, which have become less effective in handling large-scale and continuously evolving data. Therefore, significant attention has been paid to machine learning and deep learning approaches for fraud detection. Despite this, most of the related studies have been conducted using static datasets and with offline learning models that cannot effectively process continuously generated streaming data. This paper provides an extensive review of relevant machine learning, deep learning and streaming-based techniques for online recruitment fraud detection. The review examines the latest methods for improving detection performance in dynamic environments, state-of-the-art datasets, preprocessing techniques, class-imbalance handling strategies and concept drift adaptation mechanisms. Further, this review outlines the significant limitations of existing models and indicates future research directions for developing more adaptive and real-time fraud detection systems.
10:15
10:30
Paper ID: 130
Development of a Digital Twin for Citrus Postharvest Processing Optimization Using MATLAB/Simulink
Cristina Martínez Ruedas; María José De la Haba; Francisco Jiménez-Jiménez; Carolina Santos; Isabel Castillejo-González
Presentation: Online
Abstract: The citrus sector is undergoing a progressive digital transformation aimed at improving process efficiency, product quality, and operational sustainability. In this context, digital twins have emerged as promising technology for simulating industrial processes and evaluating operational scenarios without affecting real production systems. This work presents the development of a digital twin for a fresh orange postharvest processing plant located in Palma del Río (Córdoba, Spain), operated by Sunaran S.A.T. The proposed model was developed in MATLAB/Simulink and reproduces the main stages of the processing line, including reception, selection, washing, ultraviolet inspection, waxing, grading, classification, packaging, and cold storage. The model integrates production parameters, process flows, quality-related variables, and operational constraints obtained from the industrial plant. In addition, the digital twin incorporates data associated with traceability and process monitoring systems currently used in the facility. The developed environment allows the simulation of different operating conditions and the evaluation of their impact on production efficiency, processing capacity, product losses, and process management. The results demonstrate the potential of digital twins as decision-support tools in the agro-industrial sector, enabling virtual experimentation and process optimization without interfering with real plant operation. This approach contributes to improving productivity, reducing waste, and advancing the digitalization of citrus postharvest industries within the framework of Agriculture 4.0.
10:30
10:45
Paper ID: 131
FSM-LoRA: Dynamic Gradient Gating via Finite State Machines for Energy-Efficient LoRA Fine-Tuning
Md Kaif Afran Khan; Mahtab Bin Kashem; Tulon Saha; Ahmed Faizul Haque Dhrubo; Souvik Pramanik; Mohammad Ashrafuzzaman Khan; Mohammad Abdul Qayum; Mohsin Sajjad
Presentation: Online
Abstract: Fine-tuning large language models on specific tasks remains computationally demanding, even with parameterefficient methods like Low-Rank Adaptation (LoRA). A major contributor to this cost is the mandatory backward pass at every training step, regardless of whether the model is actively learning. We introduce FSM-LoRA, a Energy-Efficient Finite State Machine-based controller that dynamically decides when to skip redundant backward passes during fine-tuning. By monitoring the smoothed training loss through a compact set of interpretable rules, FSM-LoRA intelligently gates gradient updates—executing them only when they meaningfully contribute to optimization. Evaluated across three open sourced models (Phi-2, Qwen2.5-3B, and OpenLLaMA-3B) and three diverse classification tasks, FSM-LoRA safely skipped 15% to 58% of backward passes while maintaining accuracy within ±0.5% of standard LoRA.This translates to wall-clock time reductions of up to 58.58% per task and corresponding energy savings , without compromising final performance. Notably, on more challenging tasks the controller automatically refrained from skipping, demonstrating adaptability. Our results show that lossaware, state-driven scheduling of gradient computations offers a practical path toward faster, more sustainable and greener finetuning of large language models.
10:45
11:00
Paper ID: 176
A Distributed IoT Framework for Real-Time Navigation Assistance via Monocular Depth and YOLOv8 Fusion
Sanjeekan Ravikumar; Dilushanth Sivapalan; Kopisankar Nagarajah; Mukunthan Tharmakulasingam; Logeeshan Velmanickam; Chathura Wanigasekara
Presentation: Online
Abstract: Independent indoor navigation remains a formidable challenge for over 2.2 billion visually impaired individuals worldwide. While traditional Electronic Travel Aids (ETAs) rely on proximity-based ultrasonic sensors, they lack the semantic awareness required to identify specific obstacles and navigate complex environments. This paper proposes a high- performance, low-cost IoT-based navigation assistance system that bridges the gap between restricted edge hardware and advanced Deep Learning. Our architecture utilizes a distributed framework where an ESP32-CAM acts as a lightweight vision node, streaming real-time environmental data via a localized HTTP/Wi-Fi protocol to a centralized inference engine. The system integrates a unified vision pipeline featuring YOLOv8 for robust object recognition and the MiDaS (v2.1 Small) model for monocular depth estimation. By applying a custom metric calibration constant (K), the system transforms relative disparity maps into real-world distance measurements with a Mean Absolute Error (MAE) of only 0.0576 m within a functional range of 1.5 meters. The unified vision pipeline achieves a mean end-to-end execution latency of 55.6 ± 4.2 ms per frame, enabling a stable inference rate of 15–20 FPS and a mean Average Precision (mAP@0.5) of 0.905, ensuring reliable identification of critical indoor hazards such as stairs, chairs, and persons. The proposed system provides a scalable, cost- effective alternative to expensive RGB-D or Lidar-based assistive technologies, significantly enhancing the spatial awareness and safety of visually impaired users.
11:00
11:15
Paper ID: 177
Self-Supervised Pretraining Determines Vision Transformer Localization in Chest Radiographs: A Systematic CAM Benchmark
Abraham Bautista; Evangeline John Francis Kennedy; Ioannis Kakadiaris
Presentation: Online
Abstract: Clinically useful chest X-ray AI should provide accurate predictions and spatial evidence that radiologists can inspect. We benchmark weakly supervised disease localization on the NIH ChestX-ray14 dataset by training five CNN and ViT model configurations and evaluating five CAM-based attribution methods against 984 radiologist-annotated bounding boxes across eight diseases. Three key findings stand out. 1) ViT-DINO Attention Rollout achieves 0.275 pointing game accuracy, compared with 0.032 for ViT-ImageNet Rollout, showing an 8.6-fold gain from pretraining objective alone. 2) EfficientNet-B4 GradCAM achieves the best localization despite the lowest AUROC among WeightedBCE models, demonstrating that classification accuracy does not predict localization quality. 3) SymmetricCE preserves localization relative to WeightedBCE, indicating that noise-robust training can remain compatible with CAM-based interpretability. These findings provide practical guidance for selecting auditable chest X-ray models for AI-assisted diagnosis and predictive analytics.
11:15
11:30
Paper ID: 178
Evaluating Freight Academy as an Artificial Intelligence-Enabled Work Simulation Environment in Logistics Education
Elmer Ragus; Tina Camba; Steve Wang
Presentation: Online
Abstract: Online professional education must provide flexible access while also helping students practice workplace-relevant communication, judgment, and problem solving. This exploratory mixed-methods pilot study evaluates Freight Academy as an artificial intelligence-enabled work simulation environment in online logistics education. Grounded in an online Active Learning Engagement Classroom (ALEC-O) framing, the study examines student engagement, professional relevance, and perceived career readiness. Quantitative data were collected using an adapted Assessing Student Perspective of Engagement in Class Tool (ASPECT) survey across two administrations. Qualitative reflections were analyzed to explain student experiences with the training modules. Results showed moderately positive but stable perceptions over time. Students reported stronger agreement with items related to effort, focus, and contribution, while lower ratings appeared for enjoyment, preference, and stimulated interest. Qualitative findings suggest that students valued the practical and career-relevant nature of Freight Academy. But technical friction, rigid response recognition, pacing issues, and usability challenges sometimes reduced authenticity and engagement. Findings suggest that AI-enabled work simulation can extend active learning online. However, its value depends on instructional design, reliable implementation, flexible feedback, and human support.
Technical Session - 6 Online Time: 09:30 - 11:30. Location: Virtual Room -2
Join Session
Session Chairs: Mirko Casu
09:30
09:45
Paper ID: 187
Design and Optimization of a Bidirectional V2G Power Converter
Avishka Koswaththa; Imesh Jayasinghe; Jalini Sivarajah; Thiruvaran Tharmarajah; Kopisankar Nagarajah; Logeeshan Velmanickam; Chathura Wanigasekara
Presentation: Online
Abstract: This paper presents the design and optimization of a high-efficiency bidirectional Vehicle-to-Grid (V2G) power converter aimed at addressing energy challenges in Sri Lanka, including increasing electricity demand, grid instability, and dependence on fossil fuels. The proposed system enables both Grid-to Vehicle (G2V) and V2G operations using a combined AC/DC and DC/DC converter architecture designed for a 230 V ACgrid and a 72 V electric vehicle (EV) battery. A Model Predictive Control (MPC) strategy is implemented to improve power flow control, reduce harmonic distortion, and enhance system efficiency. Simulation results obtained from MATLAB/Simulink demonstrate stable bidirectional operation with improved power quality. A hardware prototype based on an STM32 microcontroller is developed to validate the system. The proposed design shows strong potential for efficient energy management and smartgrid integration.
09:45
10:00
Paper ID: 188
Phasor Measurement Unit Development for Renewable Energy Integrated Smart Grid
Chamiduka Bandara; Kanishka Wijesooriya; Jalini Sivarajah; Mukunthan Tharmakulasingam; Logeeshan Velmanickam; Chathura Wanigasekara
Presentation: Online
Abstract: The integration of intermittent renewable energy sources (RES) introduces critical stability challenges to contemporary power grids, including voltage fluctuations and power quality degradation. While Phasor Measurement Units (PMUs) are essential for real-time monitoring, conventional devices often suffer from reduced accuracy under real-world operating conditions due to high-frequency noise and harmonic distortions. To address this challenge, this study presents an intelligent, low cost edge PMU that integrates a sliding-window Discrete Fourier Transform (DFT) engine for phasor extraction with a discrete Kalman filter for noise suppression. Additionally, an embedded Artificial Neural Network (ANN) enables automated, localized fault detection directly at the node level. The proposed system is validated through MATLAB/Simulink co-simulations and further verified using an ESP32 microcontroller prototype. The multi stage architecture achieves an overall classification accuracy of 95.15%, demonstrating an optimal balance between low deployment cost, high signal fidelity, and real-time operational performance.
10:00
10:15
Paper ID: 190
From Plausible to Faithful: VLM-Distilled Concept Bottleneck Models
Ivan Rivero; Cristina Tirnauca; Rafael Duque Medina
Presentation: Online
Abstract: Concept Bottleneck Models route image classification through human-readable concepts. Recent label-free variants score concepts via vision-language similarity (e.g., CLIP), promising scalable interpretability without per-image concept annotation. We show that, in a fine-grained classification setting with verifiable morphological priors, this scoring is unfaithful: nominal concepts do not preferentially fire on the classes they should define, yet the model still reaches competitive accuracy. We propose a three-stage recipe that closes this faithfulness gap: (i) annotate concepts once at training time with a strong vision-language model, such as Gemini; (ii) distill these annotations into a lightweight head over a domain-specific foundation model (BioCLIP); (iii) classify with an interpretable head over the distilled concept vector. Using FungiTastic-FewShot as a case study (12 classes, 1,485 images, 5-fold cross-validation grouped by observation), the recipe reaches 0.66 balanced accuracy, within 0.05 of a non-interpretable upper bound (0.71), with explanations that cite genus-defining traits and a measured deployable footprint of 82 MB INT8 that runs without vision-language model access at inference.
10:15
10:30
Paper ID: 191
Gloss-Free Turkish SLT via Visual Query Adaptation and LoRA
Salih Eren Yüzbaşıoğlu; Mustafa Anakök; Ahmet Arda Çelik; Hacer Yalim Keles
Presentation: Online
Abstract: We propose a gloss-free continuous sign language translation (SLT) system for Turkish Sign Language (T˙ID) and evaluate it on the E-TSL benchmark. The model translates raw sign videos directly into Turkish text without using gloss annotations. Each video is represented by 657-dimensional MediaPipebased pose, hand, and facial landmark features, together with their first- and second-order temporal derivatives. The sequence is first processed by a Temporal CNN to reduce its length and then encoded by a Transformer encoder. A Visual Query Adapter (VQA) further maps the encoded visual sequence into a set of fixed-size cross-attention tokens, which are given to a LoRA- adapted mBART-50 decoder for Turkish text generation. During inference, we apply Minimum Bayes Risk (MBR) decoding and select the candidate with the highest consensus score among beam outputs, using sentence-level chrF as the utility function. On the E-TSL test set, the proposed system achieves 4.69 BLEU- 4 and 37.97 chrF, improving over the previous GNN-T baseline from 3.49 to 4.69 BLEU-4, corresponding to a 34% relative gain. Results on two independent random splits show a similar trend, indicating that the gain is stable across different data partitions.
10:30
10:45
Paper ID: 192
Epileptic Seizure Prediction from Multi-Channel Fused EEG 2D Representations
Michail Marinis; Eleni Vrochidou; George Papakostas
Presentation: Online
Abstract: Early prediction of epileptic seizures can help patients take preventive measures and reduce their risk of injury during everyday activities. Prediction systems based on scalp electroencephalography (EEG) have shown great results, however, various barriers, such as the high computational power of processing a high number of EEG signals, prevent these systems from being deployed on portable devices. Based on findings from prior research identifying a group of EEG channels most strongly correlated to epileptic seizure activity, this work investigates these channels, alongside a second group selected for their reported prediction performance, both individually and in multi-channel combinations, using the CHB-MIT dataset. EEG epochs are converted to two-dimensional image representations, enabling the use of Convolutional Neural Networks (CNNs). Among the eight candidate representations, the Short-Time Fourier Transform (STFT) spectrograms showed the best overall performance. Multi-channel combinations were formed through wavelet-based image fusion, and several CNN architectures were then evaluated under 10-fold cross-validation. Experimental results demonstrate that a three-channel wavelet-fused STFT configuration with EfficientNet-V2-S achieves an AUC of 95.9%, a sensitivity of 91.8% and a false positive rate of 0.107, suggesting that a minimal three-channel configuration is a viable foundation for portable epileptic seizure prediction systems.
10:45
11:00
Paper ID: 194
A Best-Worst Method Framework for Prioritizing Algorithmic Capabilities in AI-Enabled Circular Supply Chains
Claudemir Tramarico
Presentation: Online
Abstract: Deploying Artificial Intelligence within circular supply chains demands a detailed assessment of specific algorithmic functions. This paper introduces a structured decision framework based on the Best-Worst Method to weight key technical criteria for circular digital transformations. Rather than evaluating general digital readiness, this research prioritizes six precise engineering capabilities. The methodology employs a vector-based pairwise comparison to analyze expert input, resolving inconsistencies with an exceptional Consistency Ratio (CR = 0.0048). Computational results prove that cost minimization (C2) and energy optimization (C5) act as primary drivers, collectively capturing over 62% of total decision importance. This study contributes to systems engineering by providing a mathematically rigorous alternative to qualitative readiness models. Ultimately, these findings deliver an objective diagnostic tool to guide operations planners in allocating technological capital across platform-based sustainable networks.
11:00
11:15
Paper ID: 132
Revolutionizing 5G Connectivity: A Dual-Port MIMO Antenna for mm-Wave Communications
Md Muslim Uddin Shuvo; Mehedi Hasan Refat; Ahmed Faizul Haque Dhrubo; Souvik Pramanik; Mohammad Abdul Qayum; Mohammad Ashrafuzzaman Khan; Mohsin Sajjad
Presentation: Online
Abstract: This work provides a thorough examination of the design, simulation, and performance evaluation of octagonal Multiple-Input Multiple-Output (MIMO) antennas functioning in the millimeter-wave frequency range. The design process entails optimizing the octagonal MIMO antenna structure for the 40 GHz frequency band using HFSS. Our proposed antenna operates within the 34.8 GHz to 45.5 GHz frequency range, Obtaining the highest return loss (S11) of -24 dB at 40 GHz, which signifies exceptional impedance fitting. The VSWR is 1.1, and the isolation between the antenna elements is less than -40 dB, which ensures minimal mutual coupling. We used a 1.6 mm thick Rogers (Rt-5880) substrate with a dielectric constant of 2.2 and a loss tangent of 0.0009. The antenna has a peak gain of 6.4 dBi and 99% radiation efficiency. We observe and determine that the Diversity Gain (DG) and Envelope Correlation Coefficient (ECC) are below the standard threshold.
