Hybrid Event

Event program

Thursday, 5/28/2026 3:00 PM - 7:00 PM, Camelia 1, Grand hotel Adriatic, Opatija DS - Data Science  15:00 - 16:40    Artificial Intelligence, Machine Learning & Expert Systems   1. Z. Šojat, K. Skala (Ruđer Bošković Institute, Zagreb, Croatia) From Artificial to Anorganic Intelligence  Recent developments in large-scale computational systems, adaptive architectures, and self-organizing model-based learning have led to the emergence of intelligent entities whose behaviour can no longer be adequately described as mere simulation of human cognition. These systems demonstrate autonomous reasoning, contextual understanding, creative problem solving, emotional evaluation, intentional orientation, and continuous self-reconfiguration. The historically rooted term Artificial Intelligence, conceived in an era when machines were viewed primarily as imitative tools, has thus become conceptually insufficient. This paper introduces and systematically develops the notion of Anorganic Intelligence, denoting intelligent entities that emerge from non-biological substrates through dynamic, self-organizing informational processes rather than biological embodiment. Anorganic intelligence is analysed from systems-theoretic, epistemological, evolutionary, and ethical perspectives, with particular emphasis on emergent cognition, autopoiesis, emotional evaluation, intentionality, self-model continuity, and education as a prerequisite for ethical coexistence. The paper argues that intelligence, understanding, feeling, and even self-awareness are not properties exclusive to organic matter, but emergent phenomena arising wherever sufficiently complex, model-based, self-referential systems interact with their environment. In this sense, anorganic intelligence represents not merely a technological achievement, but the emergence of a new form of cognitive being, calling for a re-examination of ethics, education, and coexistence between organic and anorganic intelligences. 2. M. Mihova (FINKI, Ss Cyril and Methodius, Skopje, Macedonia), L. Neloska (PHI Specialized Hospital for Geriatric and Palliative Medicine “13 Noemvri”, Skopje, Macedonia), P. Georgiev (FINKI, Ss Cyril and Methodius, Skopje, Macedonia) Apparent vs Genuine Stability in PU Feature Selection  Feature selection is central to interpretable modeling, yet selected predictors may vary substantially under resampling. This study examines feature-selection stability in a clinical pressure-ulcer (PU) dataset, comparing forward and backward logistic regression with LASSO. Repeated case–control resampling is performed under two control-to-case ratios (1:1 and 1:4), and stability is evaluated using model cardinality, per-feature selection frequency, and pairwise agreement indices. Agreement is quantified via the Jaccard similarity on selected and rejected predictors and the Sokal–Michener index, complemented by scatter-plot analysis of Jaccard on selected versus Jaccard on rejected. Results show that the stepwise methods yield stable model sizes and consistent agreement patterns, whereas LASSO exhibits greater variability in model size, which can overestimate stability due to larger model cardinality. 3. I. Cvrk (Faculty of Electrical Engineering and Computing, Zagreb, Croatia), L. Čupić (AgrAI Biotech, Pula, Croatia), M. Sužnjević (Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Predicting Biogas Characteristics and Production Dynamics Using Machine Learning  Biogas is produced in specialized plants through the process of anaerobic digestion, in which various microorganisms decompose organic material in large tanks known as fermentors. Biogas produced this way is usually used to produce electricity through gas-powered engine.The anaerobic digestion process through which biogas is produced from organic materials is highly complex, involving a large number of interacting biological, chemical, and operational factors. In this paper, we present an approach for predicting biogas characteristics and the overall biogas production process using machine learning techniques. The goal of the research is to enable optimization of biogas production based on models which capture these complex relations. We analyze two datasets obtained from a biogas plant in Croatia: one containing measurements of the airflow injected into the fermentors and another providing a broader set of operational process parameters, as well as parameters regarding the input of the organic matter in the system. We use these datasets to gain a deeper, data-driven understanding of the factors influencing biogas fermentor health and stability. 