HEALTHINF 2026 Abstracts

Full Papers

Paper Nr:	49
Title:	Intensive Care Unit Nursing Workload Assessment with IoT and RFID: A Case Study
Authors:	René Nolio Santa Cruz, Hugo Vaz Sampaio, Carlos Becker Westphall, Maximiliano Dutra de Camargo and Daniela Couto Carvalho Barra
Abstract:	Nursing workload assessment is crucial to allocate an adequate amount of human resources in intensive care units to provide high-quality patient care and avoid stress and burnout in nursing professionals. The nursing workload assessment tools most cited in the literature usually require manual data entry, consume nurses’ time, and cover at most 80.8% of all nursing activities, resulting in an estimate that may not always accurately represent the actual workload. In order to help reduce this gap between theory and practice, we propose an Internet of Things (IoT) device to automatically document nursing workload in an intensive care unit via Radio-frequency Identification (RFID), resulting in a more transparent and less work-intensive approach. The main contribution of this paper is assessing the viability and benefits of the proposed solution by deploying a prototype in a real-world scenario, as well as highlighting challenges that should be addressed before conducting more extensive studies.
Download

Paper Nr:	52
Title:	Artificial Intelligence Adoption Dynamics among Lebanese Dietitians
Authors:	Joanne Mhawej, Nabil Georges Badr and Joumana Yeretzian
Abstract:	In the world of dietetics, Artificial Intelligence (AI) holds the potential to enhance the effectiveness of diet therapy by providing clinical support, real-time monitoring, and increased efficiency. Despite the benefits, the factors influencing dietitians’ intentions to use it remain unclear. This research explores these factors using Technology Acceptance Model 2 (TAM2). Employing a mixed-methods approach, 177 Lebanese clinical dietitians completed an online survey, and 12 underwent semi-structured interviews. Job Relevance (JR) was the strongest predictor of the Intention to Use through Perceived Usefulness (p < 0.001). Subjective Norms (SN) was a weak but significant predictor. AI adoption remains in its early-stage, with key barriers including trust concerns, limited education and the absence of standardized tools. Supporting dietitians with targeted education, relevant tools, and adequate policies will be essential to ensure AI technologies are perceived as relevant, useful, and easy to adopt, ultimately facilitating effective AI integration and improving both practitioner efficiency and patient outcomes.
Download

Paper Nr:	53
Title:	How ‘Shared’ is Shared Decision Making? The Influence of COPD Features on Healthcare Professionals’ Perceptions: A Cohort Vignette Study
Authors:	Wendy d'Hollosy, Roswita Vaseur, Tessa Beinema, Kjell Lemmen and Monique Tabak
Abstract:	Better understanding of how healthcare professionals (HCPs) judge patients’ capability for shared decision making (SDM) is needed to ensure fair access to collaborative care, particularly for patients with complex chronic conditions. This study examined the influence of 19 features of patients with Chronic Obstructive Pulmonary Disease (COPD) and comorbidities on 1) HCPs’ perceptions of SDM capability and 2) their preference to receive technical support during a SDM process. Feature importance was assessed in a vignette study where 49 HCPs judged 12 vignettes each, yielding 587 judgments. SHAP analysis showed that positive treatment attitude, higher educational level, good COPD knowledge, and native language strongly shaped HCPs’ perceptions of patients’ SDM capability. Negative treatment attitude, poor health literacy, poor COPD knowledge, low self-efficacy, and lack of social support strongly increased HCPs’ preference for technical SDM support. The low intraclass correlation ICC(1) score of 0.145 suggests considerable variation between individual HCPs in assessing patients’ SDM capability. Yet, the high ICC(1,k) score of 0.892 indicates strong overall agreement. These insights provide a valuable basis for understanding professional judgments but also reveal potential biases that may foster unequal SDM practices. Future research should focus on technical SDM support tools to promote more equitable, patient-centered SDM.
Download

Paper Nr:	58
Title:	An Innovative Infrastructure for a Pediatric Oncology Network in the Campania Region
Authors:	Erica Ambrosino, Marco Aruta, Dalia As Ad, Sara Ferrigno, Ciro Listone, Floriana Nappi, Salvatore Romano, Melania Ruopolo, Francesco Adinolfi, Michele Cimmino, Mariapia Raimondo, Francesco Serino and Aniello Murano
Abstract:	Progress in medical research has greatly improved survival among children with oncohematological diseases. Yet, for pediatric patients with leukemia or solid tumors, treatment remains long, complex, and emotionally demanding. Therapies often last years, requiring constant monitoring, frequent hospital visits, and intensive care. Geographic disparities add further challenges. Specialized oncology centers are concentrated in large cities, forcing families from smaller towns or rural areas to travel long distances. These repeated transfers generate financial strain, organizational difficulties, and emotional stress, negatively affecting adherence and outcomes. Moreover, communication between hospitals and local health services is often fragmented, creating gaps in continuity of care. The INFANT project, developed in Campania, Italy, was created to address these issues through a hub-and-spoke model. In this system, main hospitals collaborate with local facilities to deliver high-quality standardized care closer to patients’ homes. Supported by interoperable digital platforms and artificial intelligence, INFANT promotes earlier diagnosis, personalized treatments, and improved monitoring. The program also integrates educational and psychological support, fostering a holistic approach to pediatric oncology. By reducing inequalities and strengthening care pathways, the project seeks to ensure sustainable, equitable, and patient-centered treatment for children and their families.
Download

Paper Nr:	65
Title:	Modeling the Nationwide Impact of Behavioral Changes on the Number of Patients with Lifestyle-Related Diseases and Medical Costs
Authors:	Shingo Kawai, Masako Toriya and Tetsuya Toma
Abstract:	This study numerically investigates the nationwide reduction in the number of patients with lifestyle-related diseases (LDs) and the associated medical costs in 2050 through behavioral changes. Two parameters play central roles: the proportion of the population adopting behavioral changes and the annual reduction in the morbidity rates of LDs. Although the number of patients with LDs and the associated medical costs decline as these parameters improve, they may also increase, as behavioral changes can reduce mortality rates, leading to overall population growth. Under the assumption that 100% of the target population adopts behavioral changes and achieves a 10% annual reduction in LD morbidity rates, while mortality rates remain the same between the behavioral change and nonbehavioral change groups, the number of patients with LDs and associated medical costs are projected to decrease by 41.0% and 40.9%, respectively, in 2050. However, if mortality rates differ between the groups as a result of behavioral changes, the projected reductions diminish to 15.3% and 15.2%. Ultimately, whether these numbers increase or decrease depends on the balance between these opposing two effects, outcomes that will significantly influence the design of future society.
Download

Paper Nr:	80
Title:	Using a Set-Based Approach to Explore Patterns of Lost Autism Diagnoses in Hospital Data
Authors:	Roy A. Ruddle, Layik Hama, Pamela Wochner and Oliver T. Strickson
Abstract:	Health conditions such as autism do not remit (i.e., some symptoms always persist). If a patient with one of those conditions is treated as an in-patient then the hospital health record for that admission should contain a relevant diagnosis code (i.e., the diagnosis should ‘persist’), irrespective of the reason for admission (e.g., cancer). However, for years the UK’s National Health Service has been concerned that diagnostic persistence is poor for non-remitting conditions. This paper describes an investigation of diagnostic persistence for autism that used a dataset containing 25,152 hospital episodes from 6383 autism patients who were treated in 224 hospitals. Machine learning models were not accurate for predicting when autism diagnoses would be ‘lost’, but did indicate that the number of diagnoses, the primary diagnosis and hospital were the most important features. Those features provided a starting-point for a visual analytic investigation that uncovered seven patterns characterizing when autism diagnoses become lost. One pattern applied nationwide (99.6% of autism diagnoses were lost with an R69 primary diagnosis). Half of the lost diagnoses occurred when patients were admitted frequently (every 7 days or less, on average), and that included patients who had repeated treatment for a D61 primary diagnosis. A common theme in other primary diagnosis patterns was that autism diagnoses were always (or almost always) lost in some hospitals, but always or mostly persisted in other hospitals. That provides an opportunity to learn from pockets of good clinical coding practice and significantly improve persistence for autism diagnoses.
Download

Paper Nr:	82
Title:	CICADA: A Workflow Guidance for Causal Machine Learning in Precision Medicine
Authors:	Louk van Remmerden, Ilona Barańska, Ameet Jagesar, Eline Tankink-Kooijmans, Jitka Pokladníková, Paulina Wiśniewska, Daniela Fialová, Emiel Hoogendijk, Karlijn Joling, Katarzyna Szczerbińska, Vincent François-Lavet, Shujian Yu, Mark Hoogendoorn and Hein van Hout
Abstract:	Causal machine learning (ML) is increasingly applied in precision medicine to estimate individualized treatment effects, making it a powerful tool for personalized healthcare and treatment. However, many methods overlook clinical and domain-specific considerations, and rarely explain how to translate modeling assumptions into clinical practice. This often results in models that do not fully integrate clinical considerations into the modeling process, thereby limiting their usability in real-world applications such as decision support systems. Bridging this gap requires frameworks that explicitly embed clinical knowledge into each stage of causal modeling, ensuring that methodological rigor aligns with practical healthcare needs. We present Causal Inference for Clinical and Decision Assistance (CICADA), an end-to-end framework that integrates clinical expertise into key modeling stages through clinically informed design choices. CICADA provides structured guidance on cohort selection, treatment definitions, and confounder control, aligning technical objectives with clinical priorities. We evaluate CICADA across four clinical use cases and in a user study with domain experts, demonstrating that the models produced by the framework are both interpretable and practically useful.
Download

Paper Nr:	87
Title:	Giving Voice to Low-Income Older Adults with T2DM: Design and Evaluation of the JTG Dietary Management System in a Carbohydrate-Centric Culture
Authors:	Yibo Meng and Bingyi Liu
Abstract:	The management of type 2 diabetes mellitus (T2DM) relies heavily on dietary regulation, yet in northern China many low-income older adults face cultural, economic, and literacy barriers that hinder effective self-care. Most digital health solutions fail to reflect carbohydrate-heavy diets or the usability needs of this population. We present Jintang Weishi (JTG), a culturally tailored dietary management system for low-income older adults with T2DM in Shanxi Province. In a 90-day randomized controlled trial with 40 participants, JTG significantly improved diabetes knowledge, reduced daily carbohydrate intake, and lowered HbA1c, while showing high usability and acceptability. These findings demonstrate the promise of culturally adapted digital health tools for improving diet, literacy, and short-term glycemic control, and highlight opportunities for integration into broader community care.
Download

Paper Nr:	89
Title:	Hybrid Drug-Drug Interaction Prediction: Combining Knowledge Graph Embeddings and Symbolic Learning
Authors:	Duru Naz Han, Romy Vos and Annette ten Teije
Abstract:	Drug–drug interactions (DDIs) pose severe clinical and financial risks and remain difficult to predict. Current approaches either use Knowledge Graph Embeddings (KGE), which capture implicit relationships but lack explainability, or symbolic learning, which improves interpretability but struggles with complex relationships. Inspired by Decagon's graph-convolutional framework and the VISE pipeline, this paper explores a hybrid approach combining symbolic rule mining with KGE for DDI prediction. A two-stage pipeline is developed by training KGE models using the PyKEEN framework on a filtered DDI knowledge graph (KG) and enriching with Horn rules mined by AMIE+ and AnyBURL. The research questions: the effect of symbolic enrichment on predictive performance, symbolic learning method comparison, and relation‐prediction versus head/tail completion are addressed. Results show AMIE+ delivers the largest gains in entity prediction (RotatE's MRR increasing by 7%), while the broader rules of AnyBURL enhance relation prediction by around 4% in most metrics. Relation tasks remain 0.20–0.60 MRR points lower due to negative sampling biases. In conclusion, symbolic rule mining complements KGEs for accurate and interpretable DDI discovery, and future work on SHACL‐based validation and relation‐focused embedding models are proposed.
Download

Paper Nr:	130
Title:	Pattern Analysis in Oropharyngeal Cancer Using Formal Concept Analysis
Authors:	Arthur Setragni, Julio Neves and Mark Song
Abstract:	Head and neck cancers (HNC) comprise a heterogeneous group of malignancies and remain a major global public health concern. Among them, Oropharyngeal Squamous Cell Carcinoma (OPSCC) has been increasing globally, driven by shifts in exposure to classical risk factors such as tobacco use, alcohol consumption, and infection with the Human Papillomavirus (HPV). To better understand these complex interactions, this study applies Formal Concept Analysis (FCA) – a mathematical framework for extracting structured knowledge from binary relations – to a clinical-epidemiological dataset comprising 277 OPSCC patients treated at Barretos Cancer Hospital between 2008 and 2019. The analysis uncovers strong associations between behavioral risk factors (tobacco and alcohol abuse), HPV status, and clinical tumor characteristics, such as TNM stage and anatomical extension. The findings demonstrate that FCA can complement conventional statistical approaches, yielding interpretable models that enhance clinical understanding and advance data-driven oncology research.
Download

Paper Nr:	134
Title:	Emerging Clinician Readiness for Data Analysis in Augmented Era
Authors:	Neha Sundar Naik, Zhonglin Qu, Quang Vinh Nguyen, Simeon Simoff, Paul Kennedy and Daniel Catchpoole
Abstract:	Data analysis in healthcare is a complex, multi-stage process that necessitates clinician expertise and active engagement before analytical tools find their way into routine clinical practice. As immersive technologies such as virtual reality (VR) and augmented reality (AR) continue to advance, uncertainty remains about their effectiveness to aid in analytical workflows and their impact on clinicians' ability to perceive and interpret data. As part of a qualitative evaluation, we interviewed 9 medical students and 9 early-stage translational researchers who interacted with Microsoft HoloLens 2. Participants explored analytical workflow in augmented reality, identifying the concepts and contexts that are facilitated by the technology and those that posed challenges. Our findings indicate higher participant readiness in immersive domains, including interaction skills, perceived individual value, and broader impact. However, readiness was moderate in perceived analytical proficiency, such as the effective use and interpretation of statistical tools or software. Notably, readiness was lower when participants’ responses reflected third-party perspectives, specifically when analytical and immersive domains intersected, such as perceived peer commitment and clarity around long-term implementation. These findings highlight the need for nuanced exploration of data analysis processes and require integration of clinician-centred perspectives into the development and implementation of immersive analytical technologies to support data-driven decision-making in clinical research.
Download

Paper Nr:	137
Title:	Leveraging SFT and RL for Fine-Tuning LLMs to Generate Supplementary Annotated Data for Entity Recognition
Authors:	Cheng Cao, Ying Wei, Qi Li and Jay Pillai
Abstract:	We propose a framework that combines supervised fine-tuning (SFT) and reinforcement learning (RL) to fine-tune large language models (LLMs) for generating high-quality synthetic annotations for named entity recognition (NER). The model is trained in a dual-task setting that jointly optimizes document generation and structured entity annotation, producing contextually rich texts with consistent annotations. Reinforcement learning further improves generation quality by aligning model outputs with semantic fidelity and linguistic diversity objectives. Experiments on both public and proprietary biomedical datasets show that synthetic annotations generated by the proposed method significantly improve downstream BERT-based NER performance. Models trained with a combination of real and synthetic data consistently outperform those trained on human-annotated data alone, with RL-enhanced generation yielding the largest gains. These results demonstrate that LLM-generated synthetic annotations provide an effective, scalable, and cost-efficient supplement to manual annotation for domain-specific NER tasks.
Download

Paper Nr:	144
Title:	Decision Complexity in Asynchronous Patient Portal Messages: Implications for Provider Experience and EHR Usability
Authors:	Dong-Gil Ko
Abstract:	Healthcare providers face a growing burden from patient portal messages, a convenient telehealth tool that often adds to after-hours work. Some health systems have introduced billing for complex e-visits to manage this load. This study introduces a Decision Complexity (DC) Index to quantify the cognitive effort providers expend when responding to asynchronous messages. Using audit log data from a large health system, a composite metric combining Information Engagement Depth (IED) and Task Switching Intensity (TSI) was derived. Over 180,000 provider responses before and after the 2023 rollout of message billing were analyzed. Regression models, including a difference-in-differences design, assess the billing policy’s impact and other predictors of DC. Complexity per message declined slightly post-policy (≈5%, p<0.001), but increased for billable messages among providers who adopted billing (β≈+0.19, p<0.001). Primary care physicians had higher DC than specialists (β≈–0.05, p<0.001), and vague patient messages were met with lower complexity responses, especially in primary care. The DC Index offers a practical, theory-based measure of clinician workload and can inform EHR design and adaptive decision support tools aimed at easing cognitive burden and improving provider experience.
Download

Paper Nr:	151
Title:	Incorporating Interoperability for Addressing Health-Related Social Needs in the Complex Model for Screening and Referral Fulfillment
Authors:	Paulina Sockolow, Joy Doll, Yasemin Algur and Edgar Y. Chou
Abstract:	Health-related social needs (HRSN) are key determinants of health outcomes, yet healthcare organizations often lack the infrastructure to effectively manage screening and referral fulfillment (S&RF) processes. Fragmented data systems and limited interoperability hinder the ability to track patient retention and referral completion. This study aimed to refine an evidence-based complex model of the HRSN S&RF process by incorporating real-world data and identifying gaps in referral fulfillment. Drawing from literature and expert input, the model organizes 88 factors across five stages of the S&RF process. Analysis of 50 studies revealed wide variability in reported parameters and significant patient attrition, particularly between referral initiation and fulfillment. To address this, we conducted a retrospective observational study using data from a statewide health information exchange (HIE), linking electronic health records (EHRs) and HRSN platform data. The findings informed updates to the model, adding new factors related to community-based organizations (CBOs) and information systems. These enhancements improve the model’s representation of interoperability and support future intervention design and simulation modeling to improve referral fulfillment and health outcomes.
Download