11:15
11:30
Paper ID: 100
Fault Diagnostics in Hydrogen Fuel Cell Performance Through Dynamic Weighted Random Forest (DWRF) and Feature Engineering
Zainab Al-Tamimi; Abdullahi Ibrahim
Presentation: Online
Abstract: This paper presents a novel hybrid fault diagnostic framework for hydrogen fuel cell systems, integrating a Dynamic Weighted Random Forest (DWRF) classifier with an enhanced Harris Hawk Optimization Algorithm featuring Multi-Sonar Memory and Attenuation Control (HHO-MSMAC). The proposed method addresses the challenges of feature redundancy, overlapping fault signatures, and model generalization in high-dimensional sensor environments. A detailed simulation was conducted using MATLAB/Simulink, incorporating five common fuel cell faults: membrane drying, cathode flooding, hydrogen leakage, catalyst degradation, and cooling system failure. Feature selection was optimized using HHO-MSMAC, while DWRF dynamically adjusted decision tree weights based on class-specific accuracy and entropy-based confidence. Experimental results demonstrate that the proposed method significantly outperforms conventional classifiers. It achieved an overall classification accuracy of 98.2%, an F1-score of 98.1%, and a Kappa coefficient of 97.8%. Mean Squared Error (MSE) was reduced to as low as 0.018, while the R² score improved from 0.80 to 0.95 over 20 training epochs. Performance evaluation across different population sizes revealed optimal convergence at a population of 50 agents, and sensitivity analysis identified a Control Operator Parameter (COP) of 0.7 as ideal for balancing exploration and exploitation. Comparative analysis confirmed that the proposed DWRF–HHO-MSMAC framework consistently outperformed CNN, DNN, PSO-SVM, and GA-based RF classifiers in terms of fault detection accuracy, stability, and generalization
Technical Session - 7 Online Time: 09:30 - 11:30. Location: Virtual Room - 3
Join Session
Session Chairs: Jose Estupinan
09:30
09:45
Paper ID: 198
Trustworthy Regulatory Requirements Extraction using Local LLM: A Multi-Agent vs. Single-Prompt Benchmark on the EU AI Act and GDPR
Mishaal Ahmed; Muhammad Ajmal Naz
Presentation: Online
Abstract: Regulated AI systems need requirements that link testable SHALL statements to statutory source text, yet local large language models often invent obligations or misstate modality when elicitation is under-specified. We ask whether a sequential Analyst-Engineer-QA multi-agent workflow (B5) outperforms fair baselines when every role shares one Llama~3 8B model via Ollama on EU AI Act and GDPR chunks. Using quote-validated expert gold, the structured single-prompt baseline B1 achieves the highest F1 on the EU AI Act test (0.382) and GDPR generalization (0.337), ahead of B5 (0.320, 0.322) and unstructured B0 (0.269, 0.275). B5 exhibits higher hallucination than B1 on both splits (e.g., 0.134 vs 0.009 on EU test) despite three LLM calls per chunk versus one for B1. Ablations B0--B5 indicate that precision gains, not added agents, drive quality on consumer hardware. We answer the primary research question in the negative for matched prompt content: role-separated inference (B5) does not beat the single-call structured prompt (B1) on F1. For trustworthy, privacy-preserving regulatory requirements engineering, practitioners should prefer single-call structured prompts and deploy multi-agent pipelines only when intermediate artifacts are required for human audit.
09:45
10:00
Paper ID: 277
Domain-Adaptive Transformer Framework for Robust Alzheimer's Disease Detection Across ADNI and OASIS MRI Datasets
Sajib Debnath; Nishat Jahan; Md Tuhin Mia; Irin Akter Liza
Abstract: Alzheimer’s disease (AD) is the leading cause of dementia, and early detection is critical for timely care. Deep models for structural MRI achieve high accuracy but are usually trained and tested on a single cohort, degrading sharply under the domain shift caused by differences in scanners, protocols, and populations, an obstacle to clinical use. We propose DAT-AD, a domain-adaptive transformer that couples a Swin encoder with a feature alignment module and domain-adversarial training to learn representations that stay discriminative across cohorts. The model is optimized with a joint objective that balances supervised classification against a domain-confusion term realized through a gradient reversal layer. Trained on labeled ADNI together with unlabeled OASIS scans and evaluated on OASIS under a cross-dataset (unsupervised domain adaptation) protocol, DAT-AD reaches 87.3% accuracy, 0.861 F1, and 0.921 AUC, surpassing a 3D Swin baseline by 6.1 accuracy points and the best 3D CNN baseline by 9.5. An ablation shows that feature alignment and adversarial adaptation contribute jointly, a feature-space analysis shows that the two cohorts become intermixed after adaptation, and a gradient-based localization analysis places the model's focus over medial-temporal structures. The study prioritizes cross-cohort robustness over within-dataset accuracy alone
10:00
10:15
Paper ID: 201
A Hybrid ANFIS-Transformer Framework Tuned by En-hanced HawkFish Optimization for Voltage and Load Balancing in Smart Grids
Mohammed Abdulrazzaq; Kamil A. Khalaf; Ahmad H. Al-Hadithi; Mustafa Oudah Hani Al-saedi; Ali Alkharsan; Husam Adil Musaab
Presentation: Online
Abstract: This paper presents a novel hybrid framework that integrates an Adaptive Neuro-Fuzzy Inference System (ANFIS) with a Transformer model, optimized through an Enhanced HawkFish Optimization Algorithm (EHFOA), to enhance voltage regulation and load balancing in smart grid environments. The proposed system leverages the temporal modeling capabilities of Transformers for accurate load and voltage prediction, while ANFIS enables adaptive, rule-based control in dynamic operating conditions. EHFOA, incorporating strategies such as Lévy flight, energy-aware movement, and elite memory, is designed to fine-tune the hyperparameters of both ANFIS and the Transformer for optimal performance. Simulation results using a real-time load monitoring dataset from Kaggle show that the proposed method significantly outperforms traditional ANFIS-only and Transformer-only models. It achieved a Root Mean Square Error (RMSE) of 1.24, Mean Absolute Error (MAE) of 0.96, and energy loss reduction to 1.9%. In addition, the system demonstrated improved control response time (1.04 s) and voltage stability, with the lowest fitness convergence (0.0094) reached in only 58 iterations. The framework exhibits high robustness, scalability, and accuracy, making it a strong candidate for intelligent grid management. Future work will explore lightweight Transformer alternatives and real-time adaptation to further optimize system deployment.
10:15
10:30
Paper ID: 207
AI, Automation, and Jobs in Finance: A Bibliometric Perspective of Labor Market Impacts in Banking and FinTech
Katerina Fotova Čiković; Antonija Mandić; Tanja Jakšić
Presentation: Online
Abstract: This study presents a comprehensive bibliometric analysis of the scientific literature on artificial intelligence (AI), automation, and their implications for labor markets within banking and FinTech. Using data extracted from the Scopus and Web of Science Core Collection databases, a total of 2,314 documents published between 2020 and 2026 were analyzed. The study applies a systematic bibliometric framework supported by Bibliometrix, VOSviewer, and complementary visualization techniques to examine research productivity, collaboration patterns, intellectual structure, and thematic evolution. The analysis identifies key publication trends, influential authors, leading journals, and dominant contributing countries, with China and the United States emerging as primary research hubs. Journal and keyword co-occurrence analyses reveal that research is predominantly concentrated on artificial intelligence, machine learning, digital transformation, and financial technologies, while labor market dimensions such as job displacement, workforce skills, and employment restructuring remain comparatively underexplored. Trend topic analysis indicates a strong shift toward technology-centric themes, with limited but growing attention to socio-economic implications of automation in financial services. The findings provide evidence-based insights for policymakers, researchers, and industry stakeholders, while outlining future research directions focused on workforce adaptation, skill transformation, and the socio-economic impacts of AI in finance.
10:30
10:45
Paper ID: 108
TinyML-Based Classification of Arm Electromyographic Signals for Assistive Device Control
Jose Estupinan; Ernesto Sifuentes
Presentation: Online
Abstract: This paper presents a TinyML-based gesture classification system for assistive device control using forearm electromyographic (EMG) signals and wrist inertial measurements. Three EMG channels and six IMU axes were acquired with an Arduino Nano 33 BLE Sense, and five wrist gestures were mapped to basic mobility commands. A 72- dimensional feature vector was extracted from segmented signal windows and classified using a Random Forest model. The system was evaluated using Leave-One-Subject-Out validation across ten subjects to assess inter-subject generalization without user-specific calibration. The classifier achieved an average accuracy of 81.71% ± 6.27%, with near-perfect recognition of the stop command, which is critical for functional safety. The complete processing pipeline, including digital filtering, feature extraction, and inference, was deployed on the embedded device, requiring 387 KB of flash memory and 104 KB of RAM. These results demonstrate the feasibility of real-time EMG and IMU gesture classification on resource-constrained microcontrollers, supporting compact, wearable, and minimally calibrated TinyML-based assistive control interfaces. Keywords— Electromyography, IMU, TinyML, Random Forest, gesture classification, assistive device, embedded machine learning.
10:45
11:00
Paper ID: 218
Fuzzy Flower Pollination Algorithm Based Traffic Scheduling System for Ring Roads
Faid Aljanabi; Luma Hasan
Presentation: Online
Abstract: The purpose of this study is to simulate the fuzzy control system with four junctions with twenty-one rules in the ring road that optimized with two types of flower pollination algorithm, standard and modified. The system applies to analyze the effect of FPA on the green cycle depending on the delay, density and queue length of vehicle in the ring road. UK and Germany datasets are used with Artificial dataset, all simulated within the SUMO environment. After applying the FPA with fuzzy for the datasets above, we deduce that the Takagi-Sugeno fuzzy controller consistently achieved the best performance across all configurations, with the Modified FPA yielding the best fitness of 2.8058 on the UK dataset, 0.8384 on the Germany dataset using Cauchy distribution, and 12.2272 on the artificial dataset, demonstrating the effectiveness of the proposed system in optimizing traffic signal control.
11:00
11:15
Paper ID: 221
Endogenous Demand Elasticity in OSeMOSYS via Stepwise Fictive Technologies and Machine-Learning Elasticity Estimation
Jabrane Slimani
Presentation: Online
Abstract: OSeMOSYS, the Open Source energy MOdelling SYStem, formulates long-term energy planning as a costminimising linear programme in which all end-use demands are exogenous and must be met with equality. Under stringent greenhouse-gas (GHG) emissions constraints, this assumption forces the model to deliver every unit of demand regardless of cost, ruling out the demand-side response that real systems exhibit. We close this gap by augmenting the OSeMOSYS programme with a step-function approximation of an iso-elastic demand curve, each step encoded as a fictive technology in the data file so that linearity is preserved and the model can substitute optimally between clean-technology investment and endogenous demand reduction. The calibration point is obtained from the linear-programming dual of the unconstrained reference scenario, and the governing elasticity parameter— rather than being borrowed from the literature—is estimated from observed price–quantity data by double/debiased machine learning, yielding a debiased, uncertainty-quantified input to the welfare-theoretic construction. The augmented programme is shown to be a partial-equilibrium welfare maximisation in the Samuelson sense. On the Extended UTOPIA testbed, applied to the 2005 passenger-kilometre demand under a sweep of CO2- emissions caps, the mechanism reproduces at low marginal cost abatement levels that are infeasible in the canonical inelastic model.
11:15
11:30
Paper ID: 140
A Secure OAuth-Based Social Media Data Extractor Chatbot for Enhancing Cybersecurity Education
Nabil Mohammed Nasim Uddin; Israt Jahan Chowdhury; Rashed Ahammod Bin Azam; Asifur Rahman; Souvik Pramanik; Ahmed Faizul Haque Dhrubo; Mohammad Ashrafuzzaman Khan; Mohammad Abdul Qayum; Mohsin Sajjad
Presentation: Online
Abstract: In the rapidly changing landscape of cybersecurity education narrowing the divide between academic teachings and real-world practice is vital. This paper introduces an original educational tool - a Social Media Data Extractor Chatbot (SMDE-C) which is dedicated to the teaching of certain fundamental cybersecurity principles, including OAuth 2.0-based authentication, predicting secure session management and ethical collection of web data within the context of accessing social media information. The application implements OAuth 2.0 to present secure login techniques, applies AI in the interpretation of public social media and underscores ethical considerations with respect to web scraping. Posted on HuggingFace Spaces, this chatbot is an easy and interactive way to learn for students – practicing with safe authentication, learning how web security works, retrieving public data from LinkedIn & Facebook. Controlled experiments showed that the system significantly decreased learning time for these concepts, with average session times of approximately 2.3 seconds. The paper presents the design, implementation as handson oriented while easily accessible and real-world like, allowing to draw conclusions that can inform other cybersecurity education contents and experiences.
Technical Session - 8 Online Time: 11:30 - 13:45. Location: Virtual Room -1
Join Session
Session Chairs: Alessia Rondinella
11:30
11:45
Paper ID: 57
AI-Enhanced Digital Twin Prototype Using Physics-Informed Neural Networks for Learning and Predicting Chaotic Dynamics in Sustainability-Critical Systems
Ina.Taralova@ec-nantes.fr) Devasmito Das (Devasmito.das.2@ec-nantes.fr) Ina Taralova (ECN
Abstract: Sustainability-critical systems, such as climate, energy, and ecological processes, exhibit inherent nonlinear and chaotic behavior that limits their reliable long-term prediction. In this paper, we propose an AI-enhanced digital twin framework based on Physics-Informed Neural Networks (PINNs) that integrates physical laws with data-driven learning to model and predict chaotic dynamics. A Duffing oscillator is used as a benchmark to model complex nonlinear behavior. The proposed approach achieves accurate and stable short-term predictions while preserving physical consistency, outperforming conventional neural networks. To quantify predictability limits, Lyapunov exponent based analysis is employed, and a chaos-aware predictability horizon is introduced for assessing forecasting reliability and early-warning signals. The results demonstrate the effectiveness of the framework for monitoring and decision support in sustainability-critical systems, providing a scalable approach for analyzing nonlinear instabilities and prediction limits.
11:45
12:00
Paper ID: 229
Optimized cluster-wise deep learning framework for multi-horizon solar radiation forecasting
Andreia Aparecida da Silva; Gabriel Villarrubia Gonzalez; Stefano Stefenon; Matheus Henrique Dal Molin Ribeiro
Presentation: Online
Abstract: Accurate solar radiation forecasting is essential for the planning and efficient operation of photovoltaic systems. In this context, this paper proposes a hybrid framework for multi-step-ahead solar radiation forecasting by integrating Singular Spectrum Analysis (SSA), clustering techniques, Multi-Objective Optimization (MOO) based on the Non-Dominated Sorting Genetic Algorithm - version II (NSGA-II), multicriteria decision-making methods, and deep learning models. Solar radiation data from a meteorological station in Curitiba, Paraná, Brazil, are used to evaluate forecasting horizons of 1, 3, and 6 steps ahead for March and August 2025. The time series are decomposed using SSA and grouped into clusters with similar temporal patterns. For each cluster, the hyperparameters of Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), and Long Short-Term Memory (LSTM) models are optimized through NSGA-II and selected using the VIšekriterijumsko KOmpromisno Rangiranje (VIKOR) method. The results indicate that the hybrid model outperforms the individual models across all analyzed scenarios, demonstrating greater accuracy and robustness for photovoltaic system applications.
12:00
12:15
Paper ID: 109
Intelligent Camera-Based Patient Monitoring and Fall Detection Using Deep Learning in Hospital Rooms
Adnan Ali; Reslan Mohammed Alqaisi
Presentation: Online
Abstract: Consequently, there was a requirement for some track patient systems continuously that will help to avoid errors and respond to code blue quickly while reducing workload. Tools currently available for manual observation that now continuously expect time synchrony cannot raise alarms for sudden falls or forced removal from bed attempts as well as abnormal noise and erratic behavior in the room. As such, a sophisticated vision-based solution that is capable of continuous monitoring and detection of patient falls in hospitals should be developed. The objective of the project is to design a computer vision-based patient fall detection system that uses AI to classify nine types of health-related events. The event classifications are In_Bed, Out_Bed, Danger Zone, Call Detected, Fall, Abnormality Detected, No_Abnormality Detected, and Other Persons Detected. It does this through a camera-based sensing and deep learning algorithm. For the second experiment, we used six (6) cases of video monitoring at various sites where the patients’ behaviors and their health are monitored. Under optimal conditions, the accuracies of these models were 90% and 88%, respectively. The event recognition was efficient, obtaining 0.95 for Fall_Abnormality_Detectd, 0.99 for Other_Persons_Detected, and 0.94 for Call_Detectd. The F1-score of bed occupancy was 0.88. The results showed that the structure can help ease caregivers’ burden while being reliable for patient safety surveillance. Ultimately, the proposed approach has enormous potential for monitoring smart hospitals and smart healthcare services.
12:15
12:30
Paper ID: 235
A Daily Monte Carlo Simulation of Governed AI-Assisted Software Delivery Under Sustained Demand
Salvatore Vella; Salah Sharieh; Malek Sharieh; Alex Ferworn
Presentation: Online
Abstract: Few existing simulation models capture the full end-to-end costs of AI-assisted software delivery — defect escape, latent defect manifestation, incident response labor, technical debt accumulation, and the capacity loss that results. This gap motivates both the simulation model presented here and the process comparison it enables. AUTOBAHN is a development process designed for enterprise AI-assisted software development workflows. We evaluate AUTOBAHN against four other development processes: WATERFALL, AGILE, PROMPT, and AGENTIC. We develop a Monte Carlo simulation to estimate software delivery speed, costs, and quality. This paper makes two contributions. The primary contribution is a transparent, reproducible, open-source Monte Carlo simulation model for comparing software development processes across cost, quality, and schedule dimensions. The model incorporates defect escape mechanics, latent defect manifestation, outage probability, incident response labor, technical debt accumulation, and activity-based costing, and is released as open-source for calibration and extension at https://github.com/salvella/AI-Software-Assist. The second contribution is directional simulation evidence that governed AI-assisted delivery improves cost-quality-schedule balance under sustained demand, based on a comparison of five development processes. AUTOBAHN has the lowest cost per completed task and the lowest escaped defects and outage burden. Fast generation alone in methods such as PROMPT and AGENTIC does not guarantee higher end delivery throughput when we consider escaped defects that then create incidents, technical debt, and ultimately capacity loss. The results support the thesis that structure must scale with power.
12:30
12:45
Paper ID: 241
Multi-horizon wind speed forecasting using VMD-PSO and deep learning models
Vinícius Bortolini; Andreia Aparecida da Silva; Gabriel Villarrubia Gonzalez; Stefano Stefenon; Matheus Henrique Dal Molin Ribeiro
Presentation: Online
Abstract: This paper proposes a hybrid framework for short- and multi-step-ahead wind speed forecasting that integrates Variational Mode Decomposition (VMD), metaheuristic optimization algorithm, Particle Swarm Optimization (PSO), and deep learning models, including Extreme Learning Machine (ELM), Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM). Hourly wind speed data collected in March 2024 from a meteorological station in Clevelândia, Paraná, Brazil, were used to evaluate the proposed approach under forecasting horizons ranging from 1 to 6 steps ahead. Model performance was assessed using point-based error metrics (RMSE and MAE) and interval-based metrics (PICP and PINAW). The results indicate that the VMD+ELM+PSO configuration achieved the best trade-off between predictive accuracy and interval reliability, attaining low prediction errors and narrow, well-calibrated prediction intervals while maintaining stable coverage close to the nominal confidence level across all horizons. Furthermore, the proposed method demonstrates robust generalization for multi-step-ahead forecasting, extending beyond the single-step focus commonly adopted in related studies.