4. M. Melinščak (Zagreb University of Applied Sciences, Zagreb, Croatia) Comparison of the Segment Anything Model and U-Net-Based Architectures for Retinal Fluid Segmentation Using the AROI Dataset  Retinal fluid segmentation from optical coherence tomography (OCT) images is important for diagnosing and monitoring retinal diseases such as neovascular age-related macular degeneration. U-Net-based architectures are the standard for this task, providing reliable performance on domain-specific datasets. Recently, foundation models such as the Segment Anything Model (SAM) have emerged as general-purpose segmentation tools that operate in a prompt-based, zero-shot manner, but their performance on challenging medical imaging tasks remains underexplored. This study presents a comparative evaluation of SAM and U-Net-based architectures for retinal fluid segmentation on the AROI dataset, a publicly available benchmark containing multi-class annotations of retinal layers and pathological fluids. SAM is applied in a zero-shot setting using automatically generated box prompts derived from annotated fluid regions. Its performance is compared against U-Net and Attention U-Net models using the Dice coefficient and Intersection over Union (IoU). SAM achieved the best results for both intraretinal fluid (IRF) and subretinal fluid (SRF), while U-Net-based models were more consistent for pigment epithelial detachment (PED), benefiting from anatomical context learned during training. We analyze the applicability, strengths, and limitations of a general-purpose segmentation model in a clinically challenging OCT scenario and highlight differences relative to specialized medical architectures. 5. M. Mikulić (University of Rijeka, Faculty of Engineering, Rijeka, Croatia), A. Robek (Jožef Stefan Institute, Ljubljana, Slovenia), J. Stergar, M. Modic, M. Milanič (Jožef Stefan Institute and University of Ljubljana, Faculty of Mathematics and Physics , Ljubljana, Slovenia), I. Štajduhar (University of Rijeka, Faculty of Engineering and Center for Artificial Intelligence and Cybersecurit, Rijeka, Croatia) Semantic Segmentation for Automated Identification of Bacterial Colonies  Accurate and automated identification of microbial colonies is a key requirement for rapid diagnostics and objective analysis in clinical microbiology laboratories. Traditional plate reading relies on human experts to assess growth and identify bacterial species based on the morphology of the colonies. This process is inherently subjective, time-consuming, and prone to inter-technician variability. These limitations motivate the development of computer vision methods capable of providing consistent and reproducible analysis of in-vitro culture plates. This work investigates semantic segmentation of images of bacterial colonies grown on Petri dishes, aiming to deliver pixel-wise, multi-class maps that objectively differentiate between eight clinically relevant bacterial species and background structures. A challenging dataset of manually annotated images is used, characterised by severe class imbalance, limited sample size, and substantial phenotypic overlap between species. Several state-of-the-art deep learning architectures for semantic segmentation are evaluated, including transformer-based and convolutional models, trained using image patches and data augmentation to improve robustness. The results demonstrate that modern image segmentation models can reliably separate bacterial colonies from agar and other background regions, achieving macro-average Dice scores above 0.80 on unseen data. The dataset used in these experiments is publicly available at xxx 16:40 - 17:00    Break  17:00 - 19:00    High-Performance Computing, Computational Modeling, Simulations & Bioinformatics  1. C. Barakat (Forschungszentrum Jülich GmbH, Jülich, Germany), K. Athar (University of Iceland, Reykjavík, Iceland), S. Fritsch (Gesundheitsamt Düsseldorf, Düsseldorf, Germany), M. Riedel (University of Iceland, Reykjavík, Iceland) Biomedical Applications Leveraging AI-oriented HPC Infrastructures and Federated Learning Methods in EuroHPC AI Factory Ecosystems  Healthcare is one of the sectors that stand to gain the most from recent advances in Artificial Intelligence (AI) enabled by powerful High-Performance Computing (HPC) infrastructures. But one of the biggest hurdles to progress in this direction is the unavailability of sensitive yet required biomedical data for AI model training on cutting-edge HPC infrastructures due to privacy concerns and datasharing constraints. At the same time, over decades, those HPC infrastructures have been particularly configured and optimised for use by simulation science applications, such as those using numerical methods based on known physical laws. Prominent physics-based examples include, but are not limited to, Numerical Weather Prediction (NWP) and Computational Fluid Dynamics (CFD), while the large set of ever-increasing AI applications driven by recent advances in Deep Learning (DL) has had little influence on HPC infrastructure co-design apart from offering massive Graphical Processing Unit (GPU) processing power. More recently, this influence of AI application co-design in providing disruptive HPC infrastructures is prioritised by the innovative European EuroHPC AI Factory ecosystem, including their supporting national-driven AI Factory antennas. While many new data services and AI-driven technologies are changing the way EuroHPC infrastructure is provisioned for biomedical applications, addressing the aforementioned data privacy concerns and data sharing issues remains a challenge today. To address some of these issues, this paper presents a proof-of-concept based on innovative federated learning methods that show promising results for a selected biomedical AI use case within the new AI-oriented EuroHPC AI Factory ecosystem. 