Paper Nr:	155
Title:	Exploring the Feasibility of Self-Led, Personalized and Adaptive Cognitive Training after Stroke
Authors:	Teresa Paulino, Rita Costa, Joana Câmara and Mónica S. Cameirão
Abstract:	Stroke is a leading cause of disability worldwide, with survivors often experiencing persistent cognitive impairments that hinder their daily functioning. While cognitive rehabilitation is crucial for post-stroke recovery, tailoring interventions to individual needs maximizes effectiveness. Digital technologies enable automatic personalization, adaptation, and remote delivery for home-based recovery. This study explored the feasibility of the Reh@Sync, a digital system that automatically personalizes and adapts cognitive training. Ten chronic stroke survivors completed a remote, self-led intervention using the system. We measured user adherence, experience, and clinical impact on cognitive function. Findings revealed a 100% adherence and user experience evaluation yielded high levels of perceived usability, satisfaction, and user engagement with the technology. The intervention resulted in statistically significant improvements in cognitive function for a subgroup of participants with baseline cognitive levels below the normative range. Our findings highlight the potential of the Reh@Sync for enhancing post-stroke cognitive training. Future research through controlled studies with larger samples and more strict inclusion criteria is required to confirm the initial findings and establish the generalizability and long-term benefits of this approach.
Download

Paper Nr:	163
Title:	Order-Aware Metric Learning for Electronic Health Records Applications
Authors:	Mojgan Kouhounestani, Long Song, Ling Luo, Uwe Aickelin and Mark John Putland
Abstract:	Structured electronic health records (EHRs), widely used as sources of patient information, contain heterogeneous data types, including nominal, numerical, and ordinal variables. Effectively utilising these data types remains a fundamental challenge in medical data mining due to their intrinsic differences. While nominal and numerical data are typically handled using conventional encoding and arithmetic methods, ordinal data occupies a unique intermediary position. Ordinal variables embody an inherent ordered structure but often assume uniform intervals between categories, and the subjective nature of patient-reported ordinal values further complicates their utilisation. These challenges necessitate specialised frameworks to handle ordinal features accurately. In this study, we introduce and evaluate our proposed framework, applying it to MIMIC-IV-ED as a publicly available dataset and real-world data from the Emergency Department (ED) of Royal Melbourne Hospital (RMH). Our evaluation comprises two scenarios: one focusing exclusively on ordinal predictors and another involving a combination of nominal, ordinal, and numerical features. In the ordinal-only scenario of RMH, our approach enhances Logistic Regression, improving AUROC from 0.733 to 0.746 and AUPRC from 0.748 to 0.753. These gains demonstrate that explicitly modeling ordinal structure yields meaningful improvements without increasing model complexity and highlights the critical role of updating ordinal values while preserving order, thereby improving predictive accuracy and mitigating subjectivity in ED disposition decisions.
Download

Paper Nr:	176
Title:	Dynamic Risk Prediction of Hospitalized Patients with Competing Risks: A Comparative Study of DeepHit and Dynamic-DeepHit on the MDH Cohort
Authors:	Mateus Henrique Zeiser, Felipe André Zeiser, Adriana Vial Roehe and Cristiano André da Costa
Abstract:	We study dynamic risk prediction for hospitalized COVID-19 using the MyDigitalHealth dataset collected from a Brazilian hospital. The dataset comprises 1,815 RT-qPCR–confirmed admissions (March 2020–June 2022), including demographics, comorbidities, vital signs, laboratory tests, chest X-rays, and clinical records, which reflect heterogeneous and irregular real-world data. Outcomes are modeled as competing risks: discharge, transfer to intensive care, and in-hospital death. We compare proportional-hazards baselines, including time-dependent and competing-risks variants, with neural survival models (DeepSurv, PCHazard, DeepHit) and Dynamic-DeepHit. Preprocessing addresses missingness and standardizes features. Discrimination (concordance) and calibration (Brier score) are evaluated at clinically meaningful prediction times. Dynamic-DeepHit achieves the best short-term discrimination at admission and maintains competitive calibration during hospitalization, indicating that explicit modeling of longitudinal trajectories and cause-specific risks improves clinical stratification. These results show the potential for dynamic deep learning systems to streamline real-time patient triage and monitoring, leading to a more efficient allocation of hospital resources.Findings support dynamic neural survival modeling for triage and monitoring, while highlighting the need for external validation and broader multimodal integration.
Download

Paper Nr:	183
Title:	Harmonizing ELSA Data with OMOP and FHIR Standards: Enabling Longitudinal Research on Diabetes and Cardiovascular Diseases
Authors:	Francisco Lozano del Moral, Julia Sánchez Esquivel, Sergio Paraiso-Medina, Raúl Alonso Calvo, Paloma Jimeno, Inmaculada Luengo and Víctor Maojo
Abstract:	This study presents an ETL pipeline, available upon request, for converting data from the English Longitudinal Study of Ageing (ELSA) into OMOP CDM and FHIR formats. Designed to support research on diabetes and cardiovascular diseases, the pipeline addresses semantic inconsistencies and documentation gaps across the different survey periods of the longitudinal study. Over 91% of selected variables were successfully mapped to standard clinical vocabularies, enabling interoperability with hospital and clinical trial datasets. The resulting FAIR-compliant dataset supports scalable, artificial intelligence-driven healthcare research in chronic conditions like diabetes and cardiovascular disease, aligning with Horizon Europe’s goals for personalized medicine.
Download

Paper Nr:	186
Title:	Secure AI-Driven Solution for Enhancing Claim Management Systems in Healthcare
Authors:	Rita Zgheib, Yasmine Remah Sadek, Moustafa Sherif, John Aditiyah and Mohamed Abdelmksoud
Abstract:	Claims Management Systems (CMS) remain burdened by manual workflows, high error rates, and increasing fraud attempts, affecting healthcare providers, insurers, and patients. This study presents a secure, AI-enabled framework aligned with the EU AI Act definition of machine-learning-based automated decision support that integrates claim classification, fraud detection, and security hardening into a unified architecture. Using 50,000 anonymized real-world claims, the proposed system combines clustering-based segmentation, Random Forest classification, and a hybrid ensemble fraud detection module (Isolation Forest, Logistic Regression, and Gradient Boosting with majority voting). The system achieves 92% classification accuracy and 96% fraud-detection accuracy on held-out data, while security hardening and testing report complete mitigation of critical OWASP-aligned web vulnerabilities. Comparative evaluation indicates that the proposed approach improves automation reliability, fraud-case visibility, and security resilience relative to prior CMS studies. These results demonstrate that integrating machine learning with secure engineering practices provides a scalable and operationally effective enhancement to legacy CMS environments.
Download

Paper Nr:	193
Title:	MedQ: A Dynamic Meta-Model for Integrating Clinical Questionnaires and Digital Biometrics
Authors:	Nadiana K. N. Mendes, Rossana M. C. Andrade, Pedro A. M. Oliveira, Evilasio C. Junior, Ismayle S. Santos, Victor A. Silva, Wilson Castro and Victoria T. Oliveira
Abstract:	This paper presents MedQ, a system and meta-model designed to integrate clinical questionnaires with real-time health data collected through mobile Health Data Containers, addressing the limitations of traditional static health assessment tools. The aging global population and the increasing reliance on technology in healthcare highlight the need for adaptable solutions that combine subjective patient-reported outcomes with objective data from wearable devices. Traditional questionnaires, such as the WHOQOL-BREF and EQ-5D, lack flexibility, fail to keep pace with the evolving conditions of patients, and are limited by self-report bias and low adherence rates. MedQ overcomes these challenges by allowing healthcare professionals to compose or adapt questionnaire domains, items, formulas, scoring methods, and alert thresholds while integrating real-time biometric data (such as heart rate, steps, sleep, and screen time) from platforms like Google Health Connect. In this article, we outline the methodology employed to develop the system, encompassing requirements definition, system modeling, backend development, and integration with health data ecosystems. We also present a comparative analysis of related work, highlighting MedQ innovations in terms of flexibility, contextual adaptability, and support for AI-driven health analytics. The MedQ system was evaluated through a focus group of 18 participants, mainly healthcare professionals, with the inclusion of experts in the Internet of Health Things (IoHT), who assessed its usability, perceived usefulness, and potential for adoption. The results showed strong acceptance, with participants recognizing its efficiency in clinical data collection, ease of use, and applicability in personalized and remote healthcare. Our findings suggest that MedQ represents a significant advancement in health assessment tools, bridging the gap between subjective evaluations and objective metrics.
Download

Paper Nr:	200
Title:	A Modular Framework for Personalisation and Remote Configuration in Rehabilitation
Authors:	César Pita, Fábio Pereira and Sergi Bermudez I. Badia
Abstract:	Serious games hold significant promise for rehabilitation due to their engaging nature. However, their clinical utility is often limited by complex systems that lack real-time, personalised adaptation. Manual adjustments frequently require interrupting sessions, which can disrupt patient engagement and flow, reducing therapeutic effectiveness. This work presents a modular, cross-platform framework designed to address these challenges. The system enables therapists to configure serious games, manage patient profiles, and trigger in-game events in real-time using a remote device. It integrates the Abstracted Personalisation Layer (APL) as an extension of the Open Rehab Initiative (ORI) platform. The framework was validated through playtesting sessions using a serious game designed to enhance working memory. It simplifies technical deployment by operating on the same local network, eliminating the need for internet access, and providing therapists with a user-friendly interface for personalisation.
Download

Paper Nr:	209
Title:	AIHEAT+: A Hierarchical Risk-Prioritized Clinical Decision Support System for Pediatric Febrile Conditions
Authors:	Abraham Bautista-Castillo, Pierce Hentosh, Ryan Ouardaoui, Chisato Shimizu, Adriana H. Tremoulet, Jane C. Burns, Ananth V. Annapragada, Tiphanie P. Vogel and Ioannis A. Kakadiaris
Abstract:	Distinguishing pediatric febrile conditions, such as Kawasaki disease, multisystem inflammatory syndrome in children, endemic typhus, and other nonspecific febrile illnesses, poses a significant clinical challenge due to the high symptom overlap and the serious risks associated with misdiagnosis. AIHEAT+, a two-stage risk-prioritized hierarchical Clinical Decision Support System, was designed to address these challenges using only eight routinely collected clinical and laboratory features available within 12 hours of a patient's presentation in the Emergency Department. AIHEAT+ implements interpretable multi-stage confidence thresholds, enabling selective abstention at each clinical decision stage. This approach optimizes diagnostic accuracy and clinical utility, as measured under multiple clinically informed risk scenarios that reflect the costs of misclassification. The dataset used in this study comprises 2,533 patients from Rady Children's Hospital and Texas Children's Hospital, with missing data handled using KNN, MICE, and MissForest imputation. Results demonstrated that AIHEAT+ consistently outperforms baseline models in utility, significantly reducing costly misclassifications, such as cross-disease confusion between Kawasaki disease or multisystem inflammatory syndrome in children, with endemic typhus. Our findings highlight AIHEAT+'s potential to improve pediatric febrile illness diagnosis through interpretable hierarchical modeling and risk-aware decision support, delivering a clinically actionable tool that enhances patient safety and diagnostic confidence.
Download

Paper Nr:	212
Title:	A Comprehensive Infrastructure and Methodology for Multi-Modal Data Acquisition to Empower AI-Based Rehabilitation
Authors:	Katerina Tzatzimaki, Nick Portokallidis, George Drosatos, Eleni Kaldoudi and Stylianos Didaskalou
Abstract:	Accurate and synchronized motion tracking is essential for advancing quantitative assessment and personalized rehabilitation. This paper presents a comprehensive infrastructure and methodology for multi-modal data acquisition designed to power AI-based rehabilitation. The system integrates multiple inertial measurements units (IMUs) and dual-camera recordings within a unified software environment that ensures reliable connectivity, synchronization and calibration. A static calibration procedure corrects sensor-to-segment misalignments, while real-time visualization enables immediate assessment of signal quality during acquisition. Although video recordings will be used exclusively for model development, their combination with IMU data will enable the creation of multimodal datasets to train AI-models that rely solely on inertial data during clinical deployment. These models aim to enhance signal accuracy by compensating for noise, drift and alignment errors. The presented infrastructure and methodology established a robust foundation for the development of AI-based rehabilitation tools to empower unsupervised rehabilitation.
Download

Paper Nr:	230
Title:	An AI-Based Approach for Foot Progression Angle Estimation Using In-Sole Pressure Sensors
Authors:	Julien Raeker, Andreas Hein and Finn Siegel
Abstract:	This study investigates the feasibility of determining the Foot Progression Angle (FPA) using machine learning (ML) algorithms based on data from pressure sensors integrated into an insole. The primary objective is to develop a measurement method suitable for everyday use, enabling continuous FPA monitoring for patients recovering from femoral shaft fractures. For this purpose, a prototype measurement system was designed and developed, incorporating eight pressure sensors in the heel area. An experimental study involving 20 subjects yielded over 36,000 usable steps for analysis. We evaluated several time-series models, including MultiROCKET-Hydra, tsfresh+LightGBM, and tsfresh+Neural Network, to predict FPA from recorded pressure data. The results demonstrate that person-specific models can predict the FPA with high accuracy. The top-performing model, MultiROCKET-Hydra, achieved a mean absolute error (MAE) of 3.23±0.614º, which is below the Minimal Detectable Change (MDC) of 4.5º for the Inertial Measurement Unit (IMU) used as a reference. In contrast, models tested for generalizability using a leave-one-subject-out approach showed significantly higher error rates, with the best model achieving an MAE of 13.174 ± 3.502º, thereby failing to provide valid results for new users. The study concludes that while a precise, person-specific FPA estimation is possible, the development of a generalizable model is currently constrained by inter-individual gait variability and sensor placement differences.
Download

Paper Nr:	231
Title:	Unlocking Collaborative Creativity: Co-Design Smartwatch-Based Serious Games for Sleep Hygiene Intervention
Authors:	Zilu Liang, Nhung Huyen Hoang, Edward Melcer and Daeun Hwang
Abstract:	Despite growing interest in digital sleep interventions for university students, many applications suffer from poor engagement due to researcher-centric design approaches that involve users only during evaluation. This study explores how co-design methods can elicit user-driven requirements for smartwatch-based serious games supporting sleep hygiene. We conducted a workshop with 45 university students who generated 60 original smartwatch game features across three game types: virtual pet, role-playing, and building games. Through affinity diagramming, six design themes emerged: personalization (customizable avatars and environments), reward (progression systems and unlockable items), punishment (playful accountability), social interaction (collaborative challenges and peer support), multi-sensory interaction (gesture controls, haptics, audio), and sleep education (contextual micro-learning). Participants demonstrated sophisticated thinking about sustained motivation, emphasizing features that balance positive reinforcement with gentle accountability, transform sleep into social activities, and leverage smartwatch-specific interaction modalities. We derive four design implications: employ layered feedback mechanisms, prioritize personalization for ownership, leverage smartwatch-specific modalities, and integrate action-centric microlearning.
Download

Paper Nr:	246
Title:	Factors Influencing Personal Health Information Sharing for Digital Health Interventions Supporting Habit Formation: A Systematic Review
Authors:	Felix Reinsch, Florentin Hitzke and Hannes Schlieter
Abstract:	Digital health interventions show great promise in promoting long-term health behaviors; however, their effectiveness is contingent upon individuals’ willingness to share personal health data. This willingness is essential, as data sharing enables interventions to be personalized and continuously refined to better suit individual needs, thus improving their relevance and effectiveness. This systematic review investigates the key factors influencing participation in health data sharing within habit-focused digital health research. Guided by the PRISMA methodology, 16 studies were analyzed, revealing 11 critical determinants-including trust in data recipients, privacy concerns, perceived risks, transparency, incentives, and user control. To synthesize these findings, a practical framework was developed, highlighting three core dimensions: personalization, trust and privacy, and incentives. Grounded in social exchange theory, the theory of reasoned action, and the theory of planned behavior, the framework offers actionable strategies for designing secure, user-centered data-sharing platforms that support sustainable health habit formation.