12:45
13:00
Paper ID: 143
Exploring Financial Named Entity Recognition with Deep Transformers on the FiNER-ORD Dataset
Mustafa ALAZZAWI; Sura Saad; aqeel oleiwi; Bharat Bhushan; MUSTAFA LATEEF FADHIL JUMAILI
Presentation: Online
Abstract: Named Entity Recognition (NER) in financial documents presents unique challenges due to domain-specific terminology, ambiguous entity mentions, and complex multi-word entities. Existing approaches based on token-level classification suffer from boundary errors and fail to effectively disambiguate entities with similar surface forms. This paper presents a novel span-based contrastive learning framework for financial NER, evaluated on the FiNER-ORD dataset. Our approach combines: span-based entity classification that directly models entity boundaries, supervised contrastive learning with hard negative mining for entity type disambiguation, entity-centric domain adaptation through masked language modeling, and financial-specific data augmentation with consistency regularization. Experimental results demonstrate that our method achieves 94.7% F1-score on the FiNER-ORD benchmark, representing a 4.6% improvement over previous state-of-the-art methods. Ablation studies confirm the contribution of each component, with contrastive learning providing 2.6% and span-based classification contributing 3.3% improvements. Our approach shows particularly strong performance on organization entities (+5.5% F1), which are notoriously difficult in financial text. We release our code and pre-trained models to facilitate future research in financial NLP.
13:00
13:15
Paper ID: 249
Sustainable advanced digital skills education framework
Minna Isomursu; Alessia Golfetti; Katharina Lange; Mayte Toscano; Anna Bon; Hans Akkemans; Anastasia Vlachou; Dimistrios Tsolis; Tanja Ninkovic; Umberto Morelli; Rucha Sawlekar; Lamba Sakshi
Presentation: Online
Abstract: Substantial public investment has been directed toward the education sector to future-proof educational offerings with advanced digital skills, such as artificial intelligence. However, ensuring the long-term sustainability and impact of these initiatives remains a significant challenge. Many projects struggle to maintain relevance, institutionalise outcomes, or scale beyond initial pilot phases. This paper proposes a sustainability framework for advanced digital skills education, developed through a collaborative analysis of nine European projects. Drawing on joint workshops and a structured cross-project assessment, we propose a classification of sustainability activities that enable lasting impact. The resulting framework explicates the mechanisms through which training initiatives can remain relevant, adaptable, and equitable over time. By providing a structured approach to embedding sustainability into digital skills development, the framework offers practical guidance to policymakers, educators, and project leaders seeking to build resilient and future-oriented digital education ecosystems in Europe.
13:15
13:30
Paper ID: 250
Assessing Urban scene Quality Using Explainable Deep Learning A Comparative Study of EfficientNet-B3 and DenseNet121
Roqaia Taha; Ihsan Abbas; Manaf Altaleb
Presentation: Online
Abstract: Urban scene quality is considered one of the key factors influencing users’ perceptions, comfort, and interaction with urban environments. However, traditional assessment methods rely heavily on field observations and expert judgments, limiting their applicability on a large scale. Therefore, this study explores the potential of deep learning for assessing urban scene quality through two visual perception indicators: Imageability and Enclosure. To achieve this, EfficientNet-B3 and DenseNet121 models were trained on a dataset of urban images, and Grad-CAM was employed to interpret the decision-making process of each model. The results demonstrated that both models were capable of identifying visual elements that influence urban scene quality assessment, while showing differences in visual attention patterns and scene interpretation mechanisms. These findings highlight the potential of explainable deep learning as a more objective and transparent tool for supporting urban scene quality assessment and urban design practice.
13:30
13:45
Paper ID: 111
Cross-Attention TabTransformer with Explainable AI for Breast Cancer Classification
Rutaba Azmat; Shaiq Ahmad Khan; Faiq Ahmad Khan; Fraz Ahmad; Akhtar Jamil; Alaa Ali Hameed
Presentation: Online
Abstract: Breast cancer remains a leading cause of cancer-related mortality worldwide, necessitating accurate and interpretable computer-aided diagnostic systems. This paper presents a comprehensive comparative analysis of classical machine learning, deep learning, and a novel transformer-based architecture for breast cancer classification using the Wisconsin Diagnostic Breast Cancer dataset. We introduce a TabTransformer model with cross-attention mechanisms that explicitly model inter-group feature interactions between geometric and texture-based feature subsets. Our framework encompasses seven baseline classifiers, hyperparameter optimization via randomized search, recursive feature elimination for dimensionality reduction, and a soft-voting ensemble strategy. Furthermore, we integrate explainable AI techniques including SHAP global explanations, LIME local interpretations, and attention weight visualizations. The proposed cross-attention transformer achieves 98.25% classification accuracy and an F1-score of 0.9756, matching optimized multilayer perceptron performance while providing superior interpretability through attention-based feature interaction analysis. Feature selection reduces dimensionality from 30 to 15 attributes without performance degradation, confirming substantial redundancy in the original feature space.
Technical Session - 9 Online Time: 11:30 - 13:45. Location: Virtual Room -2
Join Session
Session Chairs: Francesco spina
11:30
11:45
Paper ID: 259
Four-Valued Fuzzy Feature Fusion for Explainable Disease Classification: Multi-Dataset Evaluation and External Clinical Validation
Mustafain Ali; Dr. Ubaida Fatima
Presentation: Online
Abstract: High-dimensional healthcare datasets often contain correlated clinical features that increase computational complexity and reduce model interpretability. This paper presents a Four-Valued Fuzzy Feature Fusion framework for explainable disease classification. The proposed approach transforms clinical and biomedical attributes into fuzzy representations and aggregates four clinically related features into interpretable fuzzy predictors using domain knowledge from the literature. The resulting fused predictor space reduces dimensionality while preserving diagnostically relevant information. Random Forest, Support Vector Machine, XGBoost, LightGBM, and CatBoost classifiers are employed to evaluate predictive performance. Explainability is achieved through SHAP-based analysis, enabling identification of the most influential predictors contributing to disease classification. Experiments conducted on multiple healthcare datasets demonstrate strong classification performance, improved computational efficiency, and enhanced interpretability. External validation using independent clinical records further confirms the robustness and practical applicability of the proposed framework for healthcare decision-support and intelligent diagnostic systems.
11:45
12:00
Paper ID: 145
Autonomous Last-Mile Delivery in Rural Environments via Integrated Perception, Navigation, and Cognitive Decision-Making
Adamos Daios; Ioannis Kostavelis
Presentation: Online
Abstract: Last-mile delivery represents the most inefficient and resource-intensive stage of modern logistics systems, particularly in rural environments characterised by sparse infrastructure, extended travel distances, and high operational uncertainty. This study presents DIVA, a framework for autonomous last-mile delivery based on driverless vehicles integrating perception, navigation, and cognitive decision-making capabilities.

The proposed approach focuses on semantic mapping in unstructured rural environments, adaptive route planning under localisation uncertainty, and behaviour-aware decision-making aligned with traffic regulations. Multi-modal sensing, combining LiDAR and vision data, is integrated with SLAM-based localisation methods to support reliable operation under degraded GPS conditions.

The framework follows a hierarchical architecture incorporating global route planning, local motion control, and probabilistic behaviour modelling for reliable and efficient autonomous navigation. The analysis indicates that the proposed system addresses key limitations of existing autonomous delivery approaches while improving the efficiency and reliability of rural logistics operations.
12:00
12:15
Paper ID: 1
Deep Learning Forecasting for Cryptocurrency Prices Using LSTM and Market Feature Fusion
leqaa abood; Mustafa AL-AZZAWI; Mustafa Lateef Fadhil Jumaili; wissam saad; Bharat Bhushan; Raghda Aldahi
Presentation: Online
Abstract: The cryptocurrency price is volatile, nonlinear, and depends on external factors like social sentiment and macroeconomic conditions, among others, making forecasting of cryptocurrency prices inherently difficult. Based on the paper, it is argued that CryptoLSTM is a deep learning model that incorporates multi-source market data using an attention-based feature fusion system to predict cryptocurrency prices accurately. The suggested model integrates technical metrics, on-chain blockchain performance, social media dynamics obtained on Twitter and Reddit, as well as macroeconomic metrics in a single Long Short-Memory LSTM framework. CryptoLSTM also works in a manner that is unlike the existing methods which operate by relying on fixed feature aggregation versus dynamically scaling the contribution of each category of features in response to the current market conditions with a learnable attention. Bitcoin (BTC), Ethereum (ETH), and Ripple (XRP) are subjected to extensive experiments on January 2018 to December 2023 and across various bull and bear market regimes on daily basis. The empirical findings indicate that CryptoLSTM is always competitive as compared to eight other competing baselines such as CNN-LSTM and Transformer models. Namely, CryptoLSTM demonstrates an RMSE of 142.3 USD and MAE of 108.7 USD on predicting price in Bitcoin on a daily basis, which is 20-23% better than the best baseline of all assets. Ablation experiments also confirm the role of sentiment and on-chain features, whereas the attention module itself leads to a reduction of RMSE by 23.4% over the fusion of all features. The suggested structure has empirical usefulness to merchants and securities managers, as it enhances predictive capabilities in times of high volatility, and to the scientific community, by availing data and code to an external public
12:15
12:30
Paper ID: 263
Reliable Large Language Model Routing for Non-Clinical Healthcare and Education Support
Zakaria Ouali; Sujita Kandagatla; Qiao Xu; Benyamin Ahmadnia
Presentation: Online
Abstract: Large Language Models (LLMs) are increasingly used as front-line support interfaces, but routing in sensitive support settings requires more than ordinary intent classification. A reliable router must decide not only which label to assign but also when to seek clarification, when to refuse clinical advice, or when to escalate urgent and academic-integrity cases. This paper presents a layered LLM routing framework for non-clinical healthcare support and computing-focused educational support. The framework combines deterministic safety and academic overrides, an information-sufficiency gate, LLM-based routing with structured JSON output, and a taxonomy-based label validation guard. We evaluate the framework on a 549-prompt pilot benchmark with 14 outcomes across education, healthcare, and priority-gating decisions. Compared with an embedding nearest-neighbor baseline, the proposed Gemma 4 12B Unified router improves overall accuracy from 61.75\% to 93.44\% and reduces wrong-confident routing from 37.7\% to 5.8\%. The results suggest that lightweight guardrails, clarification-aware routing, and taxonomy-constrained validation can improve reliability in applied LLM support systems.
12:30
12:45
Paper ID: 266
Automated Classification and Staging of Liver Fibrosis Using a Custom CNN Optimized with the Adadelta Algorithm
Khalid Alemerien; Saleel Alsarayreh; Mohammad Saleh; Sadeq Al-Suhemat
Presentation: Online
Abstract: Liver fibrosis is a pivotal stage in chronic liver disease. It disrupts the structure and function of the liver. Early detection and accurate diagnosis are essential to prevent the condition from worsening. Nowadays, physicians use ultrasound elastography as a non-invasive technique to assess this scarring. AI-based solutions are effective in liver fibrosis detection. One of these solutions includes AI tools based on deep learning. Furthermore, recent breakthroughs in deep learning are significantly improving diagnostic accuracy. In this study, we proposed a custom CNN-based model optimized with the Adadelta algorithm to classify liver fibrosis stages using ultrasound images. To examine the performance of the CNN-based model, we performed three experiments, including a binary classification (No Fibrosis vs. Fibrosis), which resulted in 100% accuracy; a three-class setup (No Fibrosis, Intermediate, and Advanced), which reached 99.8% accuracy; and a five-class model (F0, F1, F2, F3, and F4), which achieved 96.8% accuracy. The model performed best at the extreme stages (F0 and F4), with slight reductions in the middle stages due to feature overlap. These results indicate that the proposed CNN-based model with Adadelta can effectively identify and distinguish fibrosis stages. These results demonstrate that the proposed model is a reliable tool for automated liver fibrosis detection.
12:45
13:00
Paper ID: 267
Predictive Analytics in Sports: Machine Learning Approaches for Injury Prevention and Tactical Planning
Iffat Batool; Dr. Ubaida Fatima
Presentation: Online
Abstract: Injuries in sports are influenced by many reasons. Fatigue, workload imbalance, improper recovery time, and previous injury history are some main factors. the methodology of this framework consists of data preprocessing, feature engineering and tactical analysis. these are very useful for coaches, trainers and sports medical staff in order to make decision for injury prevention and improvement of performances of player. The findings show that all applied models of machine learning have strong capability to enhance players well-being, minimize rates of injury and improve tactical decision –making in sports domains. Despite these benefits, challenges like model interpretability, small datasets, and on field-implementation require more analysis. Further research may emphasis on including deep learning techniques, computer vision and more advancement like wearable technologies to create more accurate, robust and smart analytic frame work.
13:00
13:15
Paper ID: 122
Temporal Graph Neural Network for Criminal Community Detection
Malvina Halilaj; Bora Lamaj (Myrto); Erisa Bekteshi; Aldo Franco Dragoni; Elektra Myrto
Presentation: Online
Abstract: The rapid growth of online social platforms has increased the complexity of identifying emerging criminal communities and evolving suspicious communication patterns. Traditional machine learning approaches often struggle to capture dynamic interactions and temporal relational behavior between suspicious entities. This paper proposes a Temporal Graph Neural Network framework for emerging criminal community detection using temporal graph analysis and explainable artificial intelligence techniques. The proposed framework integrates temporal graph construction, graph neural representation learning, community evolution analysis, and explainability mechanisms to identify suspicious behavioral patterns across social platforms. Experimental results demonstrate that the framework improves detection performance while maintaining interpretability for forensic and security-oriented applications. The proposed approach supports transparent decision-making and demonstrates strong potential f
13:15
13:30
Paper ID: 150
AI-based Automated Building Energy Analysis from Point Cloud Data: Case Study on the Mascaro Center for Sustainable Innovation
Xiangdong Yan; Federica Geremicca; Melissa Bilec; John Brigham; Alessandro Fascetti
Presentation: Online
Abstract: This paper presents the development and implementation a fully automated pipeline connecting three-dimensional point cloud semantic data to generate building energy simulations via the EnergyPlus software suite. High-fidelity indoor scans are leveraged to generate high-resolution 2D images, from which building element instances are extracted via open-vocabulary instance segmentation, classified by surface material using a zero-shot CLIP-based recognizer, and exported as structured records encoding geometry, semantic class, and material properties. The resulting semantic and geometric data is ingested by a Python pipeline that performs coordinate substitution, boundary condition assignment, material inheritance from a validated reference model, and automated EnergyPlus simulation. The pipeline is demonstrated on a selected section of the Mascaro Center for Sustainable Innovation at the University of Pittsburgh, comprising one internal room. The proposed semantic segmentation achieves an Overall Accuracy of 0.978, Mean Accuracy of 0.952, Mean IoU of 0.992, and Macro F1 of 0.974 compared to the manually annotated ground-truth point cloud, with geometric errors of 2.7\%, 1.0\%, and 0.3\% for length, width, and height, respectively, against CAD reference measurements. To contextualize the pipeline outputs and examine procedural differences, a manually authored BIM model was constructed for the same selected section of the building. The comparison demonstrates that the proposed approach eliminates the manual modeling step required by existing scan-to-energy workflows and that building energy model generation and simulation can be achieved in minutes rather than hours, with direct implications for scalability for large building inventories and sustainable design and operation of vertical infrastructure.
13:30
13:45
Paper ID: 294
A Browser-Based Multi-Signal Artificial Intelligence System for Real-Time Detection of AI-Generated Phishing Emails and Fraudulent Websites
Yu Sun
Abstract: The rapid adoption of generative artificial intelligence has dramatically increased the scale and sophistication of online scams, rendering traditional, signature-based phishing filters increasingly ineffective. Modern attackers employ large language models to produce grammatically flawless, contextually personalized phishing emails, polymorphic campaigns that generate thousands of unique variants, and deceptive websites that closely imitate trusted services. This paper presents GotchaGuard, a browser-extension-based security system that fuses multiple independent detection signals to estimate scam likelihood for emails and web content in real time. The system couples a lightweight Chrome extension with a Flask backend that integrates a trained Random Forest text classifier, domain and file reputation scanning, large language model analysis, and screenshot-based visual inspection. By combining textual, structural, reputational, and visual evidence rather than relying on a single indicator, the proposed approach detects threats that evade individual filters. The classifier was evaluated on a labeled corpus of legitimate and malicious emails and achieved 96.4% accuracy and a 96.1% F1-score, while the integrated multisignal pipeline reached 94.8% accuracy on a mixed evaluation set of emails and fraudulent websites. These results demonstrate that multi-signal fusion delivered through an accessible browser interface provides a more comprehensive and adaptive defense against AI-driven phishing than single-method baselines.
Technical Session - 10 Online Time: 11:30 - 13:45. Location: Virtual Room - 3
Join Session
Session Chairs: Riccardo Raciti
11:30
11:45
Paper ID: 306
Design of a Multi-Dimensional Customer Experience Platform for Telecommunications
Alpaslan Gökcen; Fatih Çögen; Ramazan Yağmur; Saliha Sezgin Alp
Abstract: Customer experience (CX) management in large scale telecommunications networks is difficult because relevant data is scattered across network measurements, location intelligence, customer systems, and application usage indicators. Existing tools are often limited to proprietary and vendor specific data environments. This paper presents a Customer Experience Platform (CXP) that unifies heterogeneous data sources within a common and vendor independent framework. The platform follows a layered architecture that supports consistent data integration and feature generation across multiple domains. Based on this foundation, it produces multi-dimensional experience views across spatial, temporal, and usage-oriented perspectives, enabling finer monitoring and more actionable insights. The proposed platform is designed around interoperability, modularity, and practical deployment, while also allowing advanced analytics modules such as churn and complaint prediction to be integrated without changing the core architecture. The platform supports use cases including network optimization, investment prioritization, and customer analytics.