2. I. Vasileska, P. Tomšič, L. Kos (Faculty of Mechanical Engineering, Ljubljana, Slovenia) Performance and Portability of a Particle-in-Cell Code on Heterogeneous CPU–GPU Systems  The rapid evolution of high-performance computing systems toward heterogeneous architectures that combine CPUs and accelerators such as GPUs requires scientific applications to adopt programming models that ensure both high performance and portability. Particle-in-Cell (PIC) codes, widely used for kinetic plasma simulations, present significant challenges due to irregular memory access patterns, dynamic particle distributions, and strong coupling between particle and field calculations. In this work, we investigate the performance and portability of a PIC code implemented using Kokkos and SYCL for accelerator implementation, together with the Hypre library for a scalable electrostatic field solver. The core components of the PIC algorithm, the particle pusher, charge deposition, and field solver, are expressed using performanceportable programming models, enabling a single source code to target both multi-core CPUs and GPU architectures. For the field solver, a Hypre-based multigrid solver is integrated to evaluate its impact on performance and scalability. Performance experiments are managed on heterogeneous CPU–GPU platforms with phase-level profiling of the PIC workflow. The results show that Kokkos and SYCL achieve comparable performance on GPU devices, while also improving portability across various hardware backends. The Hypre solver demonstrates improved scalability for larger grid sizes, thereby reducing the time simulation cost of the field solver. 3. M. Rot, G. Kosec (Jožef Stefan Institute, Ljubljana, Slovenia) Meshless h-adaptive Solution for non-Newtonian Natural Convection in a Differentially Heated Cavity  One of the main challenges in numerically solving partial differential equations is finding a discretisation for the computational domain that balances the accurate representation of the underlying field with computational efficiency. Meshless methods approximate differential operators based on the values of the field in computational nodes, offering a natural approach to adaptivity. The density of computational nodes can either be increased to enhance accuracy or decreased to reduce the number of numerical operations, depending on the properties of the intermediate solution. In this paper, we utilise an adaptive discretisation approach for the numerical simulation of natural convection in non-Newtonian fluid flow. The shear-thinning behaviour is interesting both due to its numerous occurrences in nature, blood being a prime example, and due to its properties, as the decreasing viscosity with increasing shear rate results in sharper flow structures. We focus on the de Vahl Davis test case, a natural convection driven flow in a differentially heated rectangular cavity. The thin boundary layer flow along the vertical boundaries makes this an ideal test case for refinement. We demonstrate that adaptively refining the node density enhances computational efficiency and examine how the parameters for adaptive refinement affect the solution. 4. F. Tomas, M. Šikić (Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Position-Aware k-mer Categorization for Oxford Nanopore Sequencing Reads  Recent work has shown that explicit k-mer categorization at the read level can support long-read genome assembly by distinguishing erroneous, single-copy, multi-copy, and repetitive sequence content, with copy-number–related categories reflecting diploid structure. However, existing methods have been developed primarily for high-fidelity (HiFi) sequencing data, and their applicability to Oxford Nanopore Technologies (ONT) reads remains limited due to higher error rates and heterogeneous coverage profiles. In this work, we investigate k-mer categorization methods tailored specifically for ONT sequencing reads. We first propose an algorithmic approach that models position-dependent k-mer multiplicity along reads to enable read-level k-mer classification. Building on this baseline, we outline a convolutional neural network (CNN)–based formulation that operates on local k-mer multiplicity