Paper Nr:	268
Title:	Impact of Face Anonymization on Clinical Vision Systems: A Case Study on Head-Pose–Based TWSTRS Estimation in Cervical Dystonia
Authors:	Roland Stenger, Sebastian Loens, Tobias Bäumer and Sebastian Fudickar
Abstract:	Automated vision-based clinical assessment is increasingly used to quantify motor symptoms in cervical dystonia. However, the use of facial imagery introduces privacy concerns, since the areas required for clinical evaluation also reveal personal identity. This study examines how applying face anonymization influences the accuracy of automated symptom estimation that relies on visual analysis of head posture. Several state-of-the-art anonymization approaches that aim to preserve clinically relevant visual cues while suppressing identity are evaluated in combination with a standardized head posture estimation pipeline and clinical expert ratings. Utility is measured as the agreement between automated estimates and clinician provided assessments, while privacy is quantified using face re-identification accuracy. The results show that anonymization can be integrated into such assessment workflows, but the degree to which clinically relevant information is preserved varies substantially across anonymization strategies, even when all methods are designed to maintain task re-lated appearance attributes. These findings illustrate a clear privacy and utility trade off, and highlight the need for careful selection of anonymization techniques in clinical computer vision applications.
Download

Paper Nr:	308
Title:	Decision Threshold Optimization in Machine Learning Models for Questionnaire-Based OSA Screening
Authors:	Nhung Huyen Hoang and Zilu Liang
Abstract:	In machine learning, the choice of decision threshold is a key design parameter that converts probabilistic model outputs into binary labels. Machine learning models for clinical screening typically adopt a default decision threshold of 0.5 to convert probabilistic outputs into binary classifications, yet this assumption is rarely appropriate for imbalanced medical datasets. This study systematically examines how threshold selection affects classifier performance in questionnaire-based obstructive sleep apnea (OSA) screening. Using four independent datasets (SHHS1, SHHS2, WSC1, WSC2), and three clinically relevant AHI severity cutoffs (5, 15, 30 events/hour), we analyze how key performance metrics, sensitivity, specificity, NPV, F1-score, and Matthews correlation coefficient (MCC), vary across decision thresholds. Through comprehensive visualization of probability distributions and threshold-metric curves, we demonstrate that default thresholds consistently produce unbalanced results, with 97.5% sensitivity but only 14.5% specificity at AHI cutoff 5, reversing to 96% specificity but 4% sensitivity at cutoff 30. Threshold adjustment substantially improves diagnostic balance, maintaining 70% sensitivity while significantly improving specificity to 66%. We also observed that optimal thresholds consistently align with positive class prevalence across all datasets, providing a simple heuristic without exhaustive search. These findings underscore that threshold calibration should be standard practice in machine learning based disease screening.
Download

Paper Nr:	317
Title:	Patient Journey Framework for Digital Health Technologies and Lived Experiences for People Living with MS under the Age of 30
Authors:	Pauline Gieseler, Ariel Dora Stern and Danielly de Paula
Abstract:	Multiple Sclerosis (MS) has a highly heterogeneous pathology, and people living with MS (PlwMS) perceive themes with a different relevance than Health Care Professionals (HCPs). This qualitative study proposes a framework that identifies themes and explores possibilities for applying digital health technologies (DHTs) in the care pathways for a patient population under 30, highlighting age-related differences in information seeking and future employment topics compared to other patient populations. This study points out themes that HCPs involved in the MS care pathway should be aware of. A total of 10 interviewee partners were chosen. This study interviewed 7 People living with MS and 3 Health Care Professionals. This framework extends the discovery of themes that reflect the experience of PlwMS and HCP. It brings them in the order of pre-clinical, diagnosis, Disease-modifying Therapy (DMT) decision, lifestyle adjustment, relapse, and DMT adjustment phase. This study concludes that the most important use of DHT for patients is the provision of support for evidence-based coping strategies, personalized treatment approaches, the exchange of medical records, and verified medical information. For HCPs, the most critical leverage of DHTs is the short and long-term monitoring of clinical and digital biomarkers.
Download

Paper Nr:	347
Title:	A Method for Driver Drowsiness Identification by Means of Explainable Object Detection
Authors:	Fabio Martinelli, Francesco Mercaldo and Antonella Santone
Abstract:	Drowsiness is associated (as a direct cause or contributing factor) with one-fifth of road accidents (1 in 5) and is one of the leading causes of fatal highway accidents. As a matter of fact, sleeping less than 5 hours per night increases the likelihood of having a road accident by 4.5 times. We propose a method to automatically identify if a driver is in a drowsy state, by considering a real-time object detection model to determine whether a driver is drowsy or awake. Furthermore, to provide explainability to the model prediction, we adopt an algorithm capable of visualizing the areas of greatest interest in the image of the driver under analysis, in order to understand whether the model, in addition to making the correct prediction, is looking in the right area of the driver image. In the experimental analysis, we consider a dataset consisting of 8766 annotated images of (drowsy or awake) drivers with their respective bounding boxes to demonstrate the effectiveness of the proposed method in terms of classification, localization of the face of the driver through bounding boxes, and prediction explainability.
Download

Paper Nr:	380
Title:	Toward Data-Centric Experimentation in Prehospital AI: A Customizable Pipeline
Authors:	Tamara Krafft and Bernhard Bauer
Abstract:	Early warning in prehospital emergency care is critical for improving patient outcomes. However, applying machine learning in this context remains challenging due to the nature of prehospital data, which are often incomplete, irregularly sampled, and affected by plausibility issues. This paper introduces a data-centric experimentation pipeline that treats these challenges and their mitigation strategies as explicit experimental variables rather than hidden preprocessing choices. The pipeline structures time-series experimentation into traceable stages that enable systematic assessment of how data-quality decisions influence downstream model behavior. We illustrate the feasibility of our pipeline through a representative case study on sepsis prediction involving 96,551 encounters (4.01% prevalence). The evaluation systematically varies two experimental dimensions, imputation and usage of missingness-derived features, evaluated under differing levels of miss-ingness. Selected results show that missingness is the dominant source of performance degradation, and that imputation and missingness-derived features provide complementary benefits. These findings demonstrate how the proposed pipeline enables transparent, reproducible exploration of data-quality trade-offs, laying the groundwork for broader data-centric experimentation in prehospital machine learning.
Download

Paper Nr:	382
Title:	Towards AI-Assisted Voice Therapy: The Relevance of Facial Movements and Muscle Regions for the Recognition of Oral Facial Motor Exercises
Authors:	Rica Schulze, Sabrina Schröder, Torge Jensen, Dirk Weyhe, Verena Uslar and Sebastian Fudickar
Abstract:	Objective analysis of orofacial motor exercises is relevant for digital voice therapy, yet current evaluations rely largely on subjective clinical observation. This study investigates the relevance of facial landmark features extracted from short video snippets for the automatic classification of seven therapeutic orofacial exercises using an explainable machine-learning approach based on Google MediaPipe Face Mesh. Three feature representations were evaluated: (A) mouth-region shape, frequency, and center-movement features; (B) global perlandmark motion and spectral descriptors; and (C) a combined representation integrating both sets. Feature dimensionality was reduced using SHAP-based importance ranking. The combined feature set (C) achieved the highest performance (accuracy = 0.858, F1 = 0.859), indicating that mouth-specific and global facial motion features provide complementary information. Additionally, comparable performance was achieved using only about 25 % of the original features. A SHAP-based analysis of feature set B revealed that 124 facial landmarks contributed to the best-performing model, with relevance concentrated around the mouth but extending to other facial regions. Misclassifications mainly occurred between exercises with similar movement patterns and between side-specific variants, highlighting the challenge of subtle lateral differentiation. Overall, interpretable facial motion features combined with explainable AI enable reliable classification of orofacial exercises provide a foundation for future research.
Download

Paper Nr:	385
Title:	Towards Standardized Medical Image Quality Assessment on Projection-Based Light Field Displays
Authors:	Peter A. Kara, Duy H. Nguyen, Laszlo Bokor and Tibor Balogh
Abstract:	Light field technology holds promise for glasses-free 3D visualization, enabling a smooth, continuous parallax effect for any number of simultaneous viewers. As technology emerges from research and development, light field displays may be utilized in numerous potential fields, light field displays may be utilized in numerous potential fields of profession, including medical science and healthcare in general. The quality of such visualization is studied through empirical tests involving human observers who rate and rank what they perceive. However, at the time of writing this paper, the scientific literature on subjective studies that address light field visualization in the context of medical imaging is lacking. In this paper, a standards-informed methodological framework for assessing the quality of medical imaging using light field technology is proposed. A three-phase differentiation of research efforts prior to Day 1 deployment is introduced, along with the relevant methodological foundations for phase execution and validation in future research. Each phase addresses the characteristics of light field visualization and distinguishes subjective studies based on test environment and test participants. The paper also proposes assessment-related standardization, identifies the barriers that delay research efforts, and discusses future challenges.
Download

Paper Nr:	391
Title:	A Pipeline for Automated Phenotype Extraction from Medical Reports Using Large Language Models
Authors:	Andrea Bombarda, Martina Saletta, Matteo Bellini, Lucrezia Goisis, Maria Iascone, Paolo Cazzaniga and Domenico Fabio Savo
Abstract:	Unstructured clinical narratives are a major source of phenotypic evidence for rare-disease diagnosis and ge-nomic variant interpretation. However, their free-text nature, often multilingual, heterogeneous in format, and inconsistent in terminology, makes automated phenotype extraction and interoperability with downstream genomic pipelines difficult. This creates a practical bottleneck for scalable and reproducible phenotype curation in medical genetics, where manual review is time-consuming and prone to variability. To address this problem, we propose a robust, open-source, and fully local pipeline for automatically extracting and standardizing patient phenotypes from medical reports while preserving data privacy. The pipeline integrates: (i) OCR-based digitization and an LLM-based translation module to produce an English version of the report; (ii) a GPT-oss–based phenotype extractor using structured, few-shot prompting to identify phenotypes relevant to the index patient; and (iii) a fuzzy standardization stage that combines lexical similarity with embedding-based semantic matching to map extracted phenotypes to Human Phenotype Ontology (HPO) concepts. Our multi-stage design improves robustness to real-world documentation issues, including multilingual acronyms, variable report structure, spelling errors, and synonym variability, and it ensures privacy compliance by keeping all computation on local infrastructure. We demonstrate the pipeline end-to-end on a representative clinical report, showing that it extracts patient-relevant phenotypes and produces HPO-aligned, machine-readable outputs suitable for downstream genomic analyses. This work provides a practical foundation for privacy-preserving, scalable phenotype curation in clinical genetics and supports future integration and evaluation on larger clinical datasets.
Download

Paper Nr:	414
Title:	Optimizing DL Models for Deployment on Resource-Constrained Platforms
Authors:	Priyanka Chauhan and Kolin Paul
Abstract:	Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with an increased risk of stroke and heart failure, necessitating accurate and timely detection. Continuous cardiovascular monitoring using wearable devices provides a promising approach for early AF identification. This work presents a comparative study of four deep learning models created with building blocks such as-long short-term memory (LSTM), CNN with residual links, channel-attention–based LSTM, and NeuralODE–LSTM-for AF detection using unidimensional 12-lead ECG signals from 10,586 subjects in the Chapman University dataset. The models are trained using varying dataset sizes to estimate the data requirements for effective end-to-end learning. Performance is evaluated using accuracy, precision, sensitivity, recall, and F1-score, averaged over five-fold cross-validation. In addition, model size, parameter count, and inference latency are analyzed on x86 and simulated ARM7 platforms, with an emphasis on computational efficiency for embedded deployment. Hardware-level profiling is further conducted using the gem5 simulator, targeting ARM-based microcontrollers such as the STM32L4. Experimental results demonstrate that the ResNet model consistently achieves superior classification performance across all data sizes, while also exhibiting the lowest inference latency on both x86 and simulated ARM platforms, indicating its suitability for deployment on wearable and edge devices.All models are further compressed using suitable quantization techniques, and their accuracy–memory trade-offs are analyzed.
Download

Paper Nr:	418
Title:	Supporting Medicines Manufacturing through Semantic Technologies: Oligonucleotide Synthesis Use Case
Authors:	Moulham Alsuleman, João Gregório, Nina Perić, Jean-Laurent Hippolyte, Sean Ruane and Michael Chrubasik
Abstract:	Pharmaceutical manufacturing generates heterogeneous, siloed data across instruments and process steps, hindering interoperability and complicating compliance. A proof-of-concept ontology and knowledge graph for oligonucleotide manufacturing were designed with a modular hub-and-spoke architecture that reuses a shared core and constrains platform-specific extensions like analytical and synthesis ones. The graph integrates time-series sensor data and analytical outputs as semantic triples, using W3C standards for observation, provenance, and units along with domain knowledge aligned with manufacturing and biomedical ontologies. An extract-transform-load pipeline transformed raw instrument data files into ontology-compliant triples. The resulting knowledge graph (over 2.4M triples) encodes instrument platforms, process steps, and observation results with lineage, operators, and unit metadata. Validation used competency questions defined by domain experts; queries answered these and linked synthesis parameters to downstream analytical results, replacing manual harmonisation across spreadsheets. The approach demonstrated improved traceability and reduced manual effort, providing a structured foundation for deviation analysis, batch comparison, and decision support. FAIR implementation was assessed at Level 3 (’Pretty FAIR’) using the Pistoia Alliance FAIR Maturity Matrix. The workflow and architecture are replicable and scalable. This work illustrates a practical path to machine-actionable manufacturing data that strengthens compliance while providing a foundation for AI applications and FAIR’s machine-readability goals.
Download

Paper Nr:	446
Title:	A Deep Learning Approach for Deep-Vein Thrombosis Detection in Ultrasound Videos
Authors:	Jessé Ferreira do Nascimento Filho
Abstract:	Thrombosis is a major contributor to the global burden of disease, underlying several life-threatening cardiovascular conditions. Early and accurate detection of deep-vein thrombosis could help achieve better clinical outcomes. Even though, ultrasound-based diagnosis remains challenging due to operator variability and heterogeneous acquisition protocols. The aim of this work was to propose a deep-learning approach for thrombosis detection ResNet and Mixture-of-Experts (MoE) architecture. A dataset of 2,341 ultrasound videos was processed to combine frames with expert annotations. The datasets were resized and frame count were limited in order to be compatible with pre-trained convolutional models. ResNet backbone was used to leverage transferable representations of low-level feature relevant to ultrasound imagery. Then, we introduce MoE architecture guided by ground-truth anatomical site annotations, to enforce site-specific expert specialisation during training. At inference, soft gating is applied, allowing the model to first predict anatomical site probabilities, then combine expert outputs to estimate thrombosis risk. At the validation set, we achieved 0.2314 of log-loss, an AUC-ROC close to 0.8, 83.7% of sensitivity, and about 60% of specificity. For a test set of 578 videos, the model maintained stable calibration with 0.2909 of log-loss. These results reveal the usefulness of using deep learning as an operator-assistive tool for anatomical site identification and thrombosis detection.
Download

Short Papers

Paper Nr:	21
Title:	Enhanced Depression Detection through Sleep Data Analysis Using an Ensemble Model
Authors:	Sarah Nasir, Ayesha Seerat, Muhammad Wasim, Paulo Jorge Coelho, Ivan Miguel Pires and Nuno M. Garcia
Abstract:	Depression is a mental disorder characterized by persistent sadness, fatigue, sleep disturbances, and difficulty performing daily activities. The assessment of depression is traditionally based on self-reported data from patients. However, these reports are often subject to bias due to individual perceptions and interpretations of their experiences. With the increasing use of wearable devices for sleep monitoring, leveraging this data presents us with an opportunity for more efficient and cost-effective depression detection. To address the limitations of existing research, which often relies on detailed reports that lack generalizability and are susceptible to noise, we propose a systematic ensemble classification model for depression based on sleep data, utilizing participant-level cross-validation. Our approach achieves an improvement of 2.18% over the baseline. Specifically, we employ ensemble learning to improve the generalization and robustness of the model. We used Depresjon dataset for evaluation, and our model achieves an accuracy of 90.0%, demonstrating its effectiveness in identifying depression. Additionally, we applied SHAP (Shapley Additive Explanations) for explainability, revealing the most influential features for detection.
Download

Paper Nr:	24
Title:	A Methodology Based on Density Estimation to Model and Compare Hospital Populations in the Context of Antibiotic Resistance
Authors:	Antonio Lopez-Martinez-Carrasco and Denisse Kim
Abstract:	The digitalisation of medical data allows interoperability between medical systems, providing a valuable opportunity to perform large-scale analysis or reproduce clinical studies. However, working with large volumes of inter-hospital data presents significant challenges due to the complexity and variability of datasets from different sources. This is a relevant limitation in medical research, where it is necessary to have mechanisms to model and compare hospital populations to address certain critical problems. Machine learning offers innovative solutions to address this challenge. Therefore, this paper introduces DBPA (Distribution-Based Population Analysis), a five-step methodology based on the density estimation technique to model and compare hospital populations based on their underlying distribution. The experiments carried out in this work show the suitability of our proposal in a real use case involving populations from different Intensive Care Units with antibiotic-resistant Enterococcus bacterium. The results demonstrate the effectiveness of the methodology and its potential to support clinicians’ decision-making.
Download

Paper Nr:	29
Title:	A Proof-of-Concept Study of an AI-Supported App for Personalized Non-Pharmacological Dementia Care
Authors:	Clara Ribeiro Neves, Ana Carolina Monteiro Reina, Alexandre Bruno Brandão Esmerado Cavaleiro, Susana Isabel Gomes Lopes and Marlene Cristina Neves Rosa
Abstract:	This proof-of-concept study presents SMILIGHT, an AI-supported prototype designed to assist caregivers in delivering personalized non-pharmacological interventions for people with dementia. Developed using a user-centred design approach, the prototype integrates biometric data visualization and placeholders for AI-generated recommendations to support decision-making. Fifteen participants, including healthcare professionals, caregivers and researchers, evaluated the system through a flow questionnaire, the System Usability Scale (SUS), and a SWOT analysis. Results showed high usability (mean SUS score: 75.0) and excellent internal consistency for the custom questionnaire (Cronbach’s α = 0.9205). Key features such as manual prescription and patient data access received particularly high satisfaction scores. While participants highlighted the system’s potential to support dementia care, feedback pointed to necessary improvements in accessibility and interface clarity, especially for non-technical users. The findings validate the feasibility of SMILIGHT and support its continued development and real-world testing, with a focus on enhancing usability and personalization. This study underscores the value of involving end-users in the early design of digital health tools to improve adoption and effectiveness.
Download

Paper Nr:	37
Title:	Real-World Medication Adherence Monitoring Using Hybrid Neural Networks and Smartwatches
Authors:	Mehreen Shah and Ali Imran
Abstract:	Medication non-adherence remains a critical healthcare challenge, contributing to poor clinical outcomes and increased healthcare costs. While multiple factors underlie non-adherence, the lack of effective and unobtrusive monitoring solutions limits the ability to provide timely interventions. To address this gap, we present a pill-taking gesture recognition framework that leverages smartwatch inertial data as a scalable and non-intrusive sensing modality. A hybrid deep learning architecture is introduced, combining convolutional layers, squeeze-and-excitation blocks, attention mechanisms, and long short-term memory units to jointly capture spatial and temporal dependencies inherent in multi-step pill-taking behaviors. The framework is evaluated using a free-living dataset under a leave-one-subject-out cross-validation setting, ensuring robust assessment in real-world scenarios. Results show strong recognition performance, achieving a precision of 0.92 with an optimal window length of 9 seconds. Furthermore, a lightweight domain adaptation strategy enhances inter-user generalization, delivering up to 19% improvement in cross-user evaluations. Extensive experiments across different window sizes and demographic groups confirm the model’s robustness, scalability, and fairness. These findings demonstrate the potential of wearable-based gesture recognition for real-time, privacy-preserving monitoring of medication adherence, paving the path for the scalable deployment of such systems in both clinical and home environments.
Download