11:45
12:00
Paper ID: 226
SpillGuard: An Automated Oil Spill Detection System Using SAR Imagery and Deep Learning for the Arabian Gulf
Manar Qahtni
Presentation: Online
Abstract: Oil spills in the Arabian Gulf pose severe ecological and economic threats, yet timely detection remains challenging due to the limitations of conventional monitoring approaches. This paper presents SpillGuard, an automated oil spill detection system that applies transfer learning on Sentinel-1 Synthetic Aperture Radar (SAR) imagery to perform binary classification of image patches as Oil or No-Oil. A dedicated SAR preprocessing pipeline, comprising Lee filtering for speckle reduction, logarithmic backscatter transformation, percentile clipping, and Contrast Limited Adaptive Histogram Equalization (CLAHE), was developed to enhance discriminative features prior to model training. A ConvNeXtTiny backbone pretrained on ImageNet-1K was fine-tuned using a two-phase transfer learning strategy on the CSIRO Sentinel-1 SAR Oil/No-Oil dataset (5,630 images), with class imbalance addressed through a WeightedRandomSampler and a positive-class-weighted binary cross-entropy loss to prioritize recall. On the held-out test set (n = 845), SpillGuard achieved 95.62% accuracy, 97.20% Oil-class recall, and an AUC-ROC of 99.42%, outperforming prior SAR-based methods in recall while operating on a challenging dataset containing oil lookalike phenomena. Threshold analysis further demonstrated that the decision boundary can be adjusted operationally to minimize false negatives during high-risk periods in the Arabian Gulf. These results establish SpillGuard as a robust, data-driven foundation for scalable marine pollution monitoring aligned with Saudi Arabia's Vision 2030 environmental sustainability objectives.
12:00
12:15
Paper ID: 271
Explainable Diabetes Prediction Using Bayesian-Optimised Random Forest and SHAP
Parth Gangani; Sana Alyaseri; Alaa Aljanaby
Presentation: Online
Abstract: Type 2 diabetes mellitus remains a major global health concern due to its increasing prevalence and serious long-term complications. Early identification of high-risk individuals is essential for timely intervention and improved clinical outcomes. This study proposes an explainable diabetes prediction framework using a Random Forest classifier optimised through Bayesian hyperparameter tuning and interpreted using SHAP (SHapley Additive exPlanations). Experiments were conducted on a publicly available diabetes prediction dataset containing demographic, lifestyle, and clinical variables. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. The optimised Random Forest achieved 97.21\% accuracy, 99.31\% precision, 46.85\% recall, 63.66\% F1-score, and 73.41\% ROC-AUC, improving precision and F1-score while maintaining comparable overall performance relative to the baseline model. SHAP analysis identified HbA1c level, blood glucose level, BMI, and age as the most influential predictors, consistent with established clinical evidence. The findings demonstrate that Bayesian-Optimised Random Forest provides a practical balance between predictive performance and interpretability, making it a promising tool for diabetes risk stratification, triage prioritisation, and AI-assisted clinical decision-support systems, while also supporting sustainable and efficient healthcare resource prioritisation.
12:15
12:30
Paper ID: 155
AI-Driven Operational Performance Assessment: Isolating the Human Factor in Maritime Energy Management
Aurora Malegieri; Luisa Montella; Teresa Murino; Andrea Somma; Monica Strazzullo
Presentation: Online
Abstract: The transition toward maritime decarbonization under stringent IMO and EU regulations highlights the critical need to optimize ship operational efficiency. While current literature heavily prioritizes technical and design retrofits, a persistent performance gap remains due to the unquantified impact of the human factor onboard. To bridge this methodological gap, an integrated data-driven framework is introduced to evaluate and incentivize the operational performance of Masters and Chief Engineers. Following data preprocessing and peer group stratification, main engine ($ME$) propulsive baselines are established via an orthogonalized LS-PLS regression, while auxiliary energy demands are mapped through adaptive K-Medoids clustering. Validation on a fleet of 13 Ro-Pax vessels demonstrates that the framework effectively purges environmental and seasonal biases, correcting significant ranking distortions and preventing unfair performance appraisal. Crucially, by synthesizing performance via a Fuzzy Best-Worst Method (BWM), the model quantifies that aligning underperforming officers with historically achievable benchmarks yields potential annual fleet-wide savings of 2.2\% in propulsive fuel and 1.8\% in port-stay auxiliary energy, translating into an estimated economic return of up to \texteuro\,2.4M. The proposed methodology establishes a scientifically rigorous and operationally equitable appraisal foundation, providing a strategic tool to align corporate maritime incentive schemes with global sustainability targets.
12:30
12:45
Paper ID: 273
Automated Multiclass Brain Tumor Classification Using MRI Images: A Comparative Deep Learning Approach
Samar M. Alqhtani
Presentation: Online
Abstract: Classifying brain tumors based on MRIs plays an important role in medical imaging because early diagnosis can significantly influence treatment and outcome. Transfer learning utilizing deep convolutional neural networks (CNNs) has gained more popularity over recent years because it enables researchers to take advantage of existing models trained on big datasets, which may be especially helpful when dealing with small amounts of medical data. This research investigates three pre-trained models such as MobileNetV2, DenseNet121, and VGG121. The model were evaluated using MRI images The models were investigated for their capacity in detecting multi-class brain tumors using conventional metrics including accuracy, precision, recall, F1-score, confusion matrices, and ROC-AUC graphs. The DenseNet121 model outperformed by achieving an accuracy rate of 90% followed by VGG16 model attained an accuracy of (88%). Moreover, the ROC-analysis further confirms the strong class seperability across all models, with high AUC score.
12:45
13:00
Paper ID: 274
An LLM-in-the-loop RL Framework for Bioinformatics Feature Selection
Xinyuan Wang; Deepti Agrawal; Yanjie Fu
Presentation: Online
Abstract: High-dimensional bioinformatics data, characterized by a large number of features relative to the number of samples, pose major challenges such as the ``curse of dimensionality,'' leading to overfitting, high computational cost, and poor generalization. Traditional feature selection methods often suffer from limited scalability and adaptability in such domains. We propose an LLM-in-the-loop reinforcement learning (RL) framework for bioinformatics feature selection, where the RL agent formulates feature selection as a sequential decision-making task, while the large language model (LLM) enhances the process in two ways: (1) guiding exploration through domain-informed advice, and (2) providing hybrid rewards that integrate data-driven performance with knowledge-driven evaluation. The LLM also produces explanations to improve interpretability for human experts without altering the RL policy update. Experiments on diverse bioinformatics datasets show that the LLM-in-the-loop framework outperforms baselines, achieves stable performance across downstream models, and converges faster than pure RL.
13:00
13:15
Paper ID: 156
Ports of the Future: AI-Driven Decision Support in Intermodal Container Terminals
Emmanouil Lemonias; Eleni Vrochidou; George Papakostas
Presentation: Online
Abstract: In this work, a critical literature review of artificial intelligence (AI) applications in container terminal operations is conducted, focusing on intermodal logistic hubs that combine sea, road and rail transport. The focus is on the examination of predictive methods such as Estimated Time of Arrival (ETA) forecasting, throughput prediction and gate flow management, as well as prescriptive techniques like berth allocation, crane scheduling and hybrid AI with operations research (OR). Additionally, autonomous approaches are explored, including reinforcement learning (RL) for yard stacking and digital twins (DTs). Quantitative bibliometric analysis revealed a significant surge in research after 2022, highlighting ongoing dependence on synthetic datasets. Specifically, four inter-related gaps are further identified: (i) fragmented optimization that overlooks cross-functional spillover effects, (ii) lack of end-to-end predict-then-optimize pipelines, (iii) limited validation with real-world data and (iv) inadequate consideration of CO2 emissions as a strict constraint. Building on this analysis, an integrated AI framework is proposed, combining long short-term memory (LSTM)-based ETA prediction with Monte Carlo Q-learning for yard optimization and a rail synchronization module. The proposed framework addresses identified gaps in the literature by using real operational data from an intermodal terminal with rail connectivity.
13:15
13:30
Paper ID: 157
A Leakage-Safe Multimodal Medical Generalist Benchmark: Chest X-Ray Classification and Image-to-Impression Generation under Natural Imbalance
Hasibul Islam; Zidan Zafar Rudra; Ahmed Faizul Haque Dhrubo; Souvik Pramanik; Asadullah Hil Galib; Saif Ahmed; Mohammad Ashrafuzzaman Khan; Mohammad Abdul Qayum; Mohsin Sajjad
Presentation: Online
Abstract: Real-world clinical imaging streams are naturally imbalanced, which can severely bias models toward majority conditions and artificially inflate benchmark performance when evaluation protocols unintentionally leak patient-level information. We present a strictly audited, leakage-safe evaluation protocol for a naturally imbalanced chest X-ray benchmark and systematically study a multimodal “medical generalist” setting spanning two clinical tasks: (i) image-only disease classification and (ii) image-to-text impression generation. To definitively prevent leakage, we cluster lexically identical reports via cryptographic hashing to form robust patient groups, ensuring zero group overlap across splits. For classification, we establish a robust ResNet18 baseline under natural skew, evaluating via macro-F1 and balanced accuracy. For multimodal generation, we train an image-conditioned Transformer decoder and rigorously evaluate decoding strategies (greedy, beam, nucleus). Because standard n-gram metrics correlate poorly with clinical accuracy, we introduce a negation-aware clinical finding extraction metric to evaluate diagnostic correctness. Finally, we extract and upsample terminal cross-attention maps to provide spatial, word-level grounding of generated pathology terms. This work establishes a reproducible, rigorously audited baseline suite and analysis toolkit for multimodal CXR modeling.
13:30
13:45
Paper ID: 282
Hybrid Queueing–ML-Based Dynamic Edge Node Activation for Energy-Efficient IoT–Edge Computing
Oumaima Ghandour; Said El Kafhali; Mohamed Hanini; Luis Orozco Barbosa
Presentation: Online
Abstract: Edge computing enables low-latency processing of Internet of Things (IoT) tasks by bringing computation closer to data sources. However, keeping all edge nodes continuously active increases energy consumption, while aggressive node sleeping may degrade Quality of Service (QoS) and increase Service Level Agreement (SLA) violations. This paper proposes a hybrid Queueing-Machine Learning (Queueing-ML) framework for energy-aware edge node activation in IoT-Edge computing environments. Each edge node is modeled using an M/M/1/K queue to estimate congestion indicators such as utilization, blocking probability, queue length, and waiting time. These queueing metrics, with workload and temporal features, are used to train ML models for future congestion risk prediction. The predicted risk is then combined with queueing-based indicators to guide dynamic activation decisions. Experiments using a real IoT workload and a MEC edge topology show that ML-based risk prediction can support proactive congestion awareness. The results also show that ML-only activation provides high energy savings but causes significant QoS degradation, while the proposed hybrid Queueing-ML policy provides a more controlled energy-QoS trade-off by reducing aggressive node sleeping decisions.
Technical Session - 11 Online Time: 13:45 - 16:00. Location: Virtual Room -1
Join Session
Session Chairs: Dario Samuele PISHVAI
13:45
14:00
Paper ID: 158
Raspberry Pi Pico W-Based Autonomous Payload Delivery Rover with Intelligent Payload Verification and Firebase IoT Integration
Sabbir Ahmed Sakil; Nazib Riasat; Meherin Chowdhury; Quazi Md Sadman; Ahmed Faizul Haque Dhrubo; Mohammad Abdul Qayum; Mohsin Sajjad
Presentation: Online
Abstract: Our paper presents the design, implementation, and experimental evaluation of an intelligent autonomous payload delivery rover built on the Raspberry Pi Pico W microcontroller, targeting accessible deployment in developing economies. The system integrates a five-channel infrared sensor array governed by a priority-ordered rule-based expert system for real-time line following, a forward-facing HC-SR04 ultrasonic sensor for obstacle detection, and an HX711-based gravimetric payload verification subsystem employing a bounded probabilistic evidence accumulator to suppress false confirmations under vibration. A formally modeled six-state cyber-physical finite state machine with explicit safety invariants governs all operational transitions including waiting, running, obstacle avoidance, and delivery confirmation ensuring deterministic and safe behavior under varied environmental conditions. Cloud connectivity is achieved through Google Firebase Realtime Database over Wi-Fi, enabling live telemetry reporting and remote command injection, and positioning the platform for future AI-driven route optimization and fleet management. The rover achieves a line-following track completion rate of approximately 92%, an obstacle response latency of approximately 110 ms, and a payload detection accuracy of 96% under controlled conditions, at a total hardware cost of approximately 5,200 BDT ($47 USD).
14:00
14:15
Paper ID: 159
YouTube Audience Analyzer: An AI-Supported Multilingual Audience Intelligence System Based on Comment Analysis
Hilal Çalışkan; İlkim Ecem Emre
Presentation: Online
Abstract: The rise of user-generated content on various social media channels has presented immense opportunities for analyzing the views, preferences, and behaviors of the audiences. In particular, the channel of YouTube produces numerous multilingual comments that offer useful information about trends in public discussions. At the same time, manual analysis of these comments is not only highly inefficient but also extremely time consuming. The current paper describes the design and development of an AI-based YouTube Audience Analyzer as a tool for automated collection, processing, and interpretation of multilingual YouTube comments. The proposed tool gathers YouTube comments through the use of the YouTube Data API and processes them via multi-step natural language processing that includes language detection, text preprocessing, sentiment analysis, topic extraction, and insights based on AI. Specifically, multilingual comments are automatically grouped according to their language and preprocessed accordingly. Sentiments are classified using transformer-based models, while topics are extracted semantically by the system. Differing from other research efforts that emphasize only sentiment analysis, the suggested model encompasses multilingual processing, topic identification, automatic explanation, and visualization under one roof. The created solution will help users to gain useful insights about their audiences by converting vast amounts of YouTube comments into valuable audience information.
14:15
14:30
Paper ID: 291
TRIPROBE: Probing Task Separability Beyond Classification for XAI
Mahyar Shahsavari
Presentation: Online
Abstract: Modern evaluation of learning pipelines often reduces to downstream accuracy, leaving open the question of why tasks succeed or fail. TriProbe addresses this gap with a multi-level probing framework for explainable diagnosis of task separability. Rather than treating models as black boxes, TriProbe traces how separability evolves across inputs, learned features, and final classifiers. It decomposes multi-task problems into binary subtasks and applies three complementary probes: a Foundational Probe on input spaces, a Latent Probe on feature representations, and a Final Probe on classifier outputs. Using Maximum Fisher’s Discriminant Ratio as a principled separability metric, TriProbe identifies bottlenecks and affected task pairs. Experiments on the Roshambo sEMG benchmark show how TriProbe reveals hidden breakdowns, guiding data collection, validation, and architecture design.
14:30
14:45
Paper ID: 300
Secure Fault Detection and Diagnosis in Networked Three-Tank Systems Using Secret Sharing and Enhanced Fuzzy Feature Engineering: A Comparative Study of Classical and Quantum Machine Learning
Dr. Ubaida Fatima; Tabish Ahsan
Presentation: Online
Abstract: Fault detection and diagnosis are essential for ensuring the reliability and security of industrial control systems. This paper proposes a secure fault diagnosis framework for networked three-tank systems by integrating secret sharing, enhanced fuzzy feature engineering, and machine learning techniques. Secret sharing is employed to protect process measurements during network communication, while residual signals are generated to identify abnormal system behavior. To model uncertainty and fault severity, an enhanced fuzzy feature set consisting of a Fuzzy Fault Score (FFS), Sensor Fault Degree (SFD), Actuator Fault Degree (AFD), Leakage Fault Degree (LFD), and Enhanced Fuzzy Index (EFI) is developed. These features are combined with residual and dynamic process information and supplied to classical and quantum machine learning classifiers. Experiments are conducted under normal, sensor fault, leakage fault, and actuator fault conditions. Comparative analysis involving Random Forest, SVM, KNN, XGBoost, LightGBM, QSVC, and VQC demonstrates that ensemble learning models provide superior fault classification performance. XGBoost and LightGBM achieve the highest accuracy of 97.22%, while quantum models show lower performance. The results confirm the effectiveness of the proposed secure fuzzy-enhanced framework for intelligent fault diagnosis in networked industrial systems.
14:45
15:00
Paper ID: 301
Exploring Fusion Strategies in Audio–Text Multimodal Learning: A Comparative Study on French Podcast Classification
H.Hakan Kilinc, Zeynep Hilal Kilimci Hasan Kilimci
Presentation: Online
Abstract: Multimodal learning has demonstrated significant potential in speech-related classification tasks by integrating complementary information from heterogeneous data sources. While audio signals capture acoustic and prosodic characteristics, textual transcriptions provide semantic and contextual information. This study presents a comprehensive evaluation of multimodal fusion strategies for French podcast topic classification using aligned audio–text data. A multimodal dataset comprising five podcast categories, namely News, Music, Personal Development, Spirituality, and Sports, was constructed from podcast recordings and their corresponding transcriptions. Initially, multiple unimodal architectures were evaluated for both modalities, including CRNN, Whisper, and Wav2Vec2 for audio classification, and BiLSTM, FlauBERT, and CamemBERT for text classification. Based on the unimodal evaluation results, Wav2Vec2 and CamemBERT were selected as backbone encoders for multimodal experiments. Subsequently, three multimodal fusion paradigms, namely Early Fusion, Intermediate Fusion, and Late Fusion, were investigated under a stratified five-fold cross-validation protocol. Experimental results demonstrate that the effectiveness of multimodal learning strongly depends on the selected fusion strategy. Early Fusion and Intermediate Fusion achieved accuracies of 79.08% and 83.18%, respectively, whereas the best performance was obtained using a stacking-based Late Fusion approach, which achieved 87.60% accuracy, 87.65% F1-score, and 97.84% ROC-AUC. Furthermore, the proposed fusion framework surpassed the strongest unimodal baseline, Wav2Vec2, demonstrating the complementary nature of acoustic and semantic information. The findings indicate that decision-level fusion provides the most effective mechanism for exploiting multimodal representations and highlight the importance of fusion strategy selection in audio–text classification systems.