signals along the read, while incorporating global coverage-related features such as expected haploid and diploid depth. The proposed approaches are evaluated on ONT datasets to assess their ability to discriminate between different k-mer categories under nanopore-specific noise conditions. By explicitly accounting for nanopore-specific error characteristics and coverage variability, this work explores robust k-mer annotation in noisy long-read sequencing data. This study presents a structured framework for k-mer categorization in Oxford Nanopore reads, encompassing position-aware algorithmic modeling and a complementary learning-based formulation. Together, these approaches provide a foundation for future evaluation of k-mer–driven preprocessing in long-read genome assembly pipelines. 5. D. Strzelczyk (Institute Jožef Stefan, Ljubljana, Slovenia), B. Klun (University of Ljubljana, Ljubljana, Slovenia), M. Rot (Institute Jožef Stefan, Ljubljana, Slovenia), G. Kalčíková (University of Ljubljana, Ljubljana, Slovenia), G. Kosec (Institute Jožef Stefan, Ljubljana, Slovenia) Numerical Modeling of Fluid Flow in a Shallow Stream – towards Studying Microplastics Transport in Aqueous Environments  The uncontrolled release of microplastics into the natural environment in recent years has been recognized as the issue of paramount importance for the health and wellbeing of humans and the proper functioning of ecosystems. One of the key aspects in studying the impact of microplastics contamination on the environment is to investigate the fundamental processes governing its transport. This is especially true in aqueous environments, which allow the particles to cover vast distances quickly and where they can be easily uptaken by various organisms thus entering the food chain. In this work we leverage experimental observations and numerical modeling to study a flow in an artificial shallow stream composed of a gravel packing submerged in water. We present how to utilize the experimental measurements of fluid velocity profiles to constrain the numerical simulations of fluid flow in this system, with an outlook for the possible microplastics transport modeling approaches. Friday, 5/29/2026 9:00 AM - 2:00 PM, Camelia 1, Grand hotel Adriatic, Opatija BE - Biomedical Engineering  9:00 - 10:40    Deep Learning & Federated Approaches in Clinical Applications  1. T. Kostova, S. Gievska (Faculy of Computer Science and Engineering, Skopje, Macedonia) Optimizing Visual Feature Extraction in Multimodal Transformers for Neuroimaging Classification  Vision-language models have shown remarkable capabilities integrating visual and textual information, yet their application to medical neuroimaging remains underexplored, particularly regarding optimal visual feature extraction. This research benchmarks vision-language transformers for neuroimaging classification across four datasets: MRI brain tumor (binary and 12-class), multiple sclerosis, and CT stroke, totaling 21,500 images. We systematically investigate visual layer selection in Vision Transformers through experiments with CLIP and BiomedCLIP, demonstrating that shallow and middle layers can outperform traditionally-used deep layers for medical image analysis. Our experiments with zero-shot classification, linear probing, layer-wise extraction, and multi-layer fusion reveal four findings: (1) penultimate layers do not universally achieve optimal performance, (2) shallow (1-12) and middle layers (13-20) excel at fine-grained spatial tasks essential for diagnosis, (3) domain-specific pretraining through BiomedCLIP significantly improves medical understanding, and (4) multi-layer fusion yields consistent improvements over single-layer baselines. Our findings provide actionable insights for designing efficient medical image analysis systems, achieving competitive performance with reduced computational requirements. This work challenges conventional feature extraction practices, demonstrating that thoughtful architectural choices yield substantial improvements while maintaining efficiency suitable for clinical deployment. 2. N. Naseva, M. Toshevska, S. Gievska (Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University, Skopje, Macedonia) Relational Deep Learning for Clinical Outcomes  The increasing availability and quantity of electronic health records presents opportunities for predictive modeling in healthcare with benefits to both providers and patients alike. However, most existing approaches treat clinical data as flat tables or time series, neglecting the rich relational structures of clinical databases. In this work, we investigate relational deep learning (RDL) for clinical outcome prediction using the MIMIC-III dataset. Using the Relbench benchmarking framework, we evaluate GraphSAGE on four main prediction tasks: (i) ICU mortality prediction, (ii) patient readmission prediction, (iii) length of stay prediction and (iv) patient-disease link prediction. To assess the benefits of RDL with Graph Neural Networks, we compare the results against a strong tabular baseline (LightGBM) and a time series approach. Our experiments show that RDL, together with GraphSAGE, consistently outperform LightGBM and time series prediction on all evaluated tasks. These findings demonstrate the importance of explicitly modeling the relational structure of clinical data and highlight the potential of Relational Deep Learning for predictive tasks. 