Paper Nr:	43
Title:	Is a Music Game with a Piano Controller Effective in Improving Motor Coordination in Stroke Survivors? A Case Study
Authors:	Yasmim Fernandes Moniz, Higor Barreto Campos, Silvia R. M. S. Boschi and Luis Duarte Ferreira
Abstract:	Stroke is one of the leading causes of long-term disability, making rehabilitation essential for patient recovery. Despite advancements in rehabilitation techniques, many patients face challenges in maintaining engagement, mainly due to the need for repetitive exercises. Serious games, designed to increase motivation and provide engaging therapy, have shown potential in improving motor function. Moniz Game is a music-based game that stimulates motor coordination but has not yet been tested with stroke patients. This case study aims to evaluate whether the Moniz Game, combined with a piano controller, can improve upper limb motor coordination in stroke survivors. while also assessing the system’s usability from the patient’s perspective. The participant, a 56-year-old woman who suffered a stroke affecting her left upper limb, underwent a three-week intervention with bi-weekly sessions. Motor coordination was assessed using the Box and Blocks, Nine-Hole Peg, Tapping, and Tapping Digital Test. Results showed improvements in both hands, with notable progress in motor coordination. Usability was rated as ”Best Imaginable” on the System Usability Scale (SUS). These findings suggest that the Moniz Game, combined with the piano controller, could be an effective and engaging rehabilitation tool, showing promising results in both motor coordination and user satisfaction.
Download

Paper Nr:	44
Title:	Usability and Acceptability Study of a Cognitive Behavioral Therapy-Based Virtual Reality Application for Treatment of Stimulant Use Disorder
Authors:	Bethany K. Bracken, Dan T. Duggan, Alyssa Pigott, Brianna Kocon and Lisham Ashrafioun
Abstract:	Stimulant Use Disorders (StUD; methamphetamine, cocaine, other amphetamines) is associated with psychiatric and cardiovascular morbidity, infectious disease, crime, and housing insecurity. Cognitive-behavioral (CBT) approaches to StUD treatment, embedded into virtual reality (VR) could provide a robust supplement to current practices. In the current study, we designed a VR platform to be used as an adjunctive treatment for StUD, Stimulant Use Recovery Via Immersive Virtual Environments (SURVIVE) and evaluated its safety, usability, and acceptance by the target population. Participants (n = 15) included adults currently in treatment for StUD. We categorized errors in ascending severity from minor usability to safety violations. There were no safety violations by any participants providing evidence that the app is low-risk and safe to use. We identified multiple software bugs we are now addressing. Participants indicated they enjoyed and would like to continue using SURVIVE. Participants did not experience significant symptoms of simulator sickness. Our findings support the usability of SURVIVE.
Download

Paper Nr:	47
Title:	Creating an Adapted Piano Controller Associated with a Music Game for Stroke Survivors: The Evolution of a Prototype in Collaborative Design
Authors:	Yasmim Fernandes Moniz, Higor Barreto Campos, Silvia R. M. S. Boschi and Luis Duarte Ferreira
Abstract:	Stroke rehabilitation often relies on repetitive tasks to promote neuroplasticity and motor recovery, which can lead to patient disengagement due to monotony. Introducing interactive and motivating gamified systems into physiotherapy is a strategy to address this challenge. This study presents the collaborative design and iterative development of a piano controller integrated with a music game to enhance stroke survivors’ rehabilitation. The system combines motor stimulation with engaging musical training, leveraging physical interaction through a custom 3D-printed piano controller and gamified elements to sustain motivation through usability tests with 35 healthy individuals and 3 stroke survivors, feedback-guided design, connectivity, and functionality refinements. The final prototype demonstrates portability, adaptability, and user-centric features, making it viable for clinical and home-based applications. Future work will focus on clinical trials to evaluate the system’s therapeutic impact and explore manufacturing for broader accessibility.
Download

Paper Nr:	57
Title:	An EHR Data Standardisation Pipeline Using the MIMIC-III Dataset: Foundation for a Clinical Trial Data Quality Assessment
Authors:	Joáo Gregório, Agnieszka Lemanska, Bartlomiej Cieszynski, Nina Perić, Moulham Alsuleman, Michael Chrubasik and Paul Duncan
Abstract:	Demographic representativeness in clinical trials is essential to ensure treatments are applicable to all patients and that everyone can benefit equally. Clinical datasets, such as MIMIC-III, offer valuable opportunities for secondary research, but their fragmented structures and unstandardised content pose significant challenges for quality assessment. This paper presents a reproducible data processing pipeline designed to prepare electronic health record data for evaluating the demographic representativeness of clinical trial populations. The pipeline consolidates demographic data from multiple sources within the dataset and maps unstandardised diagnostic terms to ICD-11-compliant terms by using WHO APIs and an enhanced synonym dictionary of clinically relevant diagnostic terms to improve term-matching. It extracts disease-specific cohorts for validation. When applied to the dataset, the pipeline successfully combined patient data from multiple tables and achieved 93.92% frequency-weighted diagnostic mapping coverage. A sepsis cohort was extracted to demonstrate the ability to generate well-characterised target populations with complete scoring across key demographic features. The resulting, unified, and standardised dataset supports the future development of objective quality scores that support transparent trial evaluation and informed decision-making. By aligning with FAIR data principles and established data quality frameworks, this work contributes to broader efforts in data governance and interoperability in clinical research.
Download

Paper Nr:	61
Title:	Developing Psychological Tendency Prediction Models Using Smartphone Games
Authors:	Anna Goshima, Eriko Sugisaki, Tomohiko Kiriyama, Miyuki Nakamura, Nao Kobayashi, Hiroyoshi Ogishima, Saori C. Tanaka and Keiji Yasuda
Abstract:	Recent years have seen a surge in the development of services that predict mental health conditions using digital biomarkers. In this study, we attempted to examine the utility of estimating mental tendencies using game log data. To build prediction models, we utilized game logs, smartphone logs, and emotion questionnaires. Participants played a smartphone game once per day, and we collected records of their in-game actions, such as the number of jumps and the number of trials. At the start of the data collection experiment, approximately 100 participants completed 17 psychological scale assessments, and log data was then collected from these participants for three months. Initially, we performed factor analysis using the total scores from these psychological scale assessments. Then, we built a binary classification model using game logs, smartphone logs, and emotion questionnaires to predict the factor scores for three common factors. For the first factor, we found that the best prediction performance was achieved when using the combined smartphone logs from both weekdays and weekends and the game logs recorded at participants’ homes. For the second factor, the best prediction performance was achieved when using the combined smartphone logs from both weekdays and weekends and the three-month average of the game logs. These results suggest that incorporating game logs alongside smartphone logs can enhance the performance of the prediction model designed to assess psychological tendencies.
Download

Paper Nr:	62
Title:	Formal Concept Analysis Applied to Characterize Breast Cancer Related Lymphedema Symptoms
Authors:	Mateus Barbosa, Julio Neves and Mark Song
Abstract:	Lymphedema is a common sequela of breast cancer treatment, and early diagnosis is essential to prevent progression to severe stages. This study applied Formal Concept Analysis (FCA) to data from 286 breast cancer survivors, divided into a lymphedema group (n=67) and a control group (n=219), to identify symptoms that reliably characterize the condition. FCA generated numerous association rules, from which two distinct patterns emerged: (i) conflicting symptoms such as pain and tingling, present in both groups and therefore not specific to lymphedema; and (ii) exclusive symptoms occurring only in the lymphedema group. Among the latter, arm/hand swelling (65% support, 100% confidence) and seroma (69% prevalence across severity levels) stood out as the most consistent and clinically meaningful diagnostic markers. These findings confirm known clinical observations and demonstrate FCA’s ability to distinguish general treatment side effects from symptoms truly indicative of lymphedema. FCA thus provides a transparent, rule-based framework to support early detection and the design of predictive tools, ultimately contributing to better management and prevention of advanced disease stages.
Download

Paper Nr:	75
Title:	Patient History Visualization for Structured Medication Reviews: A Design Study
Authors:	L. Hama, R. A. Ruddle, A. S. Abuzour, M. Abaho, A. Aslam, D. Bollegala, A. Clegg, M. Gabbay, F. S. Mair, M. O'Connell, S. Relton, E. Shantsila, M. Sperrin, S. A. Wilson, A. A. Woodall, I. Buchan and L. E. Walker
Abstract:	General practitioners and pharmacists conduct Structured Medication Reviews (SMRs) to optimise prescribing for people with multiple long-term conditions (MLTC), but electronic health record systems often present information in fragmented lists and tabs. We set out to design and validate chart-based visual summaries of patient history data that can support integrated dashboards for SMRs. Using a design-study methodology, we reviewed existing approaches to visualising electronic health records, conducted four mock SMRs, and derived a data abstraction linking clinicians’ questions to patient attributes. Using visual encoding principles, we used 14 candidate chart types and sketched low-fidelity designs. A questionnaire was then used to ask eight clinicians and eleven visualization-literate researchers to rate how effectively each chart communicated its data. Across six combinations of data types, 11 chart types were consistently judged suitable for communicating key information. Finally, we implemented the outcomes of the study in a Python package using 2.1M records from the Clinical Practice Research Datalink (CPRD).
Download

Paper Nr:	77
Title:	Boosting 30-Day Emergency Department Readmission Predictions with BERT-Powered Community Health Worker Notes
Authors:	Ankita Mishra, Charmi Patel, Naveen Kumar Veeramreddy, Yiyang Wang, John Mazzeo, Jamshid Sourati, Kelly McCabe, Roselyne Tchoua, Jacob Furst and Daniela Raicu
Abstract:	Hospital readmissions pose a significant challenge to healthcare systems, impacting both care quality and costs. Predictive models for readmission rely primarily on structured Electronic Health Records (EHRs) such as checkboxes and coded fields, while often overlooking valuable information in unstructured notes. This exclusion reduces model sensitivity and limits their ability to identify high-risk patients. We propose an approach that integrates structured data with BERT-based embeddings of concatenated unstructured notes as input to classifiers to predict 30-day hospital readmissions. Using a multilayer perceptron (MLP) classifier trained on structured-only, unstructured-only, and integrated data, we show that across all patient records, the structured-only model achieved high specificity (99.5%) but very low sensitivity (1.4%). In contrast, the unstructured model substantially improved sensitivity to 65.9%, though at the cost of reduced specificity (57.9%). The integrated data model provided the best overall trade-off, with a sensitivity of 63.1%, specificity of 64.3%, and an AUROC of 0.687. These findings show that the integrated data model captures signals of readmission risk most effectively, offering the best trade-off in predictive performance. Our study underscores the importance of incorporating unstructured community health worker notes into hospital readmission models to enhance early risk identification and reduce the likelihood of readmission.
Download

Paper Nr:	85
Title:	From Conversation to Clinical Decision Supported Assistance: A Secure, Personalized Platform for Elder Care
Authors:	Ariane Méndez, Maia Aguirre, Arantza del Pozo, Teresa García-Navarro, Manuel Torralbo and Janeth Carreño
Abstract:	This paper presents the design, development and validation of a personalized conversational platform for the remote assessment and monitoring of elderly dependents. Unlike cloud-based solutions, our platform supports on-premises deployment, ensuring enhanced data privacy and compliance with regulations. The system features a personalized conversational assistant that engages care-receivers through calendar events, questionnaires and reminders, with data seamlessly feeding into a Clinical Decision Support System (CDSS) for caregivers. The validation study, conducted in a real-world senior care environment, demonstrates the platform’s feasibility and effectiveness. Health and social care professionals rated the CDSS with an average System Usability Scale score of 93.33, confirming its high usability. Care-receiver feedback for the conversational assistant was more nuanced, with a mean Chatbot Usability Questionnaire score of 66.5. While its ease of use was well-received, feedback highlighted opportunities to improve its conversational capabilities. Overall, the platform shows great promise in enhancing both patient engagement and clinical monitoring for elder care.
Download

Paper Nr:	96
Title:	Play and Care: How Mini-Games Enhance Informal Caregiver Well-Being
Authors:	Beatriz Jardim Peres, Hildegardo Noronha, Joana Câmara, Élvio Rúbio Gouveia and Pedro Campos
Abstract:	Informal caregivers are crucial in supporting individuals with chronic illnesses, disabilities, or aging-related conditions. The demands of caregiving can lead to significant physical, emotional, and financial difficulties, which can negatively affect their mental health and quality of life. This study investigates the impact of virtual reality mini-games on the well-being of informal caregivers by comparing the effects of playing chosen games versus random games. By providing caregivers with the autonomy to choose their preferred games, we learned that participants increased valence and that this effect is stronger in the chosen games but has no measurable effects on the other studied values. Participants’ insights about the chosen gaming experience corroborate the increase in valence, with most subjects reporting having fun and other positive comments. The effect of choice can also be seen when the subjects communicate that they actively look for easier or harder games, depending on their preferences.
Download

Paper Nr:	103
Title:	Evaluation of a Portable Low-Cost Depth Camera System for Diagnostic Processes in Gait Analysis
Authors:	Tabea Wollenweber and Klaus Brinker
Abstract:	Precise positional data of the human body is essential for clinical gait analysis. Data collection through marker-based methods is labor-intensive, costly, and the complex setting can lead to unnatural alterations in gait. In this paper we investigate the applicability of a low-cost depth camera system for gait analysis by comparing its output to the data provided by a high-precision marker-based motion capture system. The impact on detection accuracy of different recording conditions, including walking direction and the type of clothing worn, were tested. In addition, simultaneous gait recordings with both systems were conducted to allow for a comparison of positional data. The results reveal that a diagonal walking path and no trousers, produces the most reliable results. Although the depth camera system does not match the precision level of the motion capture system, particularly in capturing absolute values, it reliably captures temporal and dynamic movement patterns, making it suitable for identifying certain gait phases, analyzing general movement trends, or detecting gait asymmetries. But for clinical diagnostics where exact absolute values are crucial, the depth camera system cannot replace high-end marker-based systems.
Download

Paper Nr:	119
Title:	Analyzing the Mental Well-Being of Adolescents During the COVID-19 Pandemic through Formal Concept Analysis
Authors:	Bernardo Amorim, Elaine Passos, Julio Neves and Mark Song
Abstract:	This study applies Formal Concept Analysis (FCA) to a dataset on adolescent mental health (ages 12–17) collected during and following the COVID-19 pandemic to identify behavioral and psychological patterns associated with pandemic-related distress. The data were obtained through ecological momentary assessments (EMA) and patient-reported outcome measures, comprising 399 adolescents at baseline and 113 participants contributing a total of 637 follow-up responses. The FCA approach enabled the identification of conceptual clusters and association rules that describe the relationships between social isolation, lifestyle changes, and emotional well-being. The results indicate that adolescents experiencing greater social isolation and reduced physical activity exhibited stronger associations with symptoms of anxiety, loneliness, and decreased academic performance. These findings highlight the need for integrated support strategies that combine psychological counseling, peer-support initiatives, and time-management interventions. Overall, this research demonstrates the value of explainable data-mining methods such as FCA for uncovering interpretable patterns in behavioral health data, providing insights that can inform evidence-based policies and preventive strategies for adolescent mental well-being in future public health crises.
Download

Paper Nr:	120
Title:	Agentic AI in Healthcare: Architectural Model and Call to Action
Authors:	Nabil Georges Badr
Abstract:	Agentic artificial intelligence (AI) marks a paradigm shift in healthcare-moving beyond reactive automation toward autonomous, adaptive systems capable of planning, reasoning, and orchestrating multi-agent workflows. Despite growing interest, agentic AI remains conceptually ambiguous and operationally fragmented, posing challenges for clinical integration, ethical governance, and stakeholder trust. This paper synthesizes 20 peer-reviewed sources through thematic analysis to examine the definitional foundations, clinical applications, governance gaps, and human-centered design imperatives of agentic AI in healthcare. To operationalize these insights, we propose a layered architectural model followed by an illustrative use case representing how this architecture would coordinate chronic disease management across primary care, pharmacy, and home health settings. By bridging academic insight with operational relevance, this paper offers a roadmap for safe, equitable, and scalable adoption of agentic AI across healthcare systems.
Download

Paper Nr:	123
Title:	Designing an Effective Food Allergy Management App
Authors:	Adriano Galati
Abstract:	Food allergy (FA) is a growing public health concern with substantial impacts on patient safety, quality of life, and healthcare utilization. Globally, over 220 million people must avoid certain foods due to allergies, and mobile digital applications hold great potential to help prevent, identify, and manage food allergy reactions. The European Academy of Allergy and Clinical Immunology (EAACI) and the American Academy of Allergy, Asthma, and Immunology (AAAAI) recognize the advent of the mobile health technology era in medicine and actively support their development. The creation of FA specific apps offers considerable opportunities to support patients and caregivers in an efficient and cost-effective way. However, most of the available FA apps have not been proven very effective, and adoption remains low despite high general engagement with mobile applications. This highlights unmet needs in digital allergy management. This study aims to investigate FA patients’ and caregivers’ experience to inform the design of an effective FA management app, through a participatory process involving patients with food allergy, caregivers, immunologists, and computer scientists.
Download

Paper Nr:	128
Title:	Managing Energy Levels in Academia: Expanding the Conversation on Long COVID Pacing Technologies
Authors:	Audrey Girouard, Raphaëlle N. Roy and Shanel Wu
Abstract:	This report shares the experiences of two researchers who live with Long COVID (LC). LC is a chronic condition where COVID-19 symptoms linger for over 3 months. Like many other chronic illnesses and post-infectious syndromes, LC limits a person’s energy and managing the condition often focuses on “pacing”, i.e. regularly reducing activity levels to avoid post-exertion malaises. Pacing is difficult to implement, and so is the general recognition and management of LC symptoms. We report our experience with tools to help manage energy levels in an academic work context. We reflect on the recent work on pacing technologies and contribute our perspectives to advance HCI’s understanding of Long COVID and other complex, fluctuating chronic illnesses. As disabled scholars, we advocate for technology that supports pacing, an emerging topic in HCI and accessibility spaces.
Download

Paper Nr:	138
Title:	Chatbot Application Based on a Natural Language Model to Support the Improvement of Hemoglobin Levels in Anemic Preschool Children
Authors:	Natalia Cabanillas Chamochumbi, Marcelo Scerpella Zarkovic and Víctor Manuel Parasi
Abstract:	This study presents and validates a GPT-4–based nutritional chatbot designed to generate personalized meal plans aimed at improving hemoglobin levels in anemic preschool children. A 12-week pre–post study was conducted with 16 children in Lima, with hemoglobin levels increasing from 10.31 ± 0.30 to 10.61 ± 0.31 g/dL (Δ = +0.31 ± 0.20; p = 2.11×10⁻⁵). This increase is clinically meaningful, as it reflects measurable improvement within a short, non-pharmacological intervention. 81% of participants showed Hb improvement, and 19% reached increases ≥0.5 g/dL. Additionally, 192 AI-generated plans were compared with nutritionist-crafted references. Slight bias and manageable errors were observed in iron, vitamin C, and B12 content, with good overall agreement. These findings suggest that AI-assisted meal planning, validated by professionals, can complement nutritional care and contribute to reducing childhood anemia.
Download