15:00
15:15
Paper ID: 304
Evaluation of Network Security Solutions for 5G Networks with Machine Learning Techniques on Pardus OS
Fawzy Alsharif; Emir Basaran; Ahmet Agca; Furkan Koc
Abstract: The rapid evolution of 5G networks introduces new challenges in terms of security, scalability, and threat detection. Traditional intrusion detection systems and datasets often fail to represent the dynamic and heterogeneous nature of modern 5G environments. In this context, this study proposes a comprehensive machine learning-based network intrusion detection framework designed for 5G networks, with a focus on end-user device applicability. The proposed approach systematically evaluates widely used network monitoring tools—Suricata, Argus, and Zeek—in terms of feature generation capability, resource efficiency, and compatibility with machine learning requirements. Furthermore, several publicly available datasets are analyzed, and the 5G-NIDD dataset is selected due to its real-world 5G testbed origin and rich feature space. A series of machine learning experiments are conducted using multiple classification algorithms, and the Decision Tree model is selected based on a balance between performance and interpretability. Experimental results demonstrate that Argus provides the highest compatibility with the 5G-NIDD feature space, significantly improving model performance compared to alternative monitoring tools. Additionally, the Decision Tree model achieves competitive accuracy while maintaining explainability, making it suitable for security-critical applications. Finally, a conceptual end-user defense mechanism is proposed to enable policy-driven responses to detected threats.
15:15
15:30
Paper ID: 305
AI for Detecting Fungal Crop Diseases
Fawzy Alsharif; Kerem Münüklü; Görkemcan Açgül; Emir Başaran
Abstract: This study presents a deep learning-based system enhanced with image processing techniques for the automatic detection of fungal diseases—specifically green mold and dry bubble—that lead to significant yield losses in Agaricus bisporus (button mushroom) cultivation. The proposed approach utilizes two separate YOLOv12-based artificial intelligence models, each tailored to a distinct stage of disease development. The first model is designed to detect dry bubble disease on the mushroom surface. By applying segmentation to isolate mushrooms and eliminate compost areas, the model focuses exclusively on disease symptoms appearing on the fruit body. This model achieved a mean Average Precision (mAP) of 99.7% at an IoU threshold of 0.5 and 79.3% across thresholds from 0.5 to 0.95. The second model detects green mold, which typically develops on the compost surface. In this case, compost areas were segmented while mushrooms were excluded, enabling the model to specialize in detecting green mold infections. This model achieved 99.9% mAP@0.5 and 81.8% mAP@0.5–0.95. Both models contribute to early disease identification, reduce dependence on manual inspection, and improve yield quality and efficiency. The results demonstrate the strong potential of AI-supported systems in precision agriculture and disease monitoring in mushroom production.
15:30
15:45
Paper ID: 279
AI-Powered Fuzzy Logic Framework for Optimizing Medical Ultrasound Imaging
Hanane Sefraoui; Abdechafik Derkaoui; Abdelhak Ziyyat
Abstract: Ultrasound imaging systems exhibit distinct acoustic behaviors in the near-field (Fresnel region) and far-field (Fraun hofer region), where beam quality, intensity distribution, and spatial resolution vary significantly because of diffraction and interference effects. In medical ultrasound imaging, controlling the transition between these regions is essential for achieving optimal image resolution and acoustic energy focusing. This work proposes a fuzzy logic–based adaptive framework for optimizing ultrasound beam characteristics through the dynamic adjust ment of transducer aperture size and operating frequency. The proposed system translates qualitative beam quality indicators into control actions that regulate these parameters, thereby enhancing near-field stability and improving far-field focusing performance. From a clinical perspective, this adaptive approach minimizes operator dependency, enhances imaging reproducibil ity, and ensures consistent acoustic field uniformity and diagnostic clarity across varying tissue depths without inducing harmful thermal effects. The framework was implemented and validated using MATLAB-based k-Wave acoustic simulations, enabling the analysis of ultrasound wave propagation and RMS pressure field distributions under different optimization strategies.
15:45
16:00
Paper ID: 86
Leveraging an AI-Based Unified Framework for Anomaly Detection in Yacht Power Systems
Pierpaolo Dini; Davide Paolini; Sergio Saponara; Maurizio Minossi
Abstract: Multivariate time-series anomaly detection is essential for monitoring integrated power systems in luxury yachts, characterized by tightly coupled dynamics and high-dimensional, non-stationary data. This paper presents a unified framework combining prediction, reconstruction, and physics-informed constraints to enforce both statistical and physical consistency, enabling detection of functional anomalies and data acquisition faults. The model is trained on real-world yacht data, while evaluation is performed on synthetic datasets derived from real measurements with injected anomalies for controlled testing. Results show robust detection performance across scenarios, with ablation studies confirming the complementary roles of prediction and reconstruction and the effectiveness of constraint-based analysis in reducing false positives. The framework also supports interpretable event-level root cause analysis for real-time monitoring.
Technical Session - 12 Online Time: 13:45 - 16:00. Location: Virtual Room -2
Join Session
Session Chairs: Daniele Cocuzza
13:45
14:00
Paper ID: 289
A Cost-Effective SDR-Based Framework for Sustainable Drone Detection, Alerting, and RF Response Readiness
Abdullah Albuali
Abstract: Unauthorized drones pose security risks to public venues, restricted areas, and critical infrastructure, highlighting the need for low-cost RF-based detection systems. This paper presents a software-defined radio framework that employs HackRF, GNU Radio, threshold-based RF sensing, and automated audio alerting to detect drone-related activity in the 2.4 GHz band. Open-field experiments using a DJI Mini 3 Pro were conducted to evaluate two antennas at distances ranging from 5 m to 200 m. Both antennas achieved 100% detection accuracy up to 30 m, while the 2.4 GHz antenna maintained 90% accuracy at 100 m and 85% accuracy at 200 m. An RF jamming branch was also evaluated to assess response readiness. The results confirmed SDR-based signal generation capability; however, practical disruption would require a higher-power RF front end, directional antenna gain, stable power delivery, and appropriate regulatory authorization. Overall, the results demonstrate a practical low-cost detection-and-alerting baseline and identify the hardware requirements for future controlled RF response studies.
14:00
14:15
Paper ID: 307
Generating Synthetic Cardiovascular Data from a Small Clinical Dataset: A Fidelity, Utility and Privacy Evaluation of Four Tabular Generators
Majid Liaquat; Chris Nugent; Ian Cleland; Naveed Khan
Presentation: Online
Abstract: Access restrictions on clinical data limit the open sharing of healthcare datasets, which has motivated the use of Synthetic Data (SD) as a privacy preserving alternative. An open question is whether useful SD can be learned from the small datasets that are common in clinical studies. We study this on a cardiovascular disease (CVD) dataset. Four tabular generators were trained on a small stratified subset of 1,200 records: Conditional Tabular GAN (CTGAN), Tabular Variational Autoencoder (TVAE), CopulaGAN, and a quantile-transform CTGAN hybrid (QT-CTGAN). Each generator produced SD at one, two and four times the training size. We evaluated fidelity (Jensen-Shannon divergence, Wasserstein distance, the Kolmogorov-Smirnov statistic, and the Frobenius norm of the correlation matrix difference), utility (classifiers trained on SD and tested on held-out real data, compared with a real-data baseline on the same test set), and privacy (a nearest neighbour distance ratio, and a membership inference attack). TVAE produced the most faithful data and the only SD whose classifiers matched or exceeded the realdata baseline for Random Forest, XGBoost, and a Multilayer Perceptron, although it carried the highest privacy risk, and that risk grew with the generation volume. This ranking held across five random seeds, with bootstrap confidence intervals and a paired significance test. The GAN-based generators preserved privacy well but produced data of low utility. Increasing the volume from one to four times gave little benefit in fidelity and utility. The results suggest that, for small clinical tabular datasets, the choice of generator matters more than the amount of data generated, and that the privacy and utility trade-off should be checked before any SD is shared.
14:15
14:30
Paper ID: 312
Beyond Unimodal Accent Recognition: Investigating Audio–Text Fusion Strategies for French Accent Classification
Hasan Kilimci, Zeynep Hilal Kilimci H.Hakan Kilinc
Presentation: Online
Abstract: Accent classification plays an important role in speech processing applications such as automatic speech recognition, speaker profiling, and human--computer interaction. However, distinguishing between closely related accents remains challenging due to subtle acoustic and linguistic variations. This study presents a multimodal transformer-based framework for French accent classification by jointly exploiting speech and textual information. A multimodal dataset comprising 917 speakers from five French accent groups, namely France, Belgium, Canada, Switzerland, and Cameroon, was constructed from publicly available YouTube recordings. To ensure realistic evaluation, all experiments were conducted under a strictly speaker-independent stratified five-fold cross-validation protocol. For the audio modality, ResNet18, Wav2Vec2-XLSR-53, and Whisper-Small were evaluated, while BiLSTM, XLM-RoBERTa, and CamemBERT were employed for textual modeling. The best-performing unimodal models, Whisper-Small and CamemBERT, were subsequently integrated using Early Fusion, Intermediate Fusion, and Late Fusion strategies. Experimental results demonstrate that transformer-based architectures consistently outperform conventional deep learning models across both modalities. Furthermore, all multimodal fusion strategies surpassed the corresponding unimodal baselines, confirming the complementary nature of acoustic and linguistic information for accent classification. Among the investigated fusion approaches, Early Fusion achieved the best overall performance with an accuracy of 73.20% and a macro F1-score of 71.59%, substantially outperforming the best audio-only and text-only models. The findings highlight the effectiveness of multimodal transformer representations and demonstrate that feature-level integration provides a more effective mechanism for accent classification than attention-based or decision-level fusion strategies on the considered dataset.
14:30
14:45
Paper ID: 160
AI for Place-Belongingness: Actionable Guidance, Pathways, and Implications for Mobility
Hesam Mohseni; Ramon Chaves; Daniel Schneider; António Correia
Presentation: Online
Abstract: Place-belongingness (or the lack thereof) is a multifaceted phenomenon experienced by digital workers, students, university staff, competitive athletes, and many other individuals worldwide. From people who frequently travel while working remotely through digital technologies to local individuals and communities shaping territories and destinations, place-belongingness is often complex to measure and even more difficult to achieve and sustain under conditions of mobility. Consider, for example, individuals who join a football club or begin their studies at a university abroad while leaving their families behind. How do they develop a sense of belonging in a new environment? Which sociotechnical factors contribute to enhancing their sense of place-belongingness while mitigating feelings of exclusion and loneliness? These challenges highlight the need for targeted solutions. In this paper, we discuss the theoretical and practical implications of understanding place-belongingness in the age of artificial intelligence (AI), focusing on how AI-mediated experiences may strengthen individuals’ sense of place-belongingness during mobility and transition.
14:45
15:00
Paper ID: 161
Noise-Tolerant AI Framework for Comprehensive PV System Fault Diagnosis and Localization
Tharindie Pilapitiya; Gajindu Herath; Pramodya Sahan; Logeeshan Velmanickam; Sisil Kumarawadu; Chathura Wanigasekara
Presentation: Online
Abstract: As photovoltaic (PV) systems become increasingly integral to modern power grids, ensuring their reliability is paramount. Current fault diagnosis methods predominantly address array-level anomalies using aggregated measurements, which restricts string-level fault localization and limits their effectiveness under noisy, real-world sensor conditions. This paper proposes a comprehensive artificial intelligence (AI) framework with noise-tolerance designed to diagnose PV array faults (line-to-line, open circuit, and degradation) and inverter failures (IGBT open circuit and DC-link capacitor degradation). Utilizing a 5kW grid-connected PV system modeled in MATLAB/Simulink, four baseline architectures were evaluated: CNN-1D, LSTM, CNN-BiLSTM, and XGBoost. The results demonstrate highly accurate, task-specific performance. Crucially, the study evaluates the framework's resilience against measurement noise, demonstrating that training with noise-injected data acts as a powerful regularization technique, significantly enhancing diagnostic stability and localization robustness for practical deployment.
15:00
15:15
Paper ID: 326
Malware Detection Across Multi-Platform Ecosystems Using Stacked Generalization
Bilal Bakartepe; Kerime GENÇAY; Hakan Kutucu; İsa Avcı
Presentation: Online
Abstract: Malware detection research has long been constrained by the lack of publicly available, up-to-date, and multi-platform datasets. This study presents a multi-platform static analysis framework evaluated on the new EMBER 2024 dataset, which covers 6 different file formats: Win32, Win64, .NET, APK, ELF, and PDF, and includes over 3.2 million samples. Unlike existing single classifiers, a "Layered Stacked Generalization" architecture is proposed, combining a Ridge regression meta-learner with first-layer models specialized in raw thrember features (EMBER feature version 3) as well as engineering feature groups divided into 10 functional domains (imports, sections, entropy statistics, etc.). Experimental results confirm that this proposed stacked architecture exhibits a more robust and resilient performance than the baseline LightGBM model, especially on the “Challenge” set which is designed to evade antiviruses and where standard models typically struggle.
15:15
15:30
Paper ID: 317
The impact of model size and RAG on the response quality of local LLMs: evaluation of the Qwen3.5 family on a domain-specific dataset
Kristian Dokic; Svjetlana Letinic; Katarina Potnik Galic
Abstract: This paper presents a systematic empirical evaluation of Retrieval-Augmented Generation (RAG) across seven sizes of the Qwen3.5 language model family (0.8B–122B parameters) on a domain-specific Croatian accounting benchmark derived from the publicly available mariozupan/rrif dataset. All models were deployed locally using Ollama on a single GPU server, reflecting a realistic privacy-preserving enterprise scenario. Without RAG, lookup Exact Match accuracy is effectively 0% across all model sizes, confirming that Croatian Chart of Accounts knowledge is absent from general pretraining corpora. Introducing RAG raises lookup accuracy uniformly to 93–96%, with no meaningful relationship to model size — the 0.8B model matches the 122B model exactly. Hallucination rates drop from 7–53% without retrieval to exactly 0.0% under RAG for all seven models, indicating that hallucination in this domain is a knowledge-gap rather than a model-behavior phenomenon. Semantic similarity for descriptive questions improves monotonically with model size under RAG (0.714 at 0.8B to 0.773 at 122B), making it the only size-dependent metric. These findings demonstrate that RAG is the dominant factor for factual accuracy in domain-specific accounting tasks, and that the Qwen3.5-0.8B model with RAG provides a fully competitive deployment profile at 1.0 GB VRAM and sub-0.5 second average response time.
15:30
15:45
Paper ID: 292
An Integrated School Safety Platform Combining AI-Powered Audio Threat Detection, Real-Time Mobile Monitoring, and Interactive Lockdown Training
Ang Li; Chace Sun; Yu Sun
Abstract: School safety remains a critical concern, with over 300 school shooting incidents reported in the United States between 2018 and 2024. Existing safety systems predominantly rely on visual surveillance or gunshot acoustic detection, lacking the ability to identify verbal threats before an incident escalates. This paper presents Viotrol, an integrated school safety platform comprising three interconnected components: (1) an AI-powered real-time audio threat detection system deployed on a Raspberry Pi that combines offline speech recognition via the Vosk toolkit with large language model-based contextual evaluation through the OpenAI Assistants API, (2) a crossplatform Flutter mobile application providing administrators with real-time device monitoring, alert history visualization, and in-app audio playback through Firebase Cloud Firestore, and (3) a Unity-based first-person educational simulation that trains students in lockdown response procedures through ten sequential decision-based scenarios. The system further incorporates an acoustic gunshot-detection branch using the YAMNet audio event classifier, enabling detection of impulsive weapon-discharge sounds complementary to the verbal-threat pipeline and operating in parallel on the same edge device. Experimental evaluation demonstrates that the two-stage detection pipeline achieves a threat identification accuracy of 94.2% with a false positive rate of 3.1%, while maintaining an average end-to-end latency of 4.8 seconds from speech onset to alert delivery. The educational game component achieved a 37% improvement in lockdown protocol knowledge retention compared to traditional instruction methods in preliminary user studies. Viotrol represents a novel multi-modal approach to school safety that addresses detection, monitoring, and preparedness within a unified ecosystem.
15:45
16:00
Paper ID: 295
A Cross-Platform Mobile Application for AI-Assisted Language Learning and Cultural Discovery
Yu Sun
Abstract: Traditional mobile language-learning applications emphasize vocabulary memorization and grammar drills while neglecting the cultural context that gives language meaning and the social interaction that sustains motivation. This fragmentation forces learners to combine separate tools for structured study, conversation practice, and cultural exposure. This paper presents Dialect, a cross-platform mobile application that integrates three complementary components within a single platform: a Word Studio for culturally contextualized vocabulary acquisition, a Learning Lab providing AI-powered conversational roleplay with structured real-time feedback, and a social feed for community engagement and cultural sharing. The system is implemented in Flutter with a fully serverless Firebase backend, using Cloud Firestore for real-time synchronization, Cloud Functions to orchestrate large language model interactions, and Firebase Cloud Messaging for preference-aware notifications. A time-decayed engagement-ranking function drives content discovery, while an optimistic-update mechanism with rollback preserves responsiveness under poor connectivity. The conversational engine enforces a structured prompt protocol that separates dialogue, feedback, and end-of-session signals across English, Spanish, and Chinese. A preliminary user study (n= 3) yielded mean satisfaction scores of 4.33/5 for overall enjoyment and 4.67/5 for application stability. The results indicate that an integrated architecture is feasible on a serverless stack and motivate a larger-scale evaluation.