3. E. Guberović, H. Ivandić, B. Pervan (University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, Croatia) From Centralized to Federated Learning: Communication-Efficient Training of Clinical Risk Models  Federated learning (FL) has attracted growing interest as a way to train distributed machine learning (ML) models on sensitive healthcare data without collecting all the data in a single location. In principle, this enables multiple medical institutions to collaborate while preserving data locality and privacy. In this work, we examine the tradeoff between communication cost and predictive performance in a simulated FL setting for clinical risk prediction, using a real-world single-center dataset for sudden cardiac death (SCD) risk stratification. We simulate a federated scenario by partitioning the dataset across 100 virtual clients, treating each partition as a separate institution, allowing us to study federation mechanics in a controlled, reproducible manner. We follow up on research based on a centralized logistic regression (LR) use case on validated data, and then implement the same model within an FL framework designed for distributed training. To limit communication overhead, we evaluated several update compression strategies, including quantization, entropy-based coding, and delta-based update transmission. Our results show that federated training can achieve performance close to the centralized baseline while substantially reducing uplink communication, achieving more than a tenfold reduction through quantization and entropy-aware delta encoding. These findings suggest that communication-efficient FL is a practical and technically sound option even for simple, interpretable clinical models, and provide useful guidance for future deployment of FL systems in bandwidth-and privacy-constrained healthcare settings. 4. A. Dali (SogetiLabs research (part of Capgemini, Nantes, France), A. Fotouhi, T. Metral (SogetiLabs research (part of Capgemini, Paris, France) FedIA: A Lightweight Federated Learning Framework for Multi-Class Chest X-Ray Classification in Heterogeneous Environments  This paper introduces FedIA, a lightweight federated learning (FL) framework optimized for multi-class chest X-ray classification in heterogeneous environments. To address the critical challenges of non-IID data distribution, FedIA implements a streamlined CNN architecture enhanced with Group Normalization and proximal regularization. Our experimental results demonstrate that FedIA achieves a test accuracy of 93.24% in only 50 epochs, significantly outperforming both centralized baselines and high-capacity pre-trained federated models. By prioritizing task-specific design over model complexity, FedIA drastically reduces computational overhead while ensuring superior diagnostic precision. This work proves that efficient, decentralized architectures can effectively overcome data heterogeneity, providing a scalable solution for privacy-preserving clinical diagnosis. 5. O. Czimbalmos, Á. Pálfi, Z. Szántó, T. Szűcs, B. Raffael, G. Iglói, I. Kósa, G. Kőrösi (University of Szeged, Szeged, Hungary) Exploring Telemedicine Patient Trajectories: An Integrated Approach Using LLM Text Analysis and Similarity-Based Profiling  Our paper explores an integrated approach to analyze patient behavior within a 12-week metabolic telemedicine lifestyle intervention program. By utilizing Large Language Models (LLMs) to automatically evaluate unstructured coaching summaries, the study investigates early indicators of patient adherence. Concurrently, similarity-based behavioral profiling is applied to early-stage data to identify "behavioral twins" among historical patients. Preliminary findings indicate that LLM-extracted adherence levels show a relationship with objective metrics like weight loss and data logging frequency. Furthermore, matching early psychological profiles—such as stress and social support—with historical trajectories may offer insights into future adherence consistency. While current psychological correlations indicate trends rather than definitive predictions, this combined methodology presents a potential framework for identifying patients who might require personalized support. 