Paper Nr:	142
Title:	Towards Personalized and Continuously Adaptive Cognitive Training with the Reh@Sync: A Pilot Study Series
Authors:	Teresa Paulino, Joana Câmara, Diogo Branco, Luís Ferreira, David Mata, Fernando Teixeira, Ana Lúcia Faria, Sergi Bermúdez Badia, Eduardo Fermé and Mónica S. Cameirão
Abstract:	The use of interactive technologies within cognitive rehabilitation interventions following acquired brain injury, such as stroke, has rapidly evolved worldwide. The digital aspect allows the implementation of intelligent processes, enabling personalization and adaptation of the difficulty to the user’s characteristics. This work explores the implementation of supervised machine learning models trained on participants’ performance data to adjust exercise difficulty within an integrative training system, the Reh@Sync. The process involved two pilot studies with participants presenting cognitive impairments. The first pilot study, with 15 participants, aimed at collecting initial data on manual parametrization performed by the therapists. This data was then used to generate fitted models for implementing automatic parametrization in the system. The efficacy of the automation in terms of performance was evaluated in the context of a second pilot study with four participants. Results revealed that the system was able to personalize and continuously adapt the cognitive training program, maintaining participants’ performance within the range set by the therapists.
Download

Paper Nr:	143
Title:	Advancing Fairness in Clinical AI Decision-Making through a Sociotechnical Threshold Bias Audit
Authors:	Dong-Gil Ko
Abstract:	Artificial intelligence (AI) models are increasingly used in health care for risk stratification and early warnings. However, if biased, these models can underperform for certain patient groups, leading to unequal care and missed opportunities. This study audited a clinical prediction model for adverse cardiovascular events and identified significant disparities in recall across race, gender, and age. At the original alert threshold, recall for female patients was 27% versus 44% for males; for Black patients, 31% versus 39% for White patients. To address these gaps, subgroup-specific threshold recalibration was applied, lowering the alert cutoff for underserved groups. This intervention substantially improved sensitivity (by ~13–14 percentage points) for the targeted groups with minimal performance loss elsewhere. Statistical testing confirmed the gains were significant (McNemar’s p < 0.05). Through stakeholder consultation, the trade-off – more false positives in exchange for identifying more at-risk patients – was deemed acceptable. The study outlines how fairness audits and threshold adjustments can be implemented within clinical decision support systems using existing governance structures. Findings demonstrate a scalable strategy to reduce algorithmic bias in predictive models, supporting equitable care delivery and aligning with emerging standards for responsible AI in medicine.
Download

Paper Nr:	149
Title:	Development and Validation of a Domain-Specific Portuguese Chatbot for Pediatric Cardiology Literacy
Authors:	Joana Luis, Federico Guede-Fernandez, Mariana Dias, Sergio Laranjo, Miguel Santos, João Moura Pires, Eduarda Oliosi and Ana Londral
Abstract:	Understandable and easy to access information is essential for positive outcomes in healthcare. However, this clarity is often lacking, especially when it comes to complex areas such as pediatric cardiology. Patients and their families often struggle to understand medical jargon or find the necessary information in their native language. The proposed study aims to improve the access to this information and its quality, by developing a domain-specific chatbot in Portuguese, for pediatric cardiology, with a combination of Large Language Models (LLM) and Retrieval Augmented Generation (RAG), using medical documentation as support. The results show that focusing on a RAG-approach and on the retrieval of the correct information improved overall performance. The best performing configuration, SERAFIM embeddings, Portuguese prompt and the Mistral 7B model, achieved a BERTScore of 0.75 and a Hit Rate of 87% for the SERAFIM embeddings. This study highlights the potential of RAG-based approaches to develop trustworthy and accessible chatbots in the healthcare sector, ensuring accurate and reliable information is available to patients.
Download

Paper Nr:	150
Title:	Augmented Reality-Based Serious Gaming for Neurorehabilitation in Adults: A Preliminary Approach
Authors:	Eliana Silva, Miguel Teixeira, Luís Paulo Reis, Renata Ocampo, Ana Ramos, Marta Parreira, Marta Reis, Catarina Fernandes and Sara A. Silva
Abstract:	Acquired brain injury (ABI) frequently leads to long-term impairments in various domains of daily functioning. This work presents an augmented reality (AR) platform for the Meta Quest 3 that integrates psychosocial, cognitive, and motor rehabilitation within a single, immersive system. Set in a mixed-reality supermarket, the platform uses everyday tasks as therapeutic activities in the form of three minigames tailored to specific domains: 1) Customer Care for training social interaction; 2) Memory Aisle for training attention and memory; and 3) Stock Shift for training motor coordination through object manipulation. The system uses passthrough AR, spatial anchors and hand tracking to enable intuitive interaction, while adaptive difficulty ensures personalized progression. Accessibility features, including text-to-speech, gaze-assisted placement and object loss prevention, address the diverse needs of patients. Insights from sessions with a panel of healthcare professionals informed the content and design of the engagement and feedback mechanisms. A preliminary pilot study involving two healthcare professionals provided evidence that the resulting platform offers a safe, engaging and ecologically valid rehabilitation experience. By integrating multiple therapeutic domains into a single AR framework, this work demonstrates the potential of serious games to support more holistic and effective rehabilitation solutions.
Download

Paper Nr:	154
Title:	Musiquence AI Toolset: An AI-Powered Serious Game Creation Ecosystem for Health Professionals
Authors:	João Cupido, Luís Ferreira and Sergi Bermúdez Badia
Abstract:	Therapies, especially cognitive stimulation and reminiscence therapies, aim to improve an individual’s quality of life, as well as improve cognitive and social functions, through the use of various tools. Due to technological progress, including artificial intelligence (AI), there has been a revolution in the application of these tools. This evolution has influenced therapies, and can even make diagnoses with greater precision and generate personalised content. With Musiquence Toolset, the integration of AI into the Musiquence platform will allow users to have an easier experience in the process of cognitive stimulation and the insertion and use of content as memories to stimulate the memory of individuals with dementia. Therefore, this document addresses the literature review, as well as the relevant features and requirements. Additionally, it also presents system design and architecture of the app, implementation and execution of usability tests and, consequently, analysis of obtained results, and a general retrospective on the implementation of a set of systems for Musiquence with AI.
Download

Paper Nr:	162
Title:	Predictive Modelling of Tuberculosis with Environmental Factors in São Paulo, Brazil
Authors:	Roberto Eyama and Marcia Ito
Abstract:	Tuberculosis remains a global public health challenge, with Brazil being one of the countries with high disease rates. This study investigates the relationship between TB incidence in São Paulo and key environmental factors. Using data from 2001 to 2024 obtained from SINAN (disease notifications), CETESB (environmental data), and IBGE (population data), we developed and compared time series models to predict TB dynamics. The optimal model was identified as a SARIMA(2,1,1)(1,1,1)₁₂ applied to a log-transformed series to stabilize variance, which achieved a Mean Absolute Percentage Error (MAPE) of 12.58% on the test set. To enhance the predictive power and investigate the relationship with environmental factors, we developed a hybrid SARIMA-NARNNX model that incorporates the residuals of the SARIMA model with selected exogenous variables. The best-performing hybrid model included sulfur dioxide (SO₂), temperature, humidity and PM10. This model demonstrated superior performance, achieving a MAPE of 9.63%-a 23.5% improvement over the log-SARIMA model. Furthermore, long-term projections to 2035 using the model forecast an alarming increase in the TB incidence with a trajectory that significantly deviates from the reduction targets set by the WHO End TB Strategy.
Download

Paper Nr:	164
Title:	Deep Models for Blood Pressure Estimation: A Trade-Off between Accuracy and Efficiency
Authors:	Mohammad Ehsanul Alim, Vishal Singh Roha and Mehmet R. Yuce
Abstract:	Cuffless blood pressure (BP) estimation using physiological signals has become an emerging frontier in applied sensing and digital health. In this work, we present a comprehensive benchmarking of deep learning architectures for non-invasive BP estimation using spatio-temporal features derived from Electrocardiogram (ECG) and Photoplethysmogram (PPG) signals. MIMIC-II database, comprising 942 participants having 12,000 segments, has been used in this study. Four representative models are investigated: a traditional Convolutional Neural Network (CNN), a Vision Transformer (ViT), a hybrid 3D-CNN + ViT, and a ResNet-50 integrated with attention mechanism. Performances are analyzed through statistical and computational metrics and validated against the different medical standards. Among the compared architectures, the joint 3D-CNN + ViT configuration achieved the highest association and accuracy (r = 0.93, MAE = 4.91 mmHg), while the traditional CNN delivered the fastest execution with competitive accuracy (r = 0.86, MAE = 5.18 mmHg).These results clarify the accuracy–efficiency, trade-off across model families and offer guidance for real-time BP sensing.
Download

Paper Nr:	169
Title:	Generating Synthetic Behavioral Health Data: A Multi-Agent LLM System
Authors:	Dirk Maas, Maani Beigy, Laura Genga and Pieter Van Gorp
Abstract:	Mobile health (mHealth) tools offer a scalable solution for delivering behavior change interventions. In recent years, personalization has been increasingly adopted in mHealth interventions. However, developing personalized models requires user data, which is often scarce or unavailable for new interventions. This problem is often addressed by creating health behavior personas, which are virtual representations of users. Personas enable the creation of simulation environments that allow researchers to configure and train intervention models before deployment. However, developing simulation models to generate synthetic data for health behavior personas requires interdisciplinary expertise that is often inaccessible or time-intensive for domain researchers. To address this challenge, we introduce a novel cooperative multi-agent architecture based on large language models (LLMs) to empower domain experts to create and simulate health behavior personas through natural language interaction. A proof-of-concept is offered through a case study on literature-based smoking cessation personas, which shows how our approach reduces barriers for researchers while ensuring reliable synthetic behavioral health data.
Download

Paper Nr:	171
Title:	Interpretable and Fair Machine Learning for Predicting Demand in Emergency and Urgent Care Services
Authors:	Cássio Soares Carvalho, Júlio Carlos Balzano de Mattos, Marilton Sanchotene de Aguiar, Felipe Mendes Delpino, Lílian Munhoz Figueiredo and Bruno Pereira Nunes
Abstract:	Interpretability and fairness are critical for the responsible use of machine learning (ML) in healthcare. This study evaluates explanation clustering methods, LIME-EC and SHAP-EC, to assess model interpretability and algorithmic fairness in predicting the demand for urgent and emergency care services. Using a population-based cohort of 3,123 adults and 24 predictor variables covering multimorbidity, socioeconomic and behavioral factors, and health-related topics, we trained several ML models and applied EC to derive core explanations. Interpretability was quantified through the agreement between core explanations and original model predictions, and linked directly to multiple fairness metrics. Results show that SHAP-EC provides higher agreement percentages, particularly for Random Forest and Artificial Neural Networks, and better alignment with fairness measures compared to LIME-EC. Across all models, EC identifies agreement and disagreement regions, highlighting contexts where predictions are more or less reliable. These findings support explanation clustering as a robust framework for interpretability quantification and fairness-aware evaluation of ML models, with potential to inform safer and more equitable healthcare decision-making.
Download

Paper Nr:	180
Title:	Detecting Exercise Deviations in Home Physiotherapy Using Single-Camera Pose Estimation
Authors:	Matthias Maszuhn, Finn Siegel, Frerk Müller-von Aschwege and Andreas Hein
Abstract:	This work investigates the automated assessment of rehabilitation exercises using machine learning and pose estimation from a single smartphone camera. An Android application was developed to record and analyze five common physiotherapy exercises, enabling the detection of correct and incorrect execution variants without requiring specialized hardware. A study with 19 participants was conducted to evaluate two complemen-tary analytical approaches: supervised classification and unsupervised clustering. A Random Forest classifier achieved a high overall mean accuracy of 91.1%, confirming the feasibility of reliably identifying predefined movement deviations. Unsupervised K-Means clustering further demonstrated that meaningful movement patterns can be recovered without prior labels, with higher Silhouette and Adjusted Rand Index (ARI) scores observed for individualized clustering compared to generalized clustering across participants. These findings highlight the potential of combining supervised and unsupervised methods for comprehensive movement quality analysis. Future work will focus on expanding dataset diversity, improving generalization across users, and investigating which kinematic parameters are most relevant for exercise recognition to support interpretable, real-time feedback in personalized rehabilitation.
Download

Paper Nr:	198
Title:	Human-Centered Design of the NeuroAIreh@b: A Tablet-Based Platform for Remote Cognitive Training
Authors:	Teresa Paulino, Joana Câmara, Diogo Branco, Luís Ferreira, Mónica Spínola, Ana Lúcia Faria, Sergi Bermudez Badia, Eduardo Fermé and Mónica S. Cameirão
Abstract:	Interactive digital technologies are transforming cognitive rehabilitation after stroke, offering new opportunities for personalized, accessible care. The NeuroAIreh@b is an innovative tablet-based platform designed to deliver adaptive cognitive training both in clinical environments and at home, with integrated remote monitoring to support continuous care. Our development approach emphasized a highly collaborative, user-centered design process involving over 172 participants, including stroke survivors, rehabilitation clinicians, engineers, and designers, to ensure the platform meets the diverse needs of end users and healthcare providers. Methods such as surveys, focus groups, co-design workshops, prototyping, and user interaction analysis informed iterative refinements of the cognitive training activities, the management interface, and auxiliary software tools. This publication provides an overview of the multi-phase design and development journey of the Neu-roAIreh@b platform, highlighting challenges and insights from stakeholder engagement and usability testing. Preliminary findings from a four-week pilot study with 15 patients with acquired brain injury and mild cognitive impairment demonstrate the initial feasibility of the cognitive training apps and target audience acceptance in real-world settings.
Download

Paper Nr:	201
Title:	BDLBS: An Expert-Rule-Based Generator for Synthetic Complex Longitudinal Health Data
Authors:	Hugo Boisaubert, Fatoumata Dama, Corinne Lejus-Bourdeau and Christine Sinoquet
Abstract:	Advances in sensor technology have led to increasingly complex temporal data, with healthcare digitalization being a prominent example. Such data hold great potential for machine learning and deep learning research. However, privacy and data quality issues make calibrating models on real datasets challenging, creating a need for realistic synthetic data. Deep learning models require large datasets and computational resources, and they do not always reproduce ground-truth dynamics faithfully. We address the generation of realistic asynchronous data combining event sequences with multivariate time series. We propose BDLBS, a simulator based on expert-defined rules, designed to fill the gap in generating realistic anesthesia datasets. The simulator is lightweight and scalable, and extensive tests demonstrate that BDLBS realistically captures the effects of medical actions on patient dynamics, along with the temporal and inter-variable dependencies in these longitudinal data. The simulator enables the easy creation of benchmarks for calibrating machine learning and deep learning models and supports the secondary use of healthcare data in research. A public, user-friendly platform allows experts to define new surgical procedures and enables other users to generate synthetic benchmarks without any code installation.
Download

Paper Nr:	205
Title:	Counterfactual Explanations Applied to the Analysis of Violent Discipline against Children in Africa: A Machine Learning Approach
Authors:	Lívia Câmara Xavier, Hasheem Mannani and Cristiane Neri Nobre
Abstract:	Violent discipline against children and adolescents remains a significant public health challenge, especially in African countries, where physical and verbal punitive practices are still culturally embedded. This study employs machine learning and counterfactual interpretability techniques to identify the primary factors associated with this phenomenon, using data from the Multiple Indicator Cluster Surveys (MICS) provided by UNICEF and focusing on African countries. After data preprocessing, including imputation, normalization, and class balancing, several classification algorithms were evaluated, namely Decision Tree, Random Forest, Support Vector Machine, Neural Network, and XGBoost. Among them, XGBoost achieved the best predictive performance, reaching an F1-score of 0.81. Subsequently, the CSSE (Agnostic Method of Counterfactual, Selected, and Social Explanations) method was applied to generate counterfactual explanations for instances classified as having experienced violent discipline. The results indicate that minimal changes in specific variables, particularly those related to beliefs about physical punishment and patterns of family communication, are sufficient to reclassify most instances from “experienced violent discipline” to “did not experience.” These findings highlight the potential of counterfactual explanations as decision-support tools for public policy, enabling a transparent understanding of the determinants of child violence and supporting the design of more effective preventive strategies.
Download

Paper Nr:	206
Title:	Challenges in the Analysis of Clinical Text Data for Predicting Postpartum Hemorrhage: An Exploratory Study Using Brazilian Hospital Records
Authors:	Júlia M. Bomfá, Flávia R. de Oliveira, Zilma S. N. Reis and Cristiane N. Nobre
Abstract:	Postpartum hemorrhage (PPH) is a leading cause of maternal morbidity and mortality, yet early risk prediction remains challenging. This study explores clinical data from a Hospital Foundation in Brazil, combining structured variables and unstructured narratives to identify risk factors. Results show that many PPH cases occur in women initially classified as low or medium risk, highlighting limitations in current stratification. Text mining revealed complementary insights from clinical narratives, but variability and incomplete documentation pose challenges for predictive modeling. These findings emphasize the need for improved data capture and treatment in hospitals and provide a foundation for developing AI-driven tools to support timely, personalized interventions for PPH prevention.
Download