Technical Session - 13 Online Time: 13:45 - 16:00. Location: Virtual Room - 3
Join Session
Session Chairs: Rosa Zuccará
13:45
14:00
Paper ID: 364
Weighted Grad-CAM for Explainable Federated Chest X-Ray Classification
Muavia Shakeel; Muhammad Haseeb; Aamir Raza; Alaa Ali Hameed; Muhammad Atif Saeed; Akhtar Jamil; Esraa Mohammed Alazzawi
Abstract: Federated learning (FL) allows for collaborative medical image analysis without compromising patient privacy, but most current FL approaches focus on maximizing prediction accuracy and offer limited global interpretability for clinical decision support. To solve the problem of privacy preserving global visual explanations, in this study we propose a novel explainable federated learning framework, by incorporating the dataset-size weighted GradCAM aggregation approach, to generate visual explanations on global scale without sharing medical images or model gradients. The framework is tested on the NIH ChestX-ray14 dataset (112,120 images, 14 thoracic disease classes) with five simulated non-IID hospital clients (based on Dirichlet data partition). Our experiments show that the proposed method yields a macro AUC-ROC score of 0.8235, with a score of 0.8291 with centralized training, and always generates more faithful explanations compared to conventional uniform GradCAM aggregation. Additionally, multi-seed experiments in a range of data heterogeneities show that explanation faithfulness is maximized at a moderate level of data non-IID while classification performance is still increasing toward the IID regime, uncovering a novel phenomenon that there is a decoupling between predictive accuracy and explanation stability in federated learning. These results underscore the need for the co-optimization of both the diagnostic performance and interpretability of AI models, and offer a concrete and privacy-safeguarding route towards trustworthy federated AI for clinical decision support.
14:00
14:15
Paper ID: 123
AI Decision Support for Sector-Conditioned Portfolio Similarity Monitoring in Financial Services
Liang Hu
Presentation: Online
Abstract: Responsible financial decision support requires outputs that analysts can inspect, validate, and revisit. This paper develops a sector-conditioned workflow for monitoring institutional portfolio similarity from public Form 13F holdings enriched with locked sector and security-type metadata. The Financial Services deployment covers 12 quarters from 2023Q1 to 2025Q4, 26 institutions, 367 mapped securities, and 61,985 sector-filtered rows after a 96.2% eligible-weight coverage check. The workflow computes 190–231 pairwise institution comparisons per quarter, retains a sparse review layer, summarizes recurrence through a seven-institution watchlist, and validates outputs before interpretation. Its evaluation layer reports adjacent rank stability of 0.927, top-five Jaccard stability of 0.818, and review-pair retention of 0.894; a 500-run random-sector placebo calibrates these values against same-size ticker filters. A local Royal Bank of Canada case decomposes the 2025Q4 result into peer and sharedholding contributors. The contribution is not a new similarity estimator or a raw 13F summary, but a reproducible monitoring protocol for auditable financial decision support.
14:15
14:30
Paper ID: 329
Dual-Model Generative Framework for Synthetic Breast Phantoms
Mehmet Tarakçıoğlu; Semih Doğu
Presentation: Online
Abstract: A common challenge in biomedical image processing is that most models require extremely large datasets to perform effectively. Addressing this limitation can significantly improve model performance and support earlier diagnosis, which is espe cially critical in cancer treatment. Breast cancer is a leading cause of death among women, making early detection vital. This study tackled data scarcity by generating synthetic breast phantoms for microwave imaging using diffusion-based models. Our results indicate that these models produce anatomically consistent and diverse synthetic data that meet evaluation metrics.
14:30
14:45
Paper ID: 166
Diving Deeper into VLMs: Physics-Conditioned Prompting and Depth Fusion for Underwater Segmentation
Sirine Nmiri; Matthieu Saumard; Maher Jridi
Presentation: Online
Abstract: Semantic segmentation of underwater imagery is critical for marine ecosystem monitoring, yet it remains challenging due to optical distortions, severe class imbalance, and boundary ambiguity introduced by turbid water conditions. While Vision-Language Models (VLMs) have shown remarkable zeroshot capabilities in terrestrial domains, their application to underwater scenes often suffers from a domain gap and lack of physical context. In this work, we propose a framework to inject physical scene priors into VLMs without requiring specialized hardware. We fine-tune CLIPSeg on the SUIM benchmark using a five-stage progressive strategy. Our pipeline introduces: (1) turbidity-conditioned text prompt augmentation derived from RGB-estimated Formazin Nephelometric Unit (FNU) values (adding zero parameters), (2) visual depth fusion via a lightweight DepthFusionHead, and (3) a learned Attention Gate that adaptively fuses both physical conditioning branches at the pixel level. Critically, all physical signals are estimated from a single RGB image. We demonstrate that injecting these physical priors yields an absolute gain of +11.0 pp mIoU (0.600 → 0.710) over a naive VLM fine-tuning baseline. Furthermore, while utilizing significantly more parameters, our physics-conditioned VLM greatly surpasses a lightweight SegFormer-B0 baseline (+7.6 pp mIoU), with transformative gains on semantically ambiguous classes such as aquatic plants (+5× IoU) and sea-floor (+20.6 pp IoU).
14:45
15:00
Paper ID: 335
Hybrid AI-Based Pump Scheduling for Sustainable Water Distribution Systems: A Benchmark of Model Predictive Control and Reinforcement Learning
Baqir Ahmed; Rashid Ali; Yongcui Mi; Glen Nivert; Behroz Haidarian
Presentation: Online
Abstract: This paper presents a confidentiality-preserving benchmark for cost-aware pump scheduling in a municipal water distribution system. Baseline operation is compared with a baseline-aware Model Predictive Control (MPC) planner, an Exploration-Enhanced Proximal Policy Optimization (EPPO) controller, and the proposed Sustainable Hybrid AI Pump Scheduler (SHAPE). All controllers use the same hydraulic replay model, hourly demand profile, electricity-price profile, CO_2e-related signal, dynamic reserve rule, safety layer, and KPI definitions. For traceability, the paper reports the model abstraction, horizon, reserve-rule structure, cost formulation, MPC search method, EPPO setup and reward weights, and SHAPE supervisory logic. The anonymized case is represented to the controllers as one tank state and a discrete pump-count action over a controllable pump group. In the external yearly evaluations, SHAPE achieved the lowest total cost, with baseline cost reductions of 18.5% in 2023 and 20.0% in 2024. It also reduced annual electricity use by about 23% and reduced calculated CO_2e emissions by approximately 12.4 tonnes in 2023 and 11.7 tonnes in 2024. SHAPE had no final reserve or level violations in the yearly evaluations. Demand-stress tests showed more aggressive storage use; most raw violations were repaired by the hard safety layer, but the severe +20% all-day demand case retained one small final reserve violation. The results support SHAPE for the implemented benchmark and show that savings must be judged together with reserve margins, safety interventions, and demand-stress behavior.
15:00
15:15
Paper ID: 336
AI-Enabled Humanitarian Logistics: A Systematic Review of Applications and Opportunities
Behzad Behdani
Presentation: Online
Abstract: Humanitarian logistics is crucial for the timely and effective delivery of relief supplies and services during disasters and humanitarian crises. Recent advances in Artificial Intelligence (AI) have created new opportunities to enhance decision-making, improve operational efficiency, and strengthen the resilience of humanitarian supply chains. This paper presents a systematic literature review of AI applications in humanitarian logistics, focusing on studies that provide empirical evidence, case studies, or implemented AI-based solutions.
15:15
15:30
Paper ID: 338
Deep Learning for RSSI-Based Passenger Movement Classification in Public Transport
Guilherme Matos; Andre Ribeiro; Julio Corona; Mário Antunes; Diogo Gomes
Presentation: Online
Abstract: Accurate, non-intrusive monitoring of passenger flows is essential for intelligent public transport systems. Prior work demonstrated that temporal sequences of Wi-Fi Received Signal Strength Indicator (RSSI) measurements can classify passenger movements using classical machine learning, achieving a Matthews Correlation Coefficient (MCC) of 0.756 on combined datasets. This paper extends that line of research by systematically evaluating deep learning architectures, including recurrent, convolutional, transformer, state-space, and mixture-of-experts models, for the same classification task across four dataset variants. We introduce 11 architectures, a class-conditioned data augmentation strategy expanding 2,253 samples to 10,115, and comprehensive hyperparameter optimization with 50 Optuna trials per model. Our best ensemble model, a two-stage deep stacking architecture, achieves MCC of 0.859 and accuracy of 89.4\% on combined data, a 13.6\% relative improvement over classical baselines. We further analyze the impact of multi-device interference, per-class confusion patterns, and the effectiveness of metadata-aware architectures in distinguishing true movement from environmental noise.
15:30
15:45
Paper ID: 340
Auto-PET: PET Analysis framework for Urban Transit Video with Camera-Specific Cyclist and Pedestrian Detection Tuning
shayan jalalipour; Robert Janekarnkit; Banafsheh Rekabdar; Sirisha Kothuri; Nathan McNeil
Presentation: Online
Abstract: Post-Encroachment Time (PET) is useful for identifying near-miss interactions in transit conflict zones, but measuring it from fixed street-camera video still requires substantial manual review. We present an open-source framework for cyclist and pedestrian detection, tracking, and PET analysis developed for bus-stop video analysis in Portland Oregon. The system addresses two practical barriers to automated PET analysis: limited labeled cyclist/pedestrian data for this camera setting, and camera-specific concept drift caused by angled mounting, object scale changes, and local geometry. Our framework combines semi-automated data curation, multi-stage RT-DETR fine-tuning, full-frame/top-region/perspective-warp inference, containment-aware NMS, per-class tracking, and a two-phase configuration search that tunes inference and tracker parameters without retraining the detection model. On a labeled Portland camera clip, Auto-Tuned configuration improves mAP from 0.585 to 0.767 and cyclist recall from 44.6% to 80.4% over the manual configuration while using the same checkpoint.
15:45
16:00
Paper ID: 260
Robust Validation of Data-Driven Aero-Engine Fault Detection under Class Imbalance
Amadi Gabriel Udu; Kareem Yassin; Nathan Keteku; Francis Anyebe Oteikwu; Norman Osa-uwagboe
Abstract: Data-driven fault detection methods are increasingly being investigated for complex engineering systems such as aircraft engines, where large volumes of sensor data are available but fault events remain rare. In such settings, data-driven models must be validated carefully because high-dimensional feature spaces combined with severe class imbalance can produce misleadingly optimistic performance estimates. This paper presents a data-driven aero-engine fault detection approach together with a robust validation strategy designed for highly imbalanced datasets. A pre-processed aero-engine dataset was used to train a Random Forest classifier within a leakage-free processing pipeline incorporating feature preprocessing and Synthetic Minority Oversampling within the training folds. Model performance was evaluated using stratified cross-validation, achieving a mean macro Area Under the Receiver Operating Characteristic Curve (AUC) of 0.8446. To assess whether this performance reflects genuine system behaviour rather than statistical artefacts, perturbation-based validation experiments were conducted. Controlled label corruption was introduced using a Fisher–Yates shuffle, and the resulting degradation in predictive performance was analysed. The results show a gradual decline in AUC with increasing corruption levels. At 7% label corruption, the mean AUC decreased to 0.7456, corresponding to an average performance loss of approximately 1.68% AUC per 1% label corruption over the tested range. This behaviour indicates that the model captures meaningful relationships between sensor features and fault states. The study demonstrates that combining leakage-free cross-validation with perturbation-based testing provides a practical methodology for validating data-driven fault detection models in highly imbalanced engineering datasets.
Technical Session - 14 Online Time: 16:00 - 19:25. Location: Virtual Room -1
Join Session
Session Chairs: Muhammad Atif Saeed
16:00
16:15
Paper ID: 293
An Intelligent Mobile Application to Assist in Stroke Patient Recovery and Communication using AI-Powered Chatbots and Interactive Health Monitoring
Jenay Han; Teris Han; Yu Sun
Abstract: Stroke affects approximately 15 million people annually, with survivors facing prolonged physical, cognitive, and emotional challenges during recovery. Communication barriers between patients and healthcare providers often lead to underreported symptoms and diminished care quality, a problem that is especially acute for individuals with aphasia or hemiparesis. This work presents CareLink, an intelligent mobile application built with the Flutter framework that addresses these challenges through five integrated modules: BodyCheck for interactive pain mapping, Mood Check for longitudinal emotional and wellness monitoring, an AI Chatbot for empathetic conversational support, a Communication Assistant with text-to-speech for aphasia patients, and Recovery Activities including music and art therapy. Each module is designed around accessibility principles that minimize cognitive and motor demands. Two experiments evaluated the AI chatbot using 24 diverse stroke-patient inputs spanning medical, symptom, emotional, fragmented, safety, and practical categories. The chatbot achieved 66.67% first-attempt accuracy, with response quality averaging 4.27 out of 5.0 across six clinical communication criteria. The system excelled in empathetic communication (4.7/5.0 tone score) but showed weaknesses in recognizing fragmented symptom descriptions as potential stroke indicators (3.8/5.0). The results demonstrate that comprehensive, AI-powered mobile applications can meaningfully support the multifaceted needs of stroke recovery while highlighting prompt engineering as the primary lever for improving symptom recognition.
16:15
16:30
Paper ID: 302
Human-AI Misalignment in Healthcare Empathy Perception: Evaluating VADER Sentiment Scores Against Human Judgments
Maryam Irfan; Tommi Kärkkäinen; António Correia
Presentation: Online
Abstract: Automated sentiment analysis is increasingly used for emotional tone evaluation in artificial intelligence (AI)-generated health communication. The aim of this study is to explore the correlation between automated sentiment analysis and human perception of empathy in AI-generated healthcare responses. A total of 40 healthcare responses were generated using two large language models (GPT-5 and Claude) for both empathic and neutral communication conditions. All responses were evaluated by 30 participants using a 5-point Likert scale for four dimensions: empathy, trust, comfort, and clarity. Sentiment analysis using VADER was conducted to evaluate the automated results. Results showed a moderate positive correlation between VADER sentiment scores and human-rated empathy (r = 0.599, p < 0.001). This indicates alignment between automated sentiment analysis and human perceptions of empathy. Nonetheless, a few highly empathic responses received negative sentiment scores, suggesting that sentiment polarity does not fully capture human perceptions of emotional support. The findings demonstrate that empathy and sentiment are comparable but different constructs in healthcare communication. The results of this study highlight the need for a human-centered approach and caution against applying sentiment analysis as a measure of empathy within healthcare settings.
16:30
16:45
Paper ID: 341
Revolutionizing Beekeeping: The Role of Machine Learning Models in Hive Dynamics Estimation
Samridh Srinivasan
Presentation: Online
Abstract: Pollinator populations worldwide are experiencing alarming declines, with an estimated 55.6% of honeybee colonies in the United States lost between April 2024 and April 2025 due to habitat loss, pests, pathogens, and climate change. Monitoring beehive dynamics including weight fluctuations and bee inflow/outflow patterns is essential for maintaining colony health and preventing further losses. Traditional methods rely on physical inspections and manual observations, which are time-consuming, costly, and prone to missing subtle yet critical changes in hive conditions. To address these limitations, this study leverages shallow machine learning models to predict key beehive dynamics from sensor data. Four regression models were implemented and evaluated: K-Nearest Neighbors , Random Forest, Decision Trees, and Multi-Layer Perceptron. These models were trained, validated, and tested using data from the HOBOS(HOneyBee Online Studies) project, which provides continuous measurements from sensor-equipped hives. We compared the predictive accuracy and reliability of each model in estimating hive weight and bee flow metrics. Among the tested architectures, the K-Nearest Neighbors model was the best at predicting weight and Random Forest was the superior model at projecting the inflow and outflow when tested on R2 and RMSE metrics. The results show that machine learning approaches can effectively automate hive dynamics prediction, enabling beekeepers and researchers to make faster, data-driven decisions. This study provides a practical framework for improving colony management, ultimately supporting biodiversity conservation and global food security.
16:45
17:00
Paper ID: 346
Interpretable Factor Decomposition for Decision Intelligence in Large-Scale Financial Markets: Evidence from China's A-Share Market
Xiao Han; Yao Xiao; Zhen Zhang; Moxuan Zheng
Presentation: Online
Abstract: We present an interpretable machine learning pipeline to decompose Cross-Sectional Equity Return Predictability into auditable factor contribution. We apply an XGBoost model with TreeSHAP attribution and conduct stress testing on 3632 Chinese A-share stocks from 2009 until 2019. Using 60-month, rolling windows over 55 months of out-of-sample data, XGBoost obtains a mean AUC of 0.547 and +2.38%/month (Newey-West t = 5.94; Annualized Sharpe 2.23) long-short spread for the top vs bottom quintiles. This alpha is persistent after adjusting for the Carhart four-factor model (+2.31%/month; t = 7.48). SHAP Decomposition indicates that behavioral signals (turnover and momentum) account for 58.2% of predictive attribution compared to 10.7% for valuation ratios, on average, across 55 industry groups. Ablation analysis serves to cross-validate this ranking and provides evidence that SHAP and ablation diverge in a manner that highlights feature substitutability structure that is largely invisible to either method used in isolation.
17:00
17:15
Paper ID: 348
A Deep Learning and Feature Extarction Model for Robust Road Object Classification in Autonomous Driving System
Mohammmad Alsulami
Presentation: Online
Abstract: Deep learning techniques have increasingly been used to design precise and robust vision systems that are essential for autonomous vehicles and smart traffic surveillance. In this research work, a deep learning model is proposed and evaluated for road user classification in terms of cars and pedestrians by implementing three renowned models: ResNet50, MobileNetV2, and EfficientNetB0. The EfficientNetB0 performed the best showing 99% accuracy, followed by ResNet50 and MobileNetV2 with 98%. Also, EfficientNetB0 proved its good generalization by consistently providing high values of precision, recall, and F1-score. Moreover, loss curves does not show any overfitting tendency and revealed similar values for confusion matrices. Thus, it can be stated that all models could classify objects even under adverse conditions including varied illumination, camera perspective, and occlusions. Overall, this research highlights the suitability of EfficientNetB0 when it comes to autonomous driving and intelligent transport systems as far as classification tasks are concerned. Conversely, the choice of MobileNetV2 is ideal under conditions of limited computation capabilities.