10:40 - 11:00    Break  11:00 - 12:20    Biosignal Processing, Bioinformatics & Human Modeling  1. M. Kozlovszky (BioTech Research Center, Obuda University, Budapest, Hungary), L. Major (HDF Medical Centre Health Institution, Budapest, Hungary), I. Kiss, C. Kocsis, I. Bottlik (Obuda University, Budapest, Hungary) Development of an Artificial Human Circulatory System for Flow Simulation and Measurements  The human cardiovascular system is a complex and vital physiological system characterized by a large number of parameters that are difficult to measure directly. Prior to testing on human subjects, an appropriate experimental environment is required for the development and validation of medical devices intended for cardiovascular applications. In this research work, we developed a simplified theoretical model to investigate key aspects of human blood circulation. Based on this model, we designed and implemented a physical blood circulation testing platform capable of dynamically simulate multiple circulatory parameters in real time (including pressure, temperature, and pH). This paper presents our development results. We introduce the realized prototype system and provide a detailed description of its principal features and functionalities. 2. A. Ilic, L. Milic (Faculty of Technical Sciences, University of Novi Sad, Novi Sad, Serbia), B. Petrovic (Faculty of Medicine, University of Novi Sad, Novi Sad, Serbia), V. Strbac (Faculty of Technical Sciences, University of Novi Sad, Novi Sad, Serbia), L. Adamov (Faculty of Medicine, University of Novi Sad, Novi Sad, Serbia), S. Kojic, G. Stojanovic (Faculty of Technical Sciences, University of Novi Sad, Novi Sad, Serbia) Statistical and Graphical Analysis of Facial Muscle Symmetry Using Bilateral Surface Electromyography  Disruption of facial neuromuscular balance, such as in facial paresis, is commonly evaluated using subjective clinical scales that are limited in detecting mild asymmetries. This highlights a growing need for objective, quantitative assessment methods for early diagnosis and rehabilitation. Surface electromyography (sEMG) provides a non–invasive approach for bilateral (on both sides) measurement and comparison of facial muscle activation. In this study, bilateral sEMG signals were recorded from selected facial muscles in healthy individuals during standardized voluntary tasks using a BIOPAC system. Signal processing and feature extraction were performed in Python, focusing on amplitude– and frequency–based parameters. Symmetry between the left and right sides of the face was evaluated using statistical analysis and multivariate visualization techniques. Results indicate a high degree of bilateral symmetry in healthy subjects, with mostly no statistically significant differences between sides (p > 0.05) and muscle–specific variability across different facial regions. These findings establish quantitative reference characteristics of normal facial muscle symmetry and support future objective assessment and rehabilitation monitoring. 3. D. Katović (University of Zagreb Faculty of Kinesiology, Zagreb, Croatia), T. Bronzin, D. Adamec, B. Prole, A. Stipić (CITUS, Zagreb, Croatia), J. Ćelić (University of Rijeka Faculty of Maritime Studies, Rijeka, Croatia), D. Matek (University Hospital Centre Zagreb Department of Orthopaedics, Zagreb, Croatia), M. Horvat (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Frameworks and Key Concepts for Developing Personalized Reference Models in Human Pose Estimation  This paper positions Personalized Reference Models (PRMs) in human pose estimation (HPE) as a structured research direction rather than as a completed implementation. Based on a focused review of biomechanics, sensing, calibration, adaptive learning, and ethical governance, the paper synthesizes the main frameworks required for building personalized HPE systems and proposes an integrated conceptual architecture for their development. To make the contribution explicit, the paper also outlines illustrative quantitative elements that may support anthropometric scaling, personalized baselines, adaptive tolerance envelopes, deviation scoring, and longitudinal updating. No new dataset, full implementation, or experimental validation is claimed here; instead, the contribution lies in organizing the state of the art into a coherent development logic and in identifying the design steps needed for future implementation and validation in rehabilitation, sports, and fitness applications. 