Paper Nr:	216
Title:	Predicting LLM Correctness in Prosthodontics Using Metadata and Hallucination Signals
Authors:	Lucky Susanto, Anasta Pranawiyana, Cortino Sukotjo, Soni Prasad and Derry Wijaya
Abstract:	Large language models (LLMs) are increasingly adopted in high-stakes domains such as healthcare and medical education, where the risk of generating factually incorrect (i.e., hallucinated) information is a major concern. While significant efforts have been made to detect and mitigate such hallucinations, predicting whether an LLM's response is correct remains a critical yet underexplored problem. This study investigates the feasibility of predicting correctness by analyzing a general-purpose model (GPT-4o) and a reasoning-centric model (OSS-120B) on a multiple-choice prosthodontics exam. We utilize metadata and hallucination signals across three distinct prompting strategies to build a correctness predictor for each (model, prompting) pair. Our findings demonstrate that this metadata-based approach can improve accuracy by up to +7.14% and achieve a precision of 83.12% over a baseline that assumes all answers are correct. We further show that while actual hallucination is a strong indicator of incorrectness, metadata signals alone are not reliable predictors of hallucination. Finally, we reveal that prompting strategies, despite not affecting overall accuracy, significantly alter the models' internal behaviors and the predictive utility of their metadata. These results present a promising direction for developing reliability signals in LLMs but also highlight that the methods explored in this paper are not yet robust enough for critical, high-stakes deployment.
Download

Paper Nr:	221
Title:	Simulated Evaluation of Machine Learning Models for Predicting Late-Stage Diagnosis and Mortality in Prostate and Breast Cancer Patients
Authors:	Huan-ju Shih and Janusz Wojtusiak
Abstract:	This study evaluates machine learning models for predicting late-stage diagnosis and one-year cancer-specific mortality in prostate and breast cancer patients aged 65 years and older, using linked SEER-Medicare Health Outcomes Survey (SEER-MHOS) data from 1998 to 2011. Random Forest, Gradient Boosting, and Logistic Regression models were developed separately for each cancer type, with optimized lookback windows determined through stratified k-fold cross-validation and simple moving averages to balance predictive relevance and reduction of noise from historical survey data. Late-stage diagnosis models achieved modest performance (highest AUC of 0.67 for prostate cancer with Random Forest; 0.58–0.60 for breast cancer), while one-year mortality prediction demonstrated stronger results (AUC up to 0.86 in prostate cancer). A simulation-based framework was employed to explore hypothetical improvements by reclassifying high-risk late-stage cases to early-stage across probability thresholds from 0.05 to 0.50, supplemented by a perfect prediction scenario (AUC ≥ 0.9). Threshold-based reclassification generally lowered predicted mortality probabilities, with key inflection points indicating trade-offs in sensitivity, specificity, and false positives. Perfect prediction simulations revealed greater potential mortality reductions, highlighting the value of enhanced early detection. The integration of predictive modeling with simulation-based analysis strengthens model interpretability and transparency without direct clinical claims. Limitations include modest discrimination for late-stage prediction due to class imbalance and survey data constraints. Future work will incorporate additional claims data to improve model reliability. This approach advances methodological understanding of cancer progression, risk stratification, and potential intervention impacts in oncology.
Download

Paper Nr:	224
Title:	Toward a Protocol for Second Opinion Systems in Rare Diseases: Integrating Evidence, Governance, and Artificial Intelligence in a Sociotechnical Approach
Authors:	Vinícius Lima, Mariana Mozini and Domingos Alves
Abstract:	Rare diseases affect millions worldwide but remain underserved by fragmented pathways and uneven access to subspecialty expertise. This position paper argues that rare disease second opinion services should transition from ad hoc initiatives to a protocolized, auditable national digital health infrastructure. We advance a policy-ready protocol grounded in a rapid scoping review and articulated through six premises: a convergent, standardizable workflow; sociotechnical governance; transnational, data-protection-by-design (mapping roles, purposes, jurisdictions, and legal bases; standards-based interoperability; prospective evaluation aligned with established frameworks; and human-validated AI that accelerates expert work. We operationalize these premises into a pragmatic framework that encodes policy as code, routes data by jurisdiction, defines a stepwise case workflow from intake to signed recommendation and write-back, and captures key performance indicators. The approach is designed to scale from local pilots to transnational collaborations while preserving clinical responsibility with the treating team. Framing second opinions as public digital goods can redistribute scarce expertise, reduce unwarranted variation, and generate real-world evidence, advancing rare disease care.
Download

Paper Nr:	233
Title:	Agentic AI in Pharmacovigilance: A Position Paper on Opportunities, Challenges, and Implementation
Authors:	Bhaarat Pachori
Abstract:	Pharmacovigilance (PV), the guardian of drug safety, is currently overwhelmed by an escalating volume of data. The traditional methodologies are constantly challenged by the increasing quantity and complexity of the data and can pose a risk of delayed detection of adverse drug events. Agentic Artificial Intelligence (Agentic AI) is emerging as an ally with transformative potential in this domain. In this position paper, we present a forward-looking vision for realizing this potential. We assert that development must extend beyond technological capabilities to address the critical requirements of trust, safety, and policy compliance. We investigate key applications such as automated case processing and proactive signal detection, while critically examining the significant hurdles, including cross-lingual semantic understanding, designing effective Human-in-the-loop (HITL) frameworks, and ensuring system reliability. We provide preliminary empirical validation demonstrating that an agentic architecture reduces algorithmic overconfidence by verifying signals against regulatory databases. Ultimately, this paper proposes a research agenda to guide the data mining community in developing robust and ethically-sound agentic systems. Harnessing Agentic AI through focused, interdisciplinary research is essential for directing in a new era of proactive, efficient, and patient-centric pharmacovigilance.
Download

Paper Nr:	235
Title:	Exploration of Barriers and Facilitators for Utilizing Telehealth for Trauma and Emergency Medicine Patients: A Review of Qualitative Studies
Authors:	Muneef Alshammari and Fahad Alshammari
Abstract:	This review explores barriers and facilitators to telehealth utilization in trauma and emergency medicine through a thematic analysis of 20 qualitative and mixed-methods studies published between 2017 and 2025. Telehealth, defined by the WHO as the remote delivery of healthcare using information and communication technologies, has gained prominence, especially during the COVID-19 pandemic. Trauma remains a leading cause of morbidity and mortality globally and in Saudi Arabia, underscoring the importance of effective telehealth integration. The analysis identified six key themes influencing telehealth adoption: (a) Technical Infrastructure and Equipment, (b) Clinical Workflow Integration and Decision-Making, (c) Communication and Interpersonal Factors, (d) Training, Competency, and Staff Acceptance, (e) Organizational and System-Level Implementation Factors, and (f) Regulatory and Legal Barriers. Barriers included connectivity issues, clinician reluctance, communication challenges, insufficient training, organizational constraints, and complex regulatory environments. Facilitators encompassed reliable technology, improved specialist access, strong professional relationships, ongoing education, leadership support, and clear protocols. The findings highlight telehealth as a complex socio-technical transformation requiring coordinated strategies across multiple domains to optimize emergency and trauma care delivery. Future research should focus on emerging technologies, long-term outcomes, and standardized regulatory frameworks to enhance telehealth implementation and patient care.
Download

Paper Nr:	238
Title:	Avatar-Enhanced Voice Carebot for Cancer Patients: In-Hospital Supervised Pilot and Telehealth-Ready Design
Authors:	Ikuo Keshi, Fumiyo Araki, Mizuho Kato and Yoshinori Munemoto
Abstract:	Cancer patients require psychosocial support, yet oncologists have limited consultation time. We developed a voice carebot combining GPT-4 class language models, cloud-based speech recognition, and an animated physician avatar. Three sequential evaluations were conducted: healthcare professional review (about ten healthcare professionals), public field trial (n=27), and supervised clinical pilot with cancer outpatients (n=16). In the clinical pilot, 35% of patients disclosed previously unasked questions, 59% reported reduced anxiety (5-point Likert, ≥4), and 53% expressed trust in the information provided (≥4). However, 60% experienced voice interaction difficulties (“voice input was difficult”, ≤2). System improvements included migrating to a cloud speech recognizer and adopting a senior physician avatar. This supervised pilot showed feasibility indicators of AI-augmented communication in controlled oncology settings, provided older-adult-oriented UI insights, and established clinical safety protocols. Phase 4 trial (n=21) using POMS2 standardized assessment confirmed significant improvements: disclosure of hesitant concerns increased from 35% to 76%, user experience metrics improved substantially (d = 0.95–1.31), and Total Mood Disturbance reduced by 63%. While currently limited to supervised hospital deployment, the system’s telehealth-ready architecture supports planned progression toward semi-autonomous operation.
Download

Paper Nr:	240
Title:	PROMBot-HSM-AF: Design and Implementation of a Chatbot and Clinical Monitoring Platform for Patient-Reported Outcomes after Atrial Fibrillation Catheter Ablation
Authors:	Pedro Dias, Catarina Nunes-Da-Silva, Federico Guede-Fernandez, Eduarda Oliosi, Sérgio Laranjo and Ana Londral
Abstract:	Atrial fibrillation is the most frequent sustained cardiac arrhythmia, and while catheter ablation is an established treatment, the post-intervention period is critical. With routine care based on periodic consultations and on-demand monitoring, gaps remain in which deterioration or arrhythmia recurrence may not be captured. Remote patient monitoring and conversational agents such as chatbots have shown promise in enhancing continuity of care, improving patient engagement, and enabling early detection of clinical changes. In this paper, we present PROMBot-HSM-AF, a rule-based chatbot coupled with a clinical monitoring platform designed to collect patient-reported outcome measures (PROM) and symptoms following AF ablation. The paper describes a forthcoming pilot study in a real-world hospital setting and outlines PROMBot-HSM-AF digital tools’ development process. By combining structured PROMs collection with chatbot-based interaction, PROMBot-HSM-AF aims to address a critical gap in post-ablation follow-up and contribute to the safe and effective deployment of patient-centred digital tools in cardiovascular care.
Download

Paper Nr:	250
Title:	Personalised Stroke Rehabilitation: An AI Pipeline for Exercise Programmes Using a Co-Designed Decision Support Tablet Application
Authors:	Thrisha Rajkumar, Sarah Koerner, Anika Pinto, Regan Shakya, James Pope, Maria Galvez Trigo, Ali Al-Nuaimi, Michael Loizou, Kenton O'Hara and Praveen Kumar
Abstract:	Stroke rehabilitation requires personalised and continuously adapted exercise programmes, resulting in significant therapist involvement and is often impractical for patients recovering at home in community settings. This motivates the need for assistive tools and decision support systems to enhance efficiency and rehabilitation progress. This position paper presents an integrated pipeline combining a therapist-informed tablet application with artificial intelligence (AI) models to support therapists in decision-making. Co-designed with stroke therapists, human-computer interaction (HCI) researchers, AI experts, and persons with stroke (PwS), the application captures baseline and weekly reassessment data, including BBS, TUG, pain, perceived difficulty, and FITT prescriptions, across 4–6 week cycles to determine whether to progress, sustain, or regress exercises. To facilitate early model development, we created a clinically informed synthetic dataset (n = 336 sessions across 5 PwS profiles over 12 weeks) that simulates functional progression and therapist decision-making patterns. This dataset reflects key features identified through workshops with clinicians and PwS, capturing essential assessment metrics such as stroke characteristics, functional scores, therapist goals, patient feedback, exercise difficulty, repetitions, duration, body area, FITT parameters, and exercise recommendations. We trained and evaluated models to predict weekly progression decisions. Logistic regression achieved a weighted F1- score of 51.6%, while a multilayer perceptron reached 79.3% and a decision tree 90.2%. Clinical data will be collected in the next stage of the project (5–8 PwS, 4–6 weeks) and integrated with the synthetic dataset using real–synthetic fusion. This work advocates AI-augmented tools for scalable, patient-centred community stroke rehabilitation, with future efforts exploring generative AI and clinical validation.
Download

Paper Nr:	253
Title:	AI-Based Software as a Medical Device: Bridging EU MDR Compliance with ISO/IEC 80001 and 81001 Cybersecurity Governance in Healthcare Systems
Authors:	Lakshika Chandradeva and Silvana Togneri MacMahon
Abstract:	Artificial Intelligence based Software as a Medical Device (AI-SaMD) is reshaping healthcare by enhancing diagnostics and management. However, its adoption in Europe faces challenges in regulation, interoperability, and cybersecurity. This review examines the EU Medical Device Regulation (MDR), ISO/IEC 80001, and IEC 81001, highlighting regulatory ambiguities around AI, accountability issues, and cybersecurity demands from GDPR and the NIS2 Directive. It identifies gaps and synergies in the frameworks, advocating for harmonized strategies, cyber resilience, and integrated governance between manufacturers and healthcare providers. Future directions include regulatory convergence and toolkits for implementation, promoting explainability and trust in AI-SaMD.
Download

Paper Nr:	259
Title:	SEPE: Towards a Patient-Empowered, AI-Supported Optimisation of Rare-Disease Diagnostics
Authors:	William Philipp, Siddarth Venkateswaran, Nina Pillen, Jawan Kolanowski, Andreas Schäffler, Bianca Greiten, Christian Himstedt, Md Monsur Ali, Tobias Bäumer, Sebastian Fudickar and Ronald Böck
Abstract:	The SEPE project (Rare Diseases and Patient Empowerment) seeks to accelerate rare-disease diagnosis through AI-supported, patient-centred digital tools. Dedicated Centres for Rare Diseases aim to improve and accelerate the diagnostic process for these complex cases. In spite of these improvements, diagnosis remains a significant burden on patients. The SEPE platform envisions connecting patients, general practitioners, and specialists via interoperable applications for secure data exchange, participatory data enrichment, and AI-assisted decision support. In the first year, the project focused on acquiring and integrating retrospective and prospective datasets, developing a layered text extraction and anonymisation pipeline using on-premise large language models (LLMs), and implementing prototype applications for patients and clinicians. Early results suggest that privacy-preserving AI and multi-agent anonymisation frameworks can enable secure, data-driven workflow preparation, laying the foundation for future clinical evaluation and integration.
Download

Paper Nr:	274
Title:	Semantic Attribution Signals for Allergic Rhinitis Nowcasting
Authors:	Hannah Béchara, Krishnamoorthy Manohara and Slava Jankin
Abstract:	Digital epidemiology has largely treated search queries and social-media posts as volume proxies for incidence. Such counts capture attention rather than interpretation and are sensitive to media shocks and platform shifts. We argue for a semantic attribution layer that encodes how users attribute symptoms to putative causes. Focusing on allergic rhinitis (AR) in England, we type causal statements in TikTok descriptions into environmental, infectious and other triggers, and aggregate them into weekly, confidence-weighted indices that are Beta-smoothed, exponentially smoothed, and volume-gated per platform. These indices are fused with Google Trends and TikTok via exogenous regressors and confidence-weighted gating. In an illustrative nowcasting study across horizons h = 0...4, adding TikTok to Google Trends improves short-horizon performance for non-linear models; raw causal counts are unstable; and typed, gated attribution yields more consistent gains for tree-based and boosting models while often harming linear/ARIMA baselines. We outline a compact specification for attribution-aware pipelines and practical guidance for integrating this semantic layer into digital surveillance.
Download

Paper Nr:	280
Title:	Predicting Internet Addiction in Young Indian Adults: A Classification Approach Using Depression, Anxiety, and Stress Levels
Authors:	Aryaman Pathak, Shweta Sunil, Madhav Rao and Manoj Sharma
Abstract:	With increasing internet accessibility, concerns over Internet Addiction (IA) have grown, particularly among specific demographics. This study develops a machine learning (ML) model to predict IA in young Indian adults using Depression, Anxiety, and Stress (DASS) scores as key predictors. Leveraging a dataset of 725 young adult samples (mean age 21.06 years) independently collected in India, a neural network (NN) was employed for its ability to model complex non-linear relationships. Despite the dataset’s specific demographic and geographical scope, regularization techniques were applied to mitigate overfitting, with the model achieving an accuracy of 84.13% on this cohort. To the best of our knowledge, this is the first ML-based approach using affective state scores for IA prediction within this specific population. A web app was also developed to collect user feedback, enabling continuous model refinement. While the classification model is publicly available for future research, its findings are primarily relevant to young Indian adults, and further validation in diverse populations is essential. With continued enhancements in model accuracy and generalizability, such predictive tools could serve as valuable aids to current assessment techniques and hold promise for future clinical application in the early detection and intervention of Internet Addiction.
Download

Paper Nr:	288
Title:	Bridging Data and Disciplines: A Clinicians’ Dashboard for Interdisciplinary Rehabilitation
Authors:	David Snowdon, Tobias Michels, Hester Knol, Michael Buschermöhle, Dirk Möller and Christoff Zalpour
Abstract:	Postoperative rehabilitation after ankle fractures requires close coordination among healthcare professionals. Yet, communication between physicians and physiotherapists often remains fragmented, limiting continuity of care and timely recognition of complications. This position paper argues that digital rehabilitation systems should progress beyond data collection toward collaborative, user-centred platforms supporting transparency, clinical reasoning, and interdisciplinary exchange. Using the clinicians’ dashboard developed within the THE-BEA project as an example, we outline its conceptual rationale, design approach, and methodological framework. The dashboard integrates sensor-derived, clinical, and patient-reported data into coherent visualisations that promote shared situational awareness and data-informed decision-making. Developed through iterative prototyping and participatory design, it reflects a practical application of user-centred principles in digital rehabilitation. A usability study employing the Think Aloud method is currently being conducted to evaluate its feasibility and interface design. Clinician dashboards are thus positioned as essential interfaces for enabling integrated and evidence-informed rehabilitation management.
Download

Paper Nr:	292
Title:	Effects of Symbolic Methods on Biomedical Graphs for Link Prediction
Authors:	Benedek Bahrami and Annette ten Teije
Abstract:	Knowledge graph-based AI models made for link prediction have populated the field in the last years. Their natural capability to capture the confluence of data makes them powerful tools for this task. An important application of knowledge graph link prediction due to sparse information and difficult testing is the field of drug combinations, known as polypharmacy. Recently the use of embedded models have dominated this specific use case, but symbolic methods used alongside embeddings have been proven to achieve an increased performance. In this paper, we implement symbolic rule-based graph completion on the embedded model TriVec in order to improve its results. Trivec uses the dataset compiled for a graph convolutional network model Decagon, and applies tensor factorisation to it, evaluating its results with the same metrics as Decagon did. We created and evaluated multiple graph completion strategies leveraging different aspects of the dataset. Experimental results show that graph completion was able to increase graph connectivity and clustering, while statistical tests have shown consistent improvements in AUC-ROC, AUC-PR and AP@50 metrics, proving the strengths of symbolic augmentation in predictive performance. Additionally, we highlighted discrepancies between TriVec’s published performance and its publicly available implementation, causing reproducibility concerns in this domain.
Download