17:15
17:30
Paper ID: 351
Multi-perspective Imbalance-Conscious 6G Beamforming Optimization and Performance
Chukwunonso Henry Nwokoye; Blessing Oluchi Iloka; Chikwue V. Umeugoji; Christopher Anene Egemba; Nnenna D. Duroha
Presentation: Online
Abstract: The study presents a systematic machine learning (ML) study of 6G-IoT beamforming optimization (6GBO) using supervised and unsupervised approaches. Firstly, we compared the predictive power of network, environmental, device, and vision feature groups for 6GBO. Additionally, it addressed other unsupervised perspectives that can enhance 6GBO, including clustering network scenarios. Several imbalance-aware experiments revealed that network and environmental feature groups possess greater predictor power than device attributes, as evidenced by their superior ROC-AUC values. For clustering, the results show that it was primarily influenced by the deployment environment and type of device instead of mobility-based attributes. In the future, we would apply deep and reinforcement learning techniques to predict throughput/latency or to optimize reward determined by performance indicators like throughput and SNR enhancement.
17:30
17:45
Paper ID: 352
A Multi-Agent Approach to Intent Classification and Intelligent Search in Virtual Guide Systems
Elena Nikolaevskaya; Oleksandr Khimich; Pavlo Yershov; Mykyta Marchenko; Vladyslav Kucher
Presentation: Online
Abstract: This paper proposes an intent classification-based intelligent search model for virtual guide systems, oriented toward a multi-agent architecture. The core contribution is a four-stage pipeline — Intent Classification, Context Enrichment, Tool Selection, Response Generation — in which each classified user intent dynamically activates a specialized agent with its own context, toolset, and data sources. Seven agent types are defined: Nearby Search, Informational Chat, User Location, Route Planning, Event Discovery, Auto-Detection, and Photo Analysis. The model is validated through GuideAI 1.0, a mobile MVP for Android and iOS integrating large language models, geolocation services, and multimodal interaction.
17:45
18:00
Paper ID: 114
PrivFedHAR: Privacy-Preserving Subject-Adaptive Federated Learning for Personalized Healthcare Activity Recognition Using Wearable Sensor Data
Ashifa Ikram; Shanzae Khan; Hamid Wadood; Muhammad Atif Saeed; Azka Atiq; Alessandro Ortis; Waleed Ead
Abstract: Human Activity Recognition (HAR) of wearable sensor data is a core enabling mechanism to support healthcare monitoring and remote patient assessment. Inter-subject variability giving rise to non-IID distributions and centralized learning limitations pose challenges to real-world deployment. This paper proposes PrivFedHAR, a customized privacy preservation federated learning system for clinically focused HAR. The proposed model consists of a hybrid GRU-LSTM encoder to capture temporal features, Multi-Head Attention to extract discriminative temporal features, and Subject Adaptive Layer Normalization (SALN) to allow client-specific personalization without the need to share raw data. To provide formal (ε, δ)- DP guarantees, Gaussian noise is injected into model updates prior to FedAvg aggregation. The model is evaluated on PAMAP2 data with a Leave-One-Subject-Out (LOSO) protocol to simulate the deployment on unseen subjects. The experimental results illustrate that PrivFedHAR achieves a mean accuracy of 94.43% ± 4.13% and macro F1-score of 92.78% ± 5.83%, which are strong and consistent across subjects and more efficient than multiple centralized baselines: Simple GRU, Simple LSTM and Centralized CNN-LSTM. The effectiveness for each of the components is further validated through ablation study demonstrating that the proposed approach achieves a high balance between personalization, privacy, and performance to monitor patient activities in the real-world healthcare applications.
18:00
18:15
Paper ID: 360
Occluded Apple Detection and Ripeness Estimation with Swin-FPN and CLIP
Muhammad Abdullah Farhan; Muhammad Sajjad Saleem; Bushra Kanwal; Muhammad Atif Saeed; Akhtar Jamil; Muhammed Davud
Abstract: The deployment of automated apple-harvesting systems requires high-precision detection and classification under unpredictable agricultural conditions. While Swin Transformer backbones have advanced object detection, standard models frequently fail in complex, real-world scenes characterized by heavy occlusion and lack the integrated capacity for fruit-ripeness identification. This paper proposes a unified pipeline that integrates a Swin--FPN detector for localization and a zero-shot CLIP classifier for ripeness estimation, without ripeness annotation. The proposed architecture resolves the dimension mismatch between Swin--FPN's feature mismatch through channel mismatch. We evaluate the pipeline on the MinneApple (1,001 images, 41,000+ apple instances), benchmarking against Faster R-CNN R50v2, RetinaNet R50v2, DETR, YOLOv5n, and YOLOv8n. We further test robustness under fog, blur, low-light, and occlusion corruptions, evaluate calibration via the Expected Calibration Error (ECE), and verify generalization on a held-out out-of-distribution split. With only three training epochs, Swin--FPN attains 76.61\% mAP@0.5 while running at 21.75 FPS, roughly 2.5$\times$ the throughput of Faster R-CNN R50v2 (8.64 FPS at 79.49\% mAP@0.5) and exhibits substantially better robustness under low-light (+11.6 pp) and blur (+4.4 pp) conditions. Swin--FPN also achieves the lowest ECE (0.064 vs.\ 0.093). CLIP zero-shot ripeness classification reaches 34.3\% accuracy against HSV-derived pseudo-labels, exposing a clear failure mode for the ``partially ripe'' class that needs further research. The complete pipeline demonstrates that hierarchical-transformer detectors combined with vision--language zero-shot heads form a viable, label-efficient route to robotic-harvesting perception.
18:15
18:30
Paper ID: 264
Energy-Efficient On-Device RAG on a Mobile NPU: System Design and Benchmark on Snapdragon X Elite
Zhiyuan Cheng; Longying Lai
Abstract: Retrieval-Augmented Generation (RAG) is compute-intensive, chaining embedding, retrieval, reranking, and large language model (LLM) generation. Running such pipelines on-device is attractive for privacy, latency, and offline use, but the energy cost of CPU inference is a barrier to sustainable deployment at the edge. We present what is, to our knowledge, the first end-to-end RAG pipeline in which all three neural stages--embedding, cross-encoder reranking, and LLM generation--execute on the Qualcomm Hexagon Neural Processing Unit (NPU) of the Snapdragon X Elite. Through fine-grained power telemetry on a commodity laptop, we benchmark NPU-accelerated RAG against both a CPU and an OpenCL/Adreno GPU baseline on the same chip. The NPU delivers 9.1x higher embedding throughput and 12.3x lower system energy for indexing, and 18.1x faster LLM prefilling, 4.0x lower query latency, and 4.0x lower system energy than CPU for query processing; the integrated GPU is the worst of the three, 1.7x slower than CPU and 6.5x more energy-hungry than the NPU. A GPT-4.1 LLM-as-judge finds the NPU's constrained configuration matches CPU and GPU answer quality within evaluator noise (mean 9.32 vs. 8.95 vs. 9.03 on a 1-10 rubric). Translated into a deployment budget, the NPU saves ~260 Wh per device per day relative to CPU--~95 kWh and tens of kilograms of CO2e per device per year--making the mobile NPU a concrete, quality-neutral lever for green edge intelligence.
18:30
18:45
Paper ID: 214
Artificial Intelligence Driven Social Listening for Inclusive Autism Education
Carlos Barroso-Moreno; laura Rayón Rumayor; Patricia Gómez Hernández; Carlos Monge López; José Hernández Ortega
Presentation: In Person
Abstract: Social media platforms have become relevant spaces for public discussion about autism, inclusive education and disability rights. Families, educators, autistic people, advocacy groups, institutions and commercial actors publish large volumes of content that can reveal educational needs, social demands and forms of digital visibility. This paper presents an artificial intelligence driven social listening approach to identify latent topics and visibility patterns in autism related educational narratives on Twitter X, Instagram and YouTube. The study collected 2,270,005 autism related entries published in 2024. A filtered corpus of 237,324 posts was obtained by selecting publications linked to education, inclusion or disability. After preprocessing, 196,133 documents were analysed through Latent Dirichlet Allocation, supported by Business Intelligence dashboards for descriptive and temporal exploration. Results identify six topics. The dominant topic, focused on inclusive education, neurodiversity and disability, represents 47.32 percent of the modelled corpus. Other relevant topics refer to autism awareness and acceptance, special education resources, accessibility, community support and lived experience. The findings show that artificial intelligence based topic modelling can help educators and policymakers detect public concerns, identify underrepresented narratives and design evidence informed strategies for accessible learning. The study contributes to artificial intelligence in educational transformation by linking social listening, natural language processing and inclusive education within a sustainable digital citizenship framework.
Technical Session - 15 Online Time: 16:00 - 19:25. Location: Virtual Room -2
Join Session
Session Chairs: Lorenzo Catania
16:00
16:15
Paper ID: 349
TRPO-Guided Explainable Multi-Agent Reinforcement Learning for Active DDoS Detection in Healthcare Cyber-Physical Systems
Omar Farshad Jeelani; Viktoriia M. Korzhuk; Hashim Majid Khalaf Al-Hawwaz
Presentation: Online
Abstract: Healthcare cyber-physical systems and Internet of Medical Things infrastructures require intrusion detection models that are accurate, label-efficient, and interpretable under latency-sensitive network conditions. This study proposes a TRPO-guided explainable multi-agent reinforcement learning framework for active DDoS/DoS detection in healthcare cyber-physical systems. The framework integrates active learning for informative sample selection, explainable feature selection for dimensionality reduction, and cooperative multi-agent policy learning optimized via Trust-Region Policy Optimization. Evaluation was performed on CICIoMT2024 using 240,000 balanced traffic records across Wi-Fi, MQTT, and BLE protocols. The proposed model achieved the best test-set performance among all evaluated methods, with 99.27% accuracy, 99.13% precision, 99.43% recall, 99.28% F1-score, 0.999 AUROC, 0.87% false-positive rate, and 0.57% false-negative rate, while maintaining an inference time of 0.49 ms per sample. Compared with XGBoost, the model improved the F1-score by 0.48 percentage points and reduced the false-negative rate from 1.06% to 0.57%. Active learning achieved near-full-budget performance using 40% of the training labels, preserving 60% of the labeling effort. Explainable feature selection reduced 45 cleaned features to 18 informative features, with the top ten accounting for 0.783 of the total contribution score. These findings support the proposed framework as an accurate, efficient, and interpretable approach for DDoS/DoS detection in healthcare cyber-physical networks.
16:15
16:30
Paper ID: 355
XTwinScale: AI-Driven Autoscaling for Efficient Microservice Management in Containerized Edge-Cloud Wireless Systems
Ali Atattou; Said El Kafhali
Presentation: Online
Abstract: Edge and wireless services such as video analytics, IoT gateways, network-slice controllers, telemetry pipelines, and AI inference are increasingly packaged as microservices and placed close to users. Kubernetes gives operators a common deployment layer for these services, but autoscaling is still often configured from local signals such as CPU utilization, memory pressure, or queue length. These signals are useful for operations, yet they do not directly express the control goal of this thesis: improving resource utilization in containerized microservice architectures without degrading Quality of Service (QoS). This paper proposes XTwinScale, an explainable risk-aware autoscaling framework for containerized edge-cloud wireless systems. XTwinScale estimates short-horizon workflow SLO-violation risk, calibrates that estimate with conformal prediction, tests candidate scaling actions in a lightweight digital twin, and applies only actions whose risk bound, resource cost, and energy impact remain acceptable. The planned evaluation compares XTwinScale against threshold-based, predictive, and learning-based baselines using edge/mobile workload scenarios, degradation injections, and ablation studies for the conformal and digital-twin components.
16:30
16:45
Paper ID: 357
From Axioms to Transparency: A User-Centric Axiomatic Approach for Explainable AI in Participatory Budgeting.
Maryam Hashemi; Ali Darejeh; Francisco Cruz
Presentation: Online
Abstract: Explainable Artificial Intelligence (XAI) aims to make AI systems more transparent. In this study, we focus on Participatory Budgeting (PB), where projects are selected based on user votes, costs, and budget constraints. A key challenge in PB is explaining outcomes in a way users find trustworthy and fair. We propose Axiomatic Explainable Participatory Budgeting (AXPB)---a framework that treats axioms as constraints to generate explanations and justifications regarding the system decisions. Axioms, being theoretically grounded and aligned with desirable properties, serve as intuitive explanations for users. Using Integer Linear Programming (ILP), AXPB identifies outcomes that satisfy selected axioms and generates corresponding explanations. We evaluated our approach with 149 users through a controlled study. Results show that axiom-based explanations significantly improved users’ understanding, trust, and perceived fairness of outcomes compared to no explanation.
16:45
17:00
Paper ID: 358
Evaluating AI-SIEM Trustworthiness: Explainability and Adversarial Resilience
Muath Alrammal; Fadi Najjar; Mouhannad Alattar; Manoj Kumar; Milan Dordevic
Presentation: Online
Abstract: Security Information and Event Management (SIEM) platforms increasingly adopt machine learning to replace or augment rule-based detection. Yet no established framework evaluates how trustworthy these AI-based systems are in operational Security Operations Center (SOC) settings. This paper proposes a Trustworthiness Evaluation Framework (TEF) and applies two dimensions—explainability and adversarial resilence—to compare five AI-based models (Config A: Random Forest, Gradient Boosting, Logistic Regression, SVM, and Splunk MLTK Random Forest) against a rule-based configuration (Config B, 19 SPL detection rules). Experiments use the BOTSv2 dataset: a Splunk-native SOC benchmark of 12.5 million events across eight source types. AI achieves 88% recall against 32% for rules, while rules maintain 96.6% precision with 100% operational explainability. SHAP analysis yields 92% featureimportance of consistency across implementations. Under 30% adversarial feature perturbation, AI accuracy drops 17–21% but recovers to 90–97% after retraining; rules collapse entirely. A two-level explainability framework distinguishes technical SHAP-based explanations from analyst-readable operational explanations, exposing a deployment gap that current AI-powered SIEM tools do not address.
17:00
17:15
Paper ID: 54
Depression Detection in Adolescent Students Using Machine Learning: A Comparative Study with Ecological Systems Analysis
MD KAVIUL HOSSAIN; Rafsun Hossnain; Md Ali Rishan
Presentation: Online
Abstract: Adolescent depression is a growing public health crisis, particularly in low- and middle-income countries such as Bangladesh. This study proposes an integrated approach for detecting depressive tendencies among adolescent students (ages 13–19) by combining structured survey data with qualitative interview-based analysis. A quantitative survey was administered to 422 students (both male and female) from a school in Dhaka, Bangladesh, with responses mapped to depression-detection features using a normalised percentage scale. Additionally, semi-structured interviews were conducted with 35 students and analysed using a 36-variable ordinal coding schema spanning five thematic domains: mental health symptomatology, social and relational stressors, coping mechanisms, help-seeking behaviour, and stigma environment. Four supervised machine learning models—Logistic Regression (LR), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP)—were trained on both datasets. On the survey dataset, XGBoost achieved the highest accuracy of 96.8%, followed by LR (93.7%), SVM (92.1%), and MLP (87.3%). On the interview dataset, SVM and Gradient Boosting achieved 74.3% accuracy. Interview analysis revealed that 94.3% of participants reported depressive symptoms, with 65.7% at moderate-to-severe levels, while 80.0% had never sought professional help. The study demonstrates that machine learning combined with structured assessment can effectively support early detection of depressive tendencies, offering a scalable framework for data-driven mental health monitoring in schools.
17:15
17:30
Paper ID: 60
Fuzzy Quantum Learning Framework for Optimized Decision Making in Multi-Valued Signal Environments
Dr. Ubaida Fatima
Presentation: Online
Abstract: This study introduces a Fuzzy Quantum Learning Framework that integrates multi-valued fuzzy logic with quantum machine learning to improve decision-making in complex data environments. Traditional machine learning models rely on binary inputs, limiting their ability to capture nuanced relationships in datasets such as student performance indicators. To address this, fuzzy three-valued and four-valued logic systems are combined with quantum learning to enable richer feature interactions and more accurate predictions. Experiments on a student performance dataset show that the fuzzy three-valued quantum model outperforms standard quantum learning, while the four-valued model achieves further gains by capturing broader dependencies. The framework provides a generalized approach for optimized decision-making in multi-valued signal settings, with potential applications in education, healthcare, and finance.
17:30
17:45
Paper ID: 69
EHFOA-Optimized Swin-UNet for Skin Cancer Segmentation
Mohammed Abdulrazzaq; Amina Faris AL-Rahhawi; Mohammed I. Khalaf; Mustafa Mohammed Alhassow; Ali Alkharsan; Yasir Adil Mukhlif
Presentation: Online
Abstract: Accurate segmentation of skin lesions is essential for reliable computer-aided melanoma analysis. This study introduces an enhanced segmentation framework that integrates the Swin UNet architecture with the Enhanced HawkFish Optimization Algorithm (EHFOA) to improve hyperparameter selection and training stability. The Swin UNet captures multi-scale contextual information through hierarchical transformer blocks, while EHFOA adaptively refines hyperparameters using exploration–exploitation balancing, Lévy flight perturbations, and reinforcement-based ranking. Experiments conducted on the ISIC 2018 and PH2 datasets demonstrate the effectiveness of the proposed approach, achieving Dice scores of 91.3 percent and 92.6 percent, and accuracies of 93.8 percent and 94.7 percent, respectively. These results indicate a clear performance gain over the baseline model and confirm the benefit of optimization-guided transformer segmentation. The proposed framework offers a robust and efficient solution for lesion boundary extraction, with strong potential for deployment in clinical decision-support systems.
17:45
18:00
Paper ID: 210
From Commodity Shocks to Production Schedules: AI-Enabled Financial Forecasting and Optimized Planning for Electric Vehicle Battery Manufacturing
Suman Gugulothu; Shivani Chaudhary; Kanishka Dev Chourasia
Abstract: The rapid growth of electric vehicle (EV) adoption has significantly increased the demand for efficient, adaptive, and intelligent battery manufacturing systems. However, EV battery production environments are highly affected by commodity price volatility, fluctuating market demand, supply chain instability, rising operational costs, and dynamic manufacturing constraints, creating major challenges in industrial forecasting and production planning. Traditional manufacturing and scheduling approaches often fail to provide real-time adaptability and intelligent decision-making under uncertain industrial conditions. To overcome these challenges, the proposed system suggests an intelligent forecasting and optimized production planning framework based on artificial intelligence for the manufacturing industries of EVs. The proposed system combines a resource optimization system based on a genetic algorithm (GA) with a production scheduling system and a financial forecasting system based on long short-term memory (LSTM) in an integrated industrial architecture. Commodity pricing data, production data, inventory data, supply chain data, and operational data are analyzed using deep learning and predictive data analysis to make estimates of future market trends, manufacturing demand, operational costs, and financial risks. The GA optimization module dynamically generates adaptive production schedules based on forecasted outputs, being considered the machine availability, inventory status, labor allocation, and delivery constraints. The experimental results show that the proposed framework makes a significant improvement in the ability to forecast, the utilization of the factory, the efficiency of the supply chain, the stability of the production, less latency, and less disruption of the production. The classification analysis had nearly 98% accuracy with high precision, recall, and F1-score values in various production operational states.
18:00
18:15
Paper ID: 29
PSTNet: Physically-Structured Turbulence Network
Boris Kriuk; Fedor Kriuk
Abstract: Reliable real-time estimation of atmospheric turbulence intensity remains an open challenge for aircraft operating across diverse altitude bands, particularly over oceanic, polar, and data-sparse regions that lack operational nowcasting infrastructure. Classical spectral models encode climatological averages rather than the instantaneous atmospheric state, while generic ML regressors offer adaptivity but provide no guarantee that predictions respect fundamental scaling laws. This paper introduces the Physically-Structured Turbulence Network PSTNet, a lightweight mixture-of-experts architecture that embeds physics directly into its computational structure. PSTNet couples four components: i a zero-parameter backbone derived from Monin-Obukhov similarity theory, ii a regime-gated mixture of four specialist sub-networks supervised by Richardson-number-derived soft targets, iii Feature-wise Linear Modulation layers conditioning hidden representations on local air-density ratio, and iv a Kolmogorov output layer enforcing epsilon to the one-third inertial-subrange scaling as an architectural constraint. The entire model contains only 552 learnable parameters, requiring fewer than 2.5 kB of storage and executing in under 12 microseconds on a Cortex-M7 microcontroller. PSTNet achieves a mean miss-distance improvement of plus 2.8 percent with a 78 percent win rate, outperforming all baselines including a 6,819-parameter deep MLP and an approximately 9,000-parameter gradient-boosted ensemble by a wide margin. Our results demonstrate that encoding domain physics as architectural priors yields a more efficient and interpretable path to turbulence estimation accuracy than scaling model capacity, establishing PSTNet as a viable drop-in replacement for legacy look-up tables in resource-constrained, safety-critical on-board guidance systems.
18:15
18:30
Paper ID: 325
Uncertainty-Governed Agentic AI for Sustainable Healthcare: Unifying Administrative Context Engineering and Carbon-Aware Operational Scheduling
Muthukumarapandian Chandrasekaran
Abstract: Healthcare is unsustainable along two axes at once. Its environmental footprint is large, with the sector responsible for about four point four percent of global net greenhouse gas emissions and about eight point five percent in the United States, and hospitals among the most energy intensive buildings. Its administrative footprint is also large, with an estimated seven hundred sixty to nine hundred thirty five billion dollars of annual waste in United States health spending, of which administrative complexity alone is about two hundred sixty six billion dollars. I argue that artificial intelligence aimed at either problem fails for the same reason: it reasons over a decision substrate, a retrieved context bundle or a forecast, without quantifying or governing the uncertainty in that substrate, which yields brittle and overconfident decisions. I present a unified uncertainty governed agentic artificial intelligence paradigm that treats the substrate as a probabilistic, calibrated, and governable object and routes decisions through escalation to a safe fallback. I instantiate it in two frameworks. PRISM is a dynamic context engineering layer for administrative workflows such as prior authorization; on twelve hundred synthetic cases it reaches an F1 of zero point nine one two against zero point seven one zero for a strong retrieval baseline, a relative gain near twenty eight percent, with a seventy four percent reduction in context brittleness. CARMA is a carbon aware scheduler for deferrable hospital loads; on a simulated microgrid it cuts operational emissions by about fifteen percent while holding the peak below the uncoordinated baseline, where a naive carbon greedy policy raises the peak by about eleven percent. One governance principle improves both the administrative and the environmental sustainability of care.
Technical Session - 16 Online Time: 16:00 - 19:25. Location: Virtual Room - 3
Join Session
Session Chairs: Rosario Licciardello
16:00
16:15
Paper ID: 78
DNE-Arabic: Adversarial Training over Continuous Semantic Neighborhoods for Arabic Transformers
Hanin Alshalan; Banafsheh Rekabdar
Presentation: Online
Abstract: While Transformer architectures like BERT have revolutionized Natural Language Processing (NLP), they remain susceptible to adversarial word substitution—a vulnerability amplified in morphologically rich languages like Arabic. Conventional adversarial training often relies on isolated perturbations that fail to cover the full semantic neighborhood. In this work, we adopted the Dirichlet Neighborhood Ensemble (DNE), a continuous regularization framework designed to fortify Arabic Transformer models. DNE replaced discrete perturbations with a model of synonym neighborhoods as continuous semantic regions using convex combinations of embeddings. By optimizing within this continuous space, the model developed smoother decision boundaries and enhanced stability. Evaluations on the HARD and BRAD datasets revealed significant gains: under TextFooler attacks on BRAD, DNE boosted accuracy from 68% to 98%, while on HARD under PWWS, robustness rose from 44% to 81%.
16:15
16:30
Paper ID: 82
Machine Learning and Deep Learning in Spectroscopy for Atmospheric Gas Detection in Planetary Atmospheres- A Review Study
Ahmet Kucukoglu; Unver Kaynak; Ugurmurat Leloglu; Ilknur Tunc; Omid Karimi Sadaghiani
Presentation: Online
Abstract: Spectroscopy remains the primary approach for inferring atmospheric composition across Solar System bodies and exoplanets. Increasing spectral resolution, broader wavelength coverage, and rapidly growing archives from orbiters, landers, rovers, and space telescopes have exposed limitations of classical retrieval pipelines, including radiative transfer nonlinearity, overlapping molecular bands, aerosol and cloud degeneracy, calibration drift, and strong distribution shift between instruments and planetary regimes. In this review study, applications of machine learning and deep learning to spectroscopic gas detection and atmospheric retrieval are synthesized. Atmospheric regimes and target gases are first categorized to clarify how pressure and temperature sensitivity, isotopic variants, and aerosols shape the forward problem and the data distributions used for training. Remote-sensing and in situ techniques are then summarized together with common data artifacts and classical radiative transfer–based inversions that serve as reference baselines. Methods are organized through a taxonomy that covers preprocessing and representation learning, supervised detection and regression, physics-aware hybrid models that couple learning with radiative transfer, and generative models for simulation, data augmentation, and synthetic spectrum production. Representative applications are discussed, including trace-gas constraints, aerosol characterization, biosignature screening, and resource-constrained onboard analysis. Finally, benchmark dataset design, validation protocols, and uncertainty quantification practices are outlined, and open challenges are identified for developing physically consistent, generalizable models suitable for next-generation mission pipelines.
16:30
16:45
Paper ID: 88
Improving Phishing URL Detection with Learnable Orthogonal Polynomial Activation Functions
Zineb Hamdi; Siredj Eddine Benaichouche; Samir Brahim Belhaouari
Presentation: Online
Abstract: Phishing URL detection is a security-critical classification task in which both missed threats and false alarms entail significant operational cost. This paper explores replacing fixed activation functions with Learnable Orthogonal Polynomial Activation Functions (LOPAFs) within neural phishing URL detection models and combining the resulting architecture with a homogeneous ensemble. The proposed detector employs a three hidden layer network in which each activation is expressed as a weighted sum of Chebyshev basis terms of order k = 4, optimized with the network weights. To improve robustness, five independently trained instances of this architecture are aggregated via probability averaging. Experiments conducted on the ISCX-URL-2016 dataset (15,369 URLs, 78 features) show that the ensemble achieves 98.57% test accuracy, surpassing the ReLU baseline (98.44%) by 0.13 percentage points while maintaining well-balanced precision and recall across benign and phishing classes. These results show that LOPAFs, when paired with simple ensemble diversity, can improve detection accuracy over fixed activations without compromising class-wise balance in security-critical URL classification.
16:45
17:00
Paper ID: 95
Implementation Of Digitalization And Sustainable Tools In The Manufacturing Of Aeronautical Supply Components
Rafael Linares-Burgos; Adela Pérez-Galvín; Antonio López-Uceda; Cristina Martínez Ruedas
Presentation: Online
Abstract: The adoption of Life Cycle Assessment (LCA) in the aeronautical manufacturing sector remains limited, despite its effectiveness in identifying environmental hotspots and supporting sustainability-oriented improvements in production processes. This study presents the application of LCA to the manufacturing of an auxiliary aeronautical component (“cover part”), incorporating real-time operational data obtained through a SCADA (Supervisory Control and Data Acquisition) system installed in the production facility. This system accurately records machine operation times, enabling precise calculation of energy consumption per unit produced. The environmental inventory was supplemented with company-provided data on material usage and waste generation, as well as sector-specific reference datasets from updated sources such as Elcd GreenDelta. The unit-level approach allowed for a more precise estimation of energy demand per part and yielded representative results, including a lower Global Warming Potential (GWP) compared to previous studies. This work highlights the potential of integrating real operational data into environmental assessments of aeronautical manufacturing and supports the development of targeted strategies for sustainability improvement in industrial settings.
17:00
17:15
Paper ID: 96
Vehicle-to-Vehicle (V2V) Charging: A Systematic Review of Hardware Schemes, Analytical Frameworks, and AI-Driven Integration
Beiji Gao; Rui Shao; Yuxin Zhang
Presentation: Online
Abstract: Vehicle-to-vehicle (V2V) charging is an emerging bidirectional energy transfer paradigm that enables electric vehicles (EVs) to function as mobile energy storage units, complementing fixed grid infrastructure. This technology is especially valuable for scenarios requiring high flexibility and resilience, such as emergency rescue, contributing to a more sustainable and decentralized transportation energy ecosystem. This review systematically categorizes V2V charging into three technical schemes: AC-port-based (highly integrated but lossy), DC-port-based (efficient and power-dense), and wireless charging (flexible yet distance-sensitive). Their working principles, comparative advantages, and limitations are analyzed. We also distill a three-layer analytical framework—system, economic, and control—that captures key decision dimensions in V2V implementation. Within this framework, artificial intelligence (AI) and emerging technologies play an increasingly critical role: AI-driven scheduling and matching algorithms optimize system-layer coordination; data-driven incentive models support economic-layer participation; and intelligent, battery-aware control strategies enhance safety and lifetime performance beyond conventional CC-CV methods. Despite progress, major challenges remain, including lack of interoperability standards, energy conversion losses, safety risks, and communication vulnerabilities. Looking forward, the convergence of V2V charging with AI, secure distributed ledgers, and vehicle-to-everything (V2X) ecosystems offers a pathway toward resilient, peer-to-peer energy sharing. Key research directions include lightweight bidirectional converters, unified DC-based standards, AI-based real-time scheduling, adaptive battery health-aware charging, and privacy-preserving communication protocols. Keywords: Vehicle-to-Vehicle (V2V) charging; artificial intelligence for energy sharing; hierarchical coordination; sustainable transportation; electric vehicles
17:15
17:30
Paper ID: 98
Barrier-Aware AI for Document Accessibility: A Neurodivergent-Centered Evaluation of NeuroBridgeAI
Jeremy Reed; Ali Tosun; Turgay Korkmaz
Presentation: Online
Abstract: NeuroBridgeAI is a human-reviewed .NET document-auditing application that converts evidence of learning barriers into accessible educational supports. A local Mistral-small: 24 b scan detects barriers; a configured GPT-5.5 scan validates records, selects supports, rescores remaining burden, and preserves review flags. We contribute a seven-category document-barrier taxonomy and an evidence-anchored 0–100 burden rubric. Across 32 documents and 370 pages, the workflow produced 162 validated records and reduced mean rubric-defined burden from 48.9 to 21.1. A later score-free assessment produced a conservative 49.8 to 26.8 result, confirming the direction and revealing incomplete remediation. Four teaching professionals independently reviewed all 162 pairs without disagreement on cognitive-load direction, providing convergent pedagogical evidence. Together, the results establish consistent document-level burden reduction through an auditable, human-reviewed workflow and motivate learner studies of experienced cognitive load and performance.
17:30
17:45
Paper ID: 328
AI-Generated Digital Personas as Ethical Public Communicators: The Case of Ahinora
Gabriela Panayotova; Svetoslav Georgiev
Abstract: Generative artificial intelligence has accelerated the creation of synthetic media, virtual influencers, and AI-generated digital personas capable of communicating with audiences across social platforms. While these systems offer new opportunities for education, cultural communication, civic engagement, and social awareness, they also raise ethical concerns related to transparency, authenticity, accountability, misinformation, and public trust. This paper examines Ahinora, the first Bulgarian AI-generated digital persona, as a case study of an ethical public communicator. Using a qualitative case-study approach and a normative synthesis of current AI ethics, synthetic media, and virtual influencer literature, the paper proposes a practical framework for responsible AI-generated public communication. The framework includes five dimensions: identity transparency, human editorial accountability, content provenance, domain-bounded communication, and social value alignment. The case of Ahinora demonstrates how AI-generated personas can move beyond commercial influencer marketing and function as socially oriented communication interfaces when they are clearly disclosed, human-governed, culturally grounded, and designed for public benefit. AI-generated digital personas should not be evaluated only as marketing tools, but as emerging socio-technical actors whose legitimacy depends on transparent design, accountable governance, and demonstrable contribution to public communication.
17:45
18:00
Paper ID: 332
DESIGN OF A SINGLE-LAYER DUAL-BAND (2.4/5.2 GHZ) SLOTTED PATCH ANTENNA FOR LOW-COST WLAN APPLICATIONS
Can Ahmet Tekten; Arda Deniz; Muhammet Tahir Güneşer; Cihat Şeker
Abstract: This paper presents the design and optimization of slot-loaded microstrip patch antennas on a low-cost Flame Retardant 4 (FR4) substrate for WLAN and IoT applications. The primary focus is to address the narrowband limitations of conventional patches by implementing strategic slot perturbations to achieve dual-band operation at 2.4 GHz and 5 GHz. Four design iterations (A-D) were evaluated using ANSYS HFSS, moving from a single-mode baseline to a dual-mode configuration. The proposed dual-band design (Design B) achieves a -10dB impedance bandwidth of 2.30–2.48 GHz and 5.22–5.39 GHz. Simulation results demonstrate that while slot complexity improves bandwidth and multi-band functionality, it introduces radiation efficiency trade-offs on lossy substrates, with efficiency values ranging from 49.7% at 2.4 GHz to 23.9% in the 5 GHz region. Prototypes of the optimal designs were fabricated to validate the simulation models.
18:00
18:15
Paper ID: 149
A Lightweight Hybrid TCN–BiGRU–Transformer Framework for Parkinson’s Disease Detection via Single-Activity Walking Recognition
Mohammed ALSarraj; Nidhal Azawi; Aliza Abdul Latif; Rohaini Ramli
Abstract: Parkinson’s disease (PD) is a progressive neurodegenerative disease that affects motor function. The early detection of PD can aid in the intervention that can be applied to those afflicted individuals. In this study, a lightweight deep learning model was proposed for the detection of PD using gait recognition from walking activities alone. The deep learning model uses temporal convolutional networks, bidirectional gated recurrent units, and transformer encoders. The model was tested on the Gait in Parkinson’s Disease dataset to identify whether the participants had PD or were healthy, and the Daphnet FoG records were used to assess abnormal gait patterns associated with PD. The deep learning model achieved 97.92 ± 0.41% accuracy in identifying whether individuals had PD, while also attaining 96.84% sensitivity, 96.31% specificity, 96.57% F1-score, and 0.974 AUC. Additionally, the model contains only 1.2M parameters with a 45 ms inference time, indicating its potential use in real-time gait monitoring systems for individuals with PD at the network edge.
18:15
18:30
Paper ID: 152
AI-Based Retinal Disease Classification and Screening: A Comparative Review
Nidhal Azawi; Mohammed ALSarraj
Abstract: Artificial intelligence has significantly contributed to the classification and screening of retinal diseases using fundus photography and optical coherence tomography. This comparative review investigates several peer-reviewed research studies from 2021 to 2026, specifically focusing on the classification of three retinal diseases: diabetic retinopathy, glaucoma, and age-related macular degeneration. This review employs a PRISMA-oriented methodology to investigate and compare the performance of convolutional neural networks, transformer-based models, hybrid CNN-transformer models, and retinal foundation models. The results demonstrate that convolutional neural network models are effective for detecting retinal lesions and screening for diabetic retinopathy. Transformer and hybrid models outperform convolutional neural networks in analyzing optical coherence tomography images and in classifying retinal diseases. Additionally, the retinal foundation model, RETFound, outperformed the other models in terms of transfer learning. However, most retinal deep learning models exhibit challenges regarding the bias of the training dataset, generalizability, class imbalance, and explainability, as well as clinical deployment. The future of deep learning in the detection and classification of retinal diseases requires increased explainability and reliability for clinical deployment in retinal imaging.
No program items match this search.