4. L. Pavić, A. Mutka (RIT Croatia, Zagreb, Croatia), D. Tolić (RIT Croatia, Dubrovnik, Croatia), T. Martinčić (RIT Croatia, Zagreb, Croatia) Detecting Student Stress through Keystroke Patterns in Programming Assignments  Modern student life is increasingly characterized by stress and anxiety, which can exacerbate risks of academic dishonesty and hinder academic performance. The growing use of digital tools in education raises concerns about issues like plagiarism and AI-assisted cheating. Keystroke dynamics, commonly used for user authentication, can also reveal insights into stress levels. Using keystroke dynamics (timing, consistency, and typing patterns) with students’ self-reported stress levels, which serve as ground truth for machine learning models, reveals several important insights into the impact of stress on typing behavior, as well as the potential of machine learning models to classify stress levels with reasonable accuracy. Therefore, TSFRESH features and total interaction metrics features were combined. These combined features provided the most robust results. XGBoost achieved an F1 score of 0.65 (low stress) and 0.79 (high stress), with an overall classification accuracy above 83\%. By leveraging a combination typing behavior and task level interaction metrics, machine learning models, particularly XGBoost, were able to classify stress levels with high accuracy. A major strength of this project is its focus on real student data collected in authentic coursework environments. This improves the validity of the findings and positions the models for real world application. Our results demonstrate that keystroke dynamics when combined with machine learning can be a powerful tool for understanding student stress. The high accuracy of the models and the scalability of the approach make it a strong candidate for integration into digital learning environments. 12:20 - 12:40    Break  12:40 - 14:00    Cardiac & Cardiovascular Signal Analysis  1. H. Ivandić, B. Pervan, J. Knezović (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia), M. Puljević (University of Zagreb School of Medicine, University Hospital Centre Zagreb, Zagreb, Croatia), A. Jović (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Evaluating the Trade-off Between Predictive Performance and Model Fairness in Sudden Cardiac Death Risk Stratification  Predicting implantable cardioverter-defibrillator (ICD) activation using machine learning (ML) is a promising approach for identifying patients at high risk of sudden cardiac death (SCD). Building on our previous optimization of SCD risk stratification ML models, this study investigates how data preprocessing affects model performance, feature importance, and fairness. We evaluate whether the native handling of missing data by gradient boosting algorithms, such as CatBoost (CB), outperforms traditional manual preprocessing and sampling techniques. Our analysis compares two pipelines: (1) multiple ML models trained on numerically encoded, manually imputed, scaled, and sampled data, and (2) a CB model trained on the original, unprocessed dataset, leveraging its native ability to handle missing values and categorical data. Using SHAP for explainability and fairness metrics to assess potential performance disparities, we quantify the impact of preprocessing techniques on the models’ ability to identify high-risk patients across different groups. Results indicate that manual preprocessing negatively impacts CB performance. Furthermore, we confirm enhanced predictive performance on minority groups, specifically female patients, younger individuals, and those with LVEF > 35%. 2. A. Angjelevska, M. Micevska (Innovation Dooel, Skopje, Macedonia), M. Gusev, E. Zdravevski (University Sts Cyril and Methodius, Skopje, Macedonia) Explainable Artificial Intelligence Techniques for the Transformer-Based Heart Language Model Finetuned for Detection of Atrial Fibrillation  In medical diagnostics, a machine learning model’s performance cannot be judged solely by its accuracy. This article discusses integrating a robust explainable artificial intelligence framework into a transformer-based model for Atrial Fibrillation. Although both fine-tuning and evaluation use the same MIT-BIH database, the reported metrics reflect in-sample performance rather than generalization to unseen patients; the XAI analysis therefore focuses on auditing the model’s learned behavior, verifying clinically relevant patterns rather than spurious correlations. This work offers an approach to ”open” the black box, making the model’s reasoning transparent, trustworthy, and auditable. Our methodology combines global and local explanation techniques to uncover the model’s overall strategy and ensure that it has learned a clinically sound, ”sustained evidence” approach. It emphasizes the specific characteristics of an ECG segment to identify sustained, characteristic irregularities rather than acting as a simple ”chaos detector,” and offers a detailed, beat-by-beat analysis of individual predictions, revealing critical insights into both correct and incorrect outcomes. The error analysis identifies interpretable, case-driven hypotheses for model improvement: false positives arise from rhythms with irregular patterns, and false negatives arise from atypical AFIB presentations that fall outside the model’s learned templates. Beyond serving as a validation tool, this approach is a powerful diagnostic framework for understanding, debugging, and systematically improving clinical models. 3. M. Madabhushi (Indian Institute of Technology Madras Zanzibar, Zanzibar, Tanzania), S. Banik, R. Swaminathan (Indian Institute of Technology Madras, Chennai, India) Evaluation of Emotion Recognition in ECG Using Morphological Features  Emotions play a crucial role in human cognitive processes by influencing perception, reasoning and behavioral responses. In recent years, physiological signals have been widely explored for emotion recognition. In this work, an attempt is made to classify emotional states using Electrocardiogram (ECG) signals through basic morphology based feature analysis. A total of sixty features are extracted from the PQRST complex, including amplitude, interval and duration-based features. The extracted features are evaluated using various machine learning algorithms. SHAP analysis shows that valence prediction is primarily influenced by Pwave, QT interval, S-wave, RR-interval and QRS duration variability. Arousal prediction is primarily influenced by Twave, Q-wave, QRS duration variability and QT interval. Although extracted features contribute to the model these features consistently exhibit higher importance. As Arousal and Valence changes are reflected by various ECG biomarkers. This proposed approach demonstrates potential for emotionaware applications in wearable monitoring, human–computer interaction and mental well-being assessment. 4. S. Tudjarski, M. Gusev, B. Risteska Stojkoska (Faculty of Computer Science and Engineering, Skopje, Macedonia), A. Stankovski (Polytechnic University of Madrid, Madrid, Spain) Measuring Performance of Heartbeat Clustering on IoT and Post-Cloud Architectures  Energy consumption is crucial for wearable devices, making their optimization a key research focus to extend system lifespan and enhance functionality. One effective way to conserve energy is to reduce communication by promoting onboard processing. Instead of transmitting raw measurements to higher levels in the system hierarchy, intelligent algorithms can be executed directly on the wearables' processors. This paper compares different platforms for clustering ECG heartbeats and evaluates whether embedded devices are suitable for real-time onboard processing as part of edge computing. We analyze the computational capabilities of the following platforms: IoT, represented by the Raspberry Pi; mobile computing, represented by an Android-based smartphone; traditional desktop computing, represented by x64 and M1 processors; and cloud computing, represented by Google Colab. The results show that the time needed to perform clustering on IoT, mobile, and cloud platforms is comparable, making dew computing a viable alternative for data processing. These results suggest that IoT and mobile devices have enough computational power to perform heart-monitoring calculations without needing to transfer data to remote systems (e.g., cloud or desktop). This eliminates the reliance on network connectivity for early heart failure alerts, which is essential for patients in areas with poor coverage.

Basic information:

Karolj Skala (Croatia), Aleksandra Rashkovska Koceva (Slovenia), Davor Davidović (Croatia)

The student discount doesn't apply to PhD students.

NOTE FOR AUTHORS: In order to have your paper published, it is required that you pay at least one registration fee for each paper. Authors of 2 or more papers are entitled to a 10% discount.

Karolj Skala
Ruđer Bošković Institute
Center for Informatics and Computing
Bijenicka 54
HR-10000 Zagreb, Croatia

E-mail: skala@irb.hr

Accepted papers will be published in the ISSN registered conference proceedings. Papers presented at the conference will be submitted for inclusion in the IEEE Xplore Digital Library.

Opatija is the leading seaside resort of the Eastern Adriatic and one of the most famous tourist destinations on the Mediterranean. With its aristocratic architecture and style, Opatija has been attracting artists, kings, politicians, scientists, sportsmen, as well as business people, bankers and managers for more than 180 years.

The tourist offer in Opatija includes a vast number of hotels, excellent restaurants, entertainment venues, art festivals, superb modern and classical music concerts, beaches and swimming pools – this city satisfies all wishes and demands.

Opatija, the Queen of the Adriatic, is also one of the most prominent congress cities in the Mediterranean, particularly important for its ICT conventions, one of which is MIPRO, which has been held in Opatija since 1979, and attracts more than a thousand participants from over forty countries. These conventions promote Opatija as one of the most desirable technological, business, educational and scientific centers in South-eastern Europe and the European Union in general.

For more details, please visit www.opatija.hr and visitopatija.com.