Paper Nr:	303
Title:	Privacy and Security by Design: What about Africa?
Authors:	Wendgounda Francis Ouedraogo and Georges Parfait Eloundou Eloundou
Abstract:	Privacy by Design (PbD) and Security by Design (SbD) have been gaining ground in recent years with the global issue of personal data protection. In software engineering, these two concepts are complementary because while PbD advocates for proactive consideration of user privacy issues throughout the solution's lifecycle, starting from the design phase, SbD ensures that security aspects are a reality in the solution deployed from the design phase onwards. With digitalization becoming a reality in Africa (for example in the field of medicine), technological solutions are now designed or at least adapted before deployment. Taking the data-intensive field of healthcare as an example, what about the protection of personal data? Can these concepts of privacy and security by design find their place in healthcare software engineering in Africa?
Download

Paper Nr:	306
Title:	Effective Route Planning for Intra-Hospital Patient Transportation Using Discrete Combinatorial Optimization
Authors:	Saran Karthikeyan, Cord Spreckelsen and Sasanka Potluri
Abstract:	Intra-Hospital Patient Transportation (i.e., within departments in a hospital) impacts the healthcare process in daily operations and stakeholders like medical and healthcare professionals, patients, and hospital managers. Transporters execute non-continuous, subsequent transportation tasks within each shift. The task assignments among the transporters for creating effective route plans (i.e., solution) is a challenging research problem, considering multiple aspects like workload balance, uncertainty, high-quality service, and multi-objective optimization. The objectives are to optimize the number of transporters and the operational flow by effective task distribution among the transporters across different data sets. The data sets consist of tasks of an 8-hour shift on various days extracted from retrospective data. The minimization of parameters like the sum of total travel time and total idle time, along with the norm of travel time between transporters, improves the effectiveness of optimization. Our baseline method constructed a solution using multi-stage metaheuristics with constraints. Our proposed work is a hybrid of constructed multi-stage metaheuristics and a combinatorial tree, using techniques like multi-processing, varied neighbourhood search with parameter estimation, and constraints. Empirical results on different data sets show that our proposed work achieves multi-objectives and improves the effectiveness of the solution better than our baseline method.
Download

Paper Nr:	316
Title:	A Review of Methods, Outcomes, and Organizational Barriers in Healthcare Service Design
Authors:	Chenxin Ling
Abstract:	Hospitals are complex socio-technical systems, where service design is employed to enhance patient experience as well as the quality and efficiency of processes. Patient journey mapping identifies issues in cross-departmental coordination from the patient’s perspective, while service blueprinting brings patient-facing touchpoints and back-office support into a single framework. This clarifies the division of labor and information flow, provides a common language for collaboration, and focuses improvements on comparable outcomes. However, descriptive work dominates existing studies, measurement indicators lack standardization, follow-up periods are short, staff-related outcomes are insufficiently measured, and organizational and contextual factors lack consistent coding and classification. As a result, comparison and dissemination are limited. This paper reviews methodological studies on healthcare service design, discusses the main outcomes achieved in hospital settings and their commonly used indicators, briefly summarizes organizational barriers and coping strategies in clinical and managerial practice, and explores the relationships between methods, outcomes, and organizational factors as well as future directions. The value of this study lies in building an integrated framework of methods, outcomes, and organizational barriers, while examining effective evaluation indicators and implementation pathways to support the large-scale dissemination of healthcare service design and provide practical insights.
Download

Paper Nr:	322
Title:	SaúdeLink: AI-Enhanced Health Data Integration to Support Healthcare Providers
Authors:	Maurício de M. dos Santos, Evilasio C. Junior, Rossana M. C. Andrade and Pedro A. M. Oliveira
Abstract:	This paper presents the development of SaúdeLink, an intelligent platform designed to integrate physiological and behavioral data from wearable devices with large language models (LLMs), specifically leveraging the Gemini language model (gemini-2.5-flash), aiming to support contextual analysis and decision-making in digital health. The proposal explores the potential of data collected from wearable technologies, such as heart rate, sleep, and physical activity levels, to generate personalized and context-aware insights about patients’ health and well-being. The development followed the Technical Action Research (TRA) methodology, which enables iterative cycles of design, technical implementation, and empirical refinement. The proposed architecture combines vector-based representations (embeddings), semantic search, and prompt engineering, allowing clinical information to be interpreted adaptively according to each patient’s history and profile. For the initial evaluation, two synthetic profiles, Sedentary and Athlete, were created to represent contrasting lifestyle patterns. Results demonstrated that the system can distinguish between physiological profiles and generate recommendations tailored to each context. Moreover, the platform was designed to support continuous expansion, allowing daily data ingestion and the integration of external documents, such as medical guidelines and scientific publications, to enrich the model’s contextual understanding. The obtained results indicate that SaúdeLink provides a solid foundation for future applications involving longitudinal patient monitoring, validation with healthcare professionals, and integration with existing clinical systems. The platform highlights the potential of combining Internet of Things (IoT) data and generative artificial intelligence to enhance clinical screening, monitoring, and decision support.
Download

Paper Nr:	326
Title:	Evoke: A Framework for Measuring Emotional Arousal to Robotic Voice Modulation Using Electrodermal Activity
Authors:	Rowan Stratton, Ayesha Noshin, Arshia Khan and Jomara Sandbulte
Abstract:	Many neurodivergent users react strongly to sound, and a robot’s voice can become calming or overwhelming depending on how it is delivered. Designers need a clear way to measure how vocal pitch affects emotional comfort. This paper presents Evoke, an exploratory feasibility pilot that examines how robotic voice modulation influences emotional arousal. A Pepper robot read short story passages using three pitch settings. Electrodermal activity was recorded and paired with brief self-report surveys. Four participants took part in this study. Early patterns appeared in the data. Low-pitch narration produced the largest calming response, shown by reductions in tonic skin conductance. Default pitch also reduced arousal, although the results varied across individuals. High-pitched narration showed mixed effects and clear differences in how people interpreted the voice. These preliminary findings show that the protocol can capture changes in autonomic state during robot-delivered speech. Evoke provides an initial foundation for larger studies and may guide the development of comfortable robot voices for users with auditory sensitivity.
Download

Paper Nr:	328
Title:	Clustering-Based Analysis of Risk Profiles for Depression Using Quality of Life Data
Authors:	Willian Jorge Sousa Furtado, Pedro Almir Martins Oliveira, Rossana Maria Castro Andrade, Evilasio Costa Junior, Wilson Castro, Victória Tomé Oliveira and Pedro de Alcântara Santos Neto
Abstract:	Depression is a condition that affects thousands of people worldwide, significantly compromising Quality of Life (QoL). Given the urgency of early diagnosis and the need for continuous monitoring, this study employs non-invasive data from wearable devices and unsupervised learning (clustering) to segment mental health risk profiles. Using the Healful dataset and its QoL score, the research compares the effectiveness of different algorithms. The strategic choice of the clustering technique, prioritizing business interpretability over purely statistical cohesion metrics, enabled a more pragmatic risk segmentation. The model, configured with K=4, achieved a significant variation of 48.03 points in the QoL metric, clearly differentiating four distinct risk profiles: High Risk, Moderate-High, Moderate-Low, and Low Risk. This characterization provides a robust analytical framework for optimizing digital health interventions.
Download

Paper Nr:	334
Title:	Accelerometer-Based Activity Recognition to Support Calorie-Free Feedback in an Intuitive Eating Mobile Platform
Authors:	Nathan Yasnovsky, Joseph Nash, Dayyan Chaudhri, Adam Neulander and Delaram Yazdansepas
Abstract:	Calorie-centric tracking remains dominant in commercial wellness systems, despite evidence that numeric monitoring can conflict with recovery-oriented and intuitive-eating approaches. To support calorie-free, psychologically safe movement feedback, we develop a lightweight deep learning pipeline for activity recognition using wrist-mounted accelerometer data. Our method trains a temporal convolutional network (TCN) to classify walking, jogging, and stair-related activities (stair ascent and descent) from windowed raw accelerometer signals. We evaluate model generalizability using leave-one-subject-out (LOSO) validation and report both person-level and per-class performance across daily-living activity classes. This activity recognition module is intended to serve as the sensing foundation for a companion system that estimates energy-expenditure intensity qualitatively, without calorie conversion, as part of a mobile intuitive-eating application. By combining efficient wearable sensing with deep time-series modeling, this work contributes a technically rigorous and human-centered approach to health informatics, enabling movement-aware feedback that supports well-being without reinforcing calorie-focused behaviors.
Download

Paper Nr:	379
Title:	A Machine Learning Method for Facial Image-Based Autism Spectrum Disorder Detection
Authors:	Jessica Lucarelli, Mario Cesarelli, Antonella Santone, Fabio Martinelli and Francesco Mercaldo
Abstract:	Autism Spectrum Disorder is an early-onset neurodevelopmental condition for which timely diagnosis is essential to exploit the critical windows of brain plasticity. In recent years, AI-based craniofacial feature analysis has emerged as a potential support for early screening. This exploratory study evaluates the applicability of machine learning algorithms to classify facial images of subjects with and without autism spectrum disorder, without replacing traditional clinical tools. A dataset of 2241 images, cleaned and rearranged by Kaggle “Autistic Children Facial Dataset”, was processed using the Inception v3 and SqueezeNet embedders combined with multiple classifiers. The models achieved good performance, with maximum accuracy, precision, and recall values of 0.814, showing that the proposed method obtains interesting results.
Download

Paper Nr:	397
Title:	Social Isolation Detection Based on ADL Recognition and Multi-Sensor Monitoring
Authors:	Ghazi Bouaziz, Damien Brulin and Eric Campo
Abstract:	With an aging population, more and more elderly people are choosing to remain at home, promoting autonomy but increasing the risk of social isolation due to reduced mobility, limited transportation, and changing social structures. We propose a multi-sensor system to detect early signs of social isolation by analyzing mobility and eating habits. The system was tested for three months in the homes of five elderly participants. We evaluated three supervised machine learning models-decision tree, random forest, and logistic regression-using 5-fold cross-validation. Random Forest performed best, achieving a Macro-F1 of 0.6 and balanced accuracy of 0.7 for mobility-related isolation, and a Macro-F1 of 0.733 and balanced accuracy of 0.8 for eating-related isolation. These results demonstrate its potential for practical deployment in smart homes to support aging populations.
Download

Paper Nr:	400
Title:	EEG-Based Prediction of Neurological Recovery after Cardiac Arrest Using Autoencoder Features
Authors:	Amrutha Manne, Andreas Hein and Benjamin Cauchi
Abstract:	Accurately predicting neurological recovery in comatose patients post cardiac arrest remains a major clinical challenge. Electroencephalography (EEG) offers a non-invasive, bedside-accessible tool for monitoring cerebral function, yet its variability and susceptibility to artifacts hinder reliable prognostication. This paper investigates a machine learning approach using standardized preprocessing, spectrogram-based feature extraction, and self-supervised autoencoder pretraining to improve prediction of favorable vs. unfavorable outcomes based on the Cerebral Performance Category (CPC) scale. Two large datasets, the Temple University Hospital (TUH) EEG dataset and the PhysioNet I-CARE cardiac arrest dataset, were harmonized into a unified work-flow including bipolar referencing, robust scaling, and Short-Time Fourier Transform (STFT) spectrogram generation. Convolutional autoencoders were trained using Mean Squared Error (MSE) and Kullback-Leibler Divergence (KLD) objectives, and pretrained encoders were fine-tuned with a classification head. Results reveal that although reconstruction losses learn stable representations, downstream outcome prediction remains near chance, indicating the need for stronger discriminative objectives. The work establishes a reproducible baseline, highlights preprocessing harmonization across heterogeneous datasets, and outlines promising directions for contrastive learning, mutual information-inspired objectives, and multimodal integration.
Download

Paper Nr:	403
Title:	Student Mental Health: Screening for Stress, Anxiety and Depression Using Fitbit Data
Authors:	Rebecca Lopez, Avantika Shrestha, Ml Tlachac, Kevin Hickey, Xingtong Guo, Shichao Liu and Elke Rundensteiner
Abstract:	College students experience many stressors, resulting in high levels of anxiety and depression. Wearable technology provides unobtrusive sensor data that can be used for the early detection of mental illness. However, current research is limited concerning the variety of psychological instruments administered, physiological modalities, and time series parameters. In this research, we collect the Student Mental and Environmental Health (StudentMEH) Fitbit dataset from students at our institution during the pandemic. We assess the ability of predictive machine learning models to screen for depression, anxiety, and stress using different Fitbit modalities. Our findings indicate potential in physiological modalities such as heart rate and sleep to screen for mental illness with the F1 scores as high as 0.79 for anxiety, the former modality reaching 0.77 for stress screening, and the latter modality achieving 0.78 for depression. This research highlights the potential of wearable devices to support continuous mental health monitoring, the importance of identifying best data aggregation levels, and appropriate modalities for screening for different mental ailments.
Download

Paper Nr:	445
Title:	A Dataset for Benchmarking Machine Learning Models for Autonomous Deep Vein Thrombosis Detection Based on Compression Ultrasound Videos
Authors:	Stylianos Didaskalou, Nick Portokallidis, Katerina Tzatzimaki, Maria Liapi, Nikolas Moustakidis, Theofilos Moustakidis, Neringa Balciuniene, Andrius Macas, Rytis Kijauskas, Adomas Aladaitis, Amalia Sotiriadou, Filippos Sarafis, Georgios Kynigopoulos, Michail Potoupnis, Elvira Grandone, Giovanni Mastrangelo, Silvio Maresca, Maxime Gautier, Djamila Chaiba, Hiba Boussaha, Sandrine Goulvent, Chrysovalantis Stylianou, Emine Nezntet Oglou, Kostantinos Chouchos, Savas Deftereos, Kostantinos Papatheodorou, Ioanna Drougka, Penelope Anagnostopoulou, Hong Qing Yu, Eleni Kaldoudi and ThrombUS+
Abstract:	Deep vein thrombosis (DVT) is a major vascular condition associated with substantial morbidity, mortality and healthcare burden. Compression ultrasonography, performed and interpreted by medical experts, is the primary diagnostic method. Advances in machine learning (ML) and deep learning (DL) offer promising op-portunities to support automated and real-time DVT assessment by non-experts. However, existing approach-es rely on pixel-wise vessel annotations, which are costly to generate, posing difficulties in developing models that generalize across devices and acquisition protocols. To address these limitations, we introduce the Throm-bUS+ Dataset #1. The Dataset is based on 2919 segmentation-free compression ultrasound videos from 742 patients suspected of DVT, acquired from a multicenter cohort study across 5 European hospitals.
Download

Paper Nr:	55
Title:	Application of Formal Concept Analysis to Study the Factors Associated with Cholera among Children Aged 1-5 Years During an Epidemic in Rural Haiti
Authors:	Clara Franco, Julio Neves, Luis Zarate and Mark Song
Abstract:	Diarrheal illness remains a leading cause of morbidity and mortality among children in Haiti, exacerbated by the cholera epidemic from 2010 to 2019. This study applies Formal Concept Analysis (FCA) to investigate behavioral, environmental and sociodemographic factors associated with cholera infection among children aged 1 to 5 years in central Haiti (2012–2016), using data from a study on the effectiveness of the cholera vaccine. After preprocessing, FCA generated implication rules with 100% confidence, revealing distinct patterns in infected and non-infected groups. Protective rules consistently highlighted the importance of vaccination, hand hygiene, safe food preparation, water treatment, and access to sanitation facilities. In contrast, combinations of dirty hands, untreated water, dirt floors, agricultural activities, and female gender were strongly associated with infection, even when partial preventive measures (e.g., chlorine treatment) were present. These results underscore that individual interventions, while valuable, may be insufficient in isolation when compounded by environmental and occupational risks. By providing an interpretable rule-based framework, FCA offers a transparent method to identify critical combinations of protective and risk factors. The findings emphasize the need for integrated public health strategies that combine vaccination campaigns, hygiene promotion, nutritional supplementation, environmental improvements, and caregiver education to reduce pediatric diarrheal disease in epidemic contexts.
Download

Paper Nr:	63
Title:	A Formal Concept Analysis Approach to Mental Health Surveillance: Isolation Effects in COVID-19 Times
Authors:	Victor Moraes, Julio Neves and Mark Song
Abstract:	By March 2020, the World Health Organization (WHO) had declared COVID-19 a global health emergency. To contain viral transmission, governments imposed prolonged isolation and strict social restrictions, which disrupted daily routines and raised widespread concerns about mental health. This study examines the psychological and behavioral impacts of extended confinement among Bangladeshi residents, focusing on individuals who remained indoors for more than 15 days. Using Formal Concept Analysis (FCA), a mathematical framework for extracting interpretable rules from data, we analyzed responses from a structured questionnaire that captured demographic, behavioral, and psychosocial indicators. The analysis revealed a core triad of symptoms - growing stress, changes in eating and sleeping habits, and quarantine frustrations - that consistently co-occurred, suggesting a common syndrome of psychological distress. Additional rules linked these symptoms to loss of work interest, mood swings, and weight changes, demonstrating cascading effects on physical health, social interaction, and professional functioning. Importantly, prolonged isolation emerged as a strong predictor of this distress triad, quantitatively linking confinement to negative mental health outcomes. These findings demonstrate FCA’s ability to provide transparent, rule-based insights into pandemic-related psychological patterns. The results underscore the urgent need to integrate mental health surveillance and support into public health strategies during extended lockdowns.
Download

Paper Nr:	68
Title:	Organizational Factors in Health ICT Implementation Readiness: A Scoping Review
Authors:	Parul Nagar and Marieke Sijm-Eeken
Abstract:	Health ICT interventions have the potential to enhance patient outcomes and reduce healthcare cost. However, their successful implementation depends on various factors. In particular, organizational factors can influence success of implementation since interventions often necessitate organizational change. This paper presents a qualitative analysis of relevant literature to identify key organizational factors to influence ICT implementation success before go-live. The insights provided could support a better understanding of these factors prior to implementing a health ICT intervention. This review may serve as a practical resource for assessing and optimizing organizational readiness ahead of project go-live.
Download

Paper Nr:	86
Title:	Non-Invasive Digital Technologies on the Biomechanical Behavior of Office Users, Literature Review
Authors:	Mário Gomes, Jorge Ribeiro, Cristiano de Jesus and Sérgio Lopes
Abstract:	According with a European survey on working conditions, 43% of European workers report back pain, often accompanied by muscular discomfort in the neck and upper limbs. Office-based work is increasingly associated with a range of health issues, including work-related musculoskeletal disorders, computer vision syndrome, cardiovascular strain, and cognitive or psychological fatigue. In the context of Lean Office, Industry 4.0 and Assisted Living Environments and exploring Artificial Intelligence (AI), particularly Computer Vision (CV) and data science, this paper aims to conduct a comprehensive literature review of existing non-invasive digital technologies used to monitor biomechanical behaviour in office settings. Addresses three core challenges: the risk of musculoskeletal disorders, emotional instability and mental fatigue, and the lack of objective data to support performance and productivity management. Based on literature review, we aim to assess the potential of vision-based and sensor-based approaches for smart office applications. We identified a gap in the availability of open-source CV implementations and conclude that the combination of CV and smart sensor technologies enables accurate, real-time, and non-invasive monitoring of posture and fatigue in office environments.
Download

Paper Nr:	113
Title:	Formal Concept Analysis for Characterizing Pediatric Respiratory Diseases: A Case Study on a Moroccan Cohort
Authors:	Luís Oliveira, Julio Neves and Mark Song
Abstract:	Respiratory diseases rank among the top five causes of global mortality and remain a major public health challenge, especially in children, whose immature immune and respiratory systems increase susceptibility to infections. Differential diagnosis between acute respiratory infections (ARIs) such as pneumonia and bronchiolitis, and chronic conditions with acute exacerbations such as asthma, is often difficult due to overlapping symptoms and etiological agents. This study applies Formal Concept Analysis (FCA) to a public dataset of 801 Moroccan children admitted with symptoms of severe clinical pneumonia, aiming to identify clinical–etiological patterns that distinguish these conditions. The dataset includes detailed clinical histories and laboratory detection of 19 respiratory pathogens. FCA enabled the extraction of association and implication rules describing symptom–pathogen relationships. The analysis revealed distinct combinatory profiles: for asthma, wheezing and rhinovirus co-occurred with rhinorrhea and nasal flaring; bronchiolitis was strongly associated with rhonchi and respiratory syncytial virus (RSV); and pneumonia showed higher support for fever, elevated C-reactive protein, and Streptococcus pneumoniae. These findings demonstrate that combinations of attributes, rather than individual symptoms, provide discriminative signatures for differential diagnosis. The study underscores the potential of explainable data-mining methods such as FCA for uncovering interpretable patterns in pediatric respiratory diseases and supporting the development of transparent diagnostic decision-support tools.
Download

Paper Nr:	140
Title:	A Formal Concept Analysis Approach to Evaluating the Sentiment of Mexico City Residents Regarding COVID-19
Authors:	Mateus Sobreira, Luiz Zarate and Mark Song
Abstract:	This study applies Formal Concept Analysis (FCA) to a dataset on emotions and socioeconomic factors among residents of Mexico City during the COVID-19 pandemic. After preprocessing and constructing the formal context, 133 inference rules were extracted using predefined support and confidence thresholds. The rules reveal associations between negative emotions, such as fear and uncertainty, and sleep and mental health issues. FCA proved effective at identifying relevant patterns, despite the limitations posed by the subjectivity of the responses, as the data were collected through a self-assessment questionnaire.
Download

Paper Nr:	147
Title:	From Perceived Risk to Trusted Care: Lessons from Tele-Mental Health Adoption in Lebanon
Authors:	Layal N. Mohtar and Nabil Georges Badr
Abstract:	Tele mental health (TMH) offers scalable solutions for fragile health systems, yet adoption in Lebanon remains constrained by perceived risk, regulatory ambiguity, and weak infrastructure. Building on Mohtar et al. (2025), this paper develops a Trust Moderated Adoption Framework that integrates behavioral theory with governance strategy. While perceived usefulness and subjective norms drive adoption, legal uncertainty, platform insecurity, and fragmented oversight suppress clinician confidence. Trust is positioned as a dynamic moderator that transforms vulnerability into reliability. Sequencing quick wins-peer led training, interim guidelines, incremental platform upgrades-with long term reforms-binding legislation, centralized audits, systemic capacity building-provides a roadmap for sustainable TMH integration and transferable insights for LMICs.
Download

Paper Nr:	148
Title:	Salem: Developing a Localized Drug Food Interaction Checker
Authors:	Najd A. Al-Mouh, Moneira Al-Yahya, Alaa Al-Sleam, Sarah Al-Saab and Manar Al-Awajy
Abstract:	Drug-food interactions (DFIs) pose a significant challenge to the safety of the patients and potentially leading to altered drug efficacy, increased toxicity, or nutritional deficiencies. Current strategies for identifying DFIs largely rely on manual screening, which is both time-intensive and prone to error, particularly in polypharmacy cases. Although several digital tools exist, most lack comprehensive integration of pharmacological and nutritional databases or real time personalization. This paper proposes the conceptual design and potential impact of a digital drug-to-food interaction checker that aligns with the principles of digital medicine. The proposed application leverages pharmacokinetics and pharmacodynamics modeling, coupled with extensive databases of drug and nutrient constituents, to generate personalized alerts and dietary guidance. This system aims to empower healthcare professionals and patients by enhancing medication safety, supporting informed dietary decisions, and reducing preventable drug related complications.

Paper Nr:	172
Title:	Usability Study of the "THEBEA App" Prototype for Therapy Support and Analysis in Orthopedics and Trauma Surgery, Including a Qualitative Survey of Stakeholders Regarding the User Interface
Authors:	Tobias Michels, David Snowdon, Dirk Möller, Nikolaus Ballenberger and Christoff Zalpour
Abstract:	Introduction: Ankle fractures are the seventh most common fractures in Germany and the fourth most common in people over 60. Improved networking between healthcare professionals and digital care options are essential for modern care. The THEBEA project addresses these needs by developing a mobile app with an external sensory feedback system to digitally support rehabilitation after Weber B ankle fractures. Methods: To guide app development, two studies were conducted: a semi-structured stakeholder interview study and a usability study using the think-aloud method. The interviews explored clinicians’ and physiotherapists’ data and communication needs, while the usability study evaluated self-administered physiotherapy exercises recorded via smartphone-based pose tracking and pressure-measuring soles. Results: Stakeholders identified advantages in improved care and time savings. The usability study revealed two main issues: data accuracy and insufficient explanations before use. Key insights included recruitment strategies, applied qualitative methods, real-world testing, and iterative feedback incorporation. Discussion: These findings highlight the importance of user-centered design and multidisciplinary collaboration in digital health innovations and provide a transferable framework for other digital rehabilitation solutions.

Paper Nr:	174
Title:	Adults’ Perceptions and Experiences with Mental Health Applications in Sweden: A Questionnaire Study
Authors:	Ahmad Jabri, Shweta Premanandan and Sofia Ouhbi
Abstract:	The increasing prevalence of stress, anxiety, and depression has heightened demand for accessible mental health support. Mobile mental health applications offer scalable solutions, yet adoption remains uneven. This questionnaire study surveyed 61 adults in Sweden (11 users, 50 non-users) to examine perceptions and factors influencing engagement. Non-users primarily cited lack of awareness, doubts about effectiveness, and privacy concerns, while users valued accessibility and low cost but reported limited long-term engagement due to low perceived value and motivation. Both groups emphasized the need for professional endorsement, personalization, and human interaction. Overall, adoption of digital mental health tools in Sweden is shaped by awareness, trust, cultural expectations, and design quality.
Download

Paper Nr:	185
Title:	Formal Concept Analysis of Social and Behavioral Factors Associated with COVID-19 Vaccine Refusal
Authors:	Cecília Fernandes Silva Costa, Julio Neves, Luiz Zarate and Mark Song
Abstract:	This study applies Formal Concept Analysis (FCA) to explore the social and behavioral factors underlying vaccine hesitancy during the COVID-19 pandemic. The analysis uses data from 999 participants in the United States, capturing variables related to social media use, education, income level, trust in scientists, and belief in conspiracy theories. By extracting association rules from this dataset, FCA uncovers structured relationships between misinformation exposure, institutional distrust, and vaccine refusal. The findings reveal that low institutional trust and high endorsement of conspiracy beliefs are central factors in explaining resistance to vaccination, often mediated by intensive social media use. These results demonstrate the potential of FCA to complement traditional statistical approaches by providing interpretable models of social behavior, thereby supporting the design of more effective public health communication and trust-building strategies.
Download

Paper Nr:	195
Title:	Evaluating Open-Source Transformer Models for Unsupervised Sentiment Analysis in Medical Dialogue Datasets
Authors:	Joseph Kalinzi, James Adinkrah, Khaled Ahmed and Richard Selinfreund
Abstract:	This study presents a systematic evaluation of seven open-source transformer models for unsupervised sentiment analysis in medical dialogues, addressing real-world complexities such as clinical jargon, nested negation, limited annotated datasets, and subtle emotional cues. Utilizing two complementary corpora MedDialog and MTS-Dialog, the research benchmarks both general-purpose (BERT, RoBERTa, BART, DeBERTa, Sentence Transformers) and domain-adapted (BioBERT, ClinicalBERT) transformer models through zero-shot classification and embedding-based clustering. While zero-shot results show BART achieving higher confidence scores (0.646) than domain-adapted models, clustering analysis reveals fundamental challenges: all models achieved low silhouette scores (< 0.25), indicating weak cluster separability and structural limitations of unsupervised embedding-based approaches for clinical sentiment detection. Traditional baseline comparisons demonstrate that transformer embeddings offer marginal improvements over classical methods, emphasizing that this work primarily diagnoses the difficulty of unsupervised clinical sentiment analysis rather than providing a complete solution. These findings highlight the need for hybrid supervised-unsupervised approaches and domain-expert validation in real-world healthcare applications.
Download

Paper Nr:	202
Title:	The Neuropsychiatric Roots of Dementia: Evidence from Depression, Bipolar Disorder, Schizophrenia, PTSD and Anxiety
Authors:	Ayesha Noshin and Arshia Khan
Abstract:	Dementia and Alzheimer’s disease are growing public health concerns. Emerging research suggests that major mental disorders may increase the risk of cognitive decline. This review examines 30 human studies published between 2015 and 2025 that explore the link between psychiatric conditions and dementia. The disorders include major depression, bipolar disorder, schizophrenia, post-traumatic stress disorder and anxiety. Findings from cohort studies, biomarker analyses and neuroimaging consistently show higher dementia risk among individuals with these conditions. Late-onset and treatment-resistant mood disorders often precede dementia which may reflect shared mechanisms or prodromal changes. Schizophrenia is linked to cognitive decline through pathways that may differ from classical Alzheimer’s disease. Biological evidence points to tau and amyloid abnormalities in some older adults with mood disorders. These results highlight the need to consider psychiatric history in dementia risk assessments and support early intervention. This review encourages a more integrated approach to mental and cognitive health for both research and clinical practice.
Download

Paper Nr:	204
Title:	Gender Differences in Cognitive Reserve
Authors:	Yagna Manasa Boyapati and Arshia Khan
Abstract:	This study examines how gender disparities and cognitive reserve affect cognitive aging, with an emphasis on resilience to neurodegenerative diseases such as Alzheimer’s. We evaluate how lifestyle choices, education, and work involvement affect cognitive reserve differentially in men and women based on results from interdisciplinary research. The data demonstrates how sex-specific brain connections and metabolic variations affect cognitive function, with women demonstrating superiority in emotional detection and nonverbal reasoning, and males in visuospatial tasks. The study emphasizes the necessity of gender-sensitive approaches in cognitive health care.
Download

Paper Nr:	211
Title:	Assessing Trustworthiness in Digital Health: Insights from the Brazilian Case of “Meu SUS Digital”
Authors:	Giovana Nunes Inocêncio, Liverson Paulo Furtado Severo and Jean Everson Martina
Abstract:	Trustworthiness in Digital Health Systems (DHS) depends on balancing transparency with data security and privacy. This study proposes a patient-centric evaluation framework to assess this balance and applies it to the Brazilian Meu SUS Digital application. Based on a qualitative literature review, four trustworthiness factors were identified: Data Security, Privacy and Control, Explainability, and Feedback Mechanisms. These factors were operationalized into an auditable checklist focused on patient-accessible transparency beyond regulatory compliance. The case study shows strong performance in Data Security (9/10) and Privacy and Control (7/10), supported by encryption and consent mechanisms, but reveals notable gaps in Explainability (3/10) and Feedback Mechanisms (5/10). The proposed framework translates legal requirements such as GDPR and HIPAA into measurable transparency indicators, offering practical guidance for improving patient trust without compromising data protection.
Download

Paper Nr:	215
Title:	Interpretable Machine Learning for Personalized Profiling of Mild Cognitive Impairment from Daily Activities
Authors:	Budhitama Subagdja, Ah-Hwee Tan, Kenneth Kwok and Iris Rawtaer
Abstract:	Continuous monitoring of individual daily activities is essential to detect mild cognitive impairment (MCI) wherein timely intervention can still be applied to prevent more severe mental decline. Recent approaches in predicting MCI are mostly considering digital biomarkers across individuals but often neglecting specific indicators from a single person over a long period of time. Making this personalized, dynamic, and highly noisy prediction model with irregular distribution of missing information to be explainable and actionable for clinical use, remains a challenge. This paper presents a study on a personalized MCI prediction and profiling from an in-home and mobile cognitive health monitoring integrating data on Activities of Daily Living (ADLs), digital biomarkers, and spatial-temporal features. To address the challenge, two interpretable neural network models for MCI detection are proposed: (i) STEM-MCI integrating spatial-temporal mobility data and ADL sequences; and (ii) Fusion ART, a multi-channel digital biomarker-based predictive model, leveraging multi-modal digital biomarkers and ADL features. Empirical evaluations demonstrate that Fusion ART, based on biomarker features, produces better performance than STEM-MCI, trained on spatial-temporal mobility data. However, augmenting them with ADL data improves predictive performance of both models. Beyond prediction, the models based on biomarkers enable interpretive analysis of internal representations, offering insights into behavioral patterns that differentiate MCI from Normal Cognition (NC).
Download

Paper Nr:	234
Title:	Biochemical Self-Tracking as Personal Informatics: Sensemaking and Actionability of At-Home Micronutrient Tests
Authors:	Zilu Liang, Nhung Huyen Hoang and Thilini Savindya Karunarathna
Abstract:	At-home micronutrient testing is emerging as a new form of personal informatics, allowing individuals to collect biological data and receive digital lifestyle recommendations. However, little is known about how users make sense of such biochemical self-tracking results or whether these insights translate into behavior change intension. In this pilot study, young adults (n = 13) completed a urine-based micronutrient test and reviewed their results through a smartphone app. Although none achieved target levels across all nutrients, participants’ interpretation and prioritization of test results were shaped by familiarity, not physiological importance. Surprisingly, some deficiencies occurred despite consuming nutrient-rich foods or supplements, revealing misalignment between dietary behavior and metabolic outcomes. Our analysis uncovered multiple sensemaking barriers, including difficulty interpreting data visualizations, uncertainty about the relevance of metabolites, and recommendations that did not align with users’ food preferences or resources. This study expands personal informatics to biochemical self-tracking and highlights opportunities to improve data interpretability, personalization, and follow-through in consumer digital nutritional health.
Download

Paper Nr:	248
Title:	From Scoring to Learning: A Smart Framework for Mortality Prediction in Liver Transplantation
Authors:	Zeinab Delaram and Dinesh K. Bhatia
Abstract:	This study introduces a transformative framework that redefines liver transplantation assessment by shifting from traditional rule-based scoring systems to an integrated, knowledge-driven learning approach. The proposed architecture enhances accuracy, interpretability, and personalization in mortality prediction across all phases of transplantation. The clinical reasoning incorporates established clinical scores and expert knowledge to maintain consistency with current medical practice. The Temporal deep learning applies modeling to the UNOS registry data to uncover complex relationships among donor, recipient, and procedural variables, enabling continuous learning from patient outcomes over time. The final personalization step leverages the features extracted from both last phases to generate individualized mortality predictions and treatment recommendations, aligning predictive insights with each patient’s unique clinical profile. Together, these analyses form an Integrated Hybrid Liver Transplantation Intelligence Platform that unites medical expertise with artificial intelligence. By transforming the paradigm from ’using data to score’ to ’using data to learn,’ the framework establishes a transparent, adaptive, and interpretable system. Through its focus on fairness, explainability, and intelligent communication, it empowers clinicians to make informed, patient-centered decisions that enhance predictive precision and support personalized care in liver transplantation.
Download

Paper Nr:	258
Title:	Transfer Learning from Clinical PSG to Real-World Wearables for Sleep Staging
Authors:	Andres Hernandez-Matamoros, Kenji Doya, Sutashu Tomonaga and Haruo Mizutani
Abstract:	Accurate sleep staging is essential for assessing health, fatigue, and circadian regulation, yet polysomnography (PSG) is impractical for long-term, large-scale monitoring. Consumer wearables scale well but usually rely on cardiac and motion signals and rarely include EEG, reducing the fidelity of whole-night sleep staging. We study a neck-mounted wearable (XHRO) that records biopotentials from non-standard electrodes together with PPG in real-world conditions. We propose a transfer learning framework that combines 5-minute EEG spectral features and heart rate (HR) with CEBRA latent embeddings and multinomial logistic regression (MLR). Models trained on PSG-labeled are transferred to XHRO and evaluated against Fitbit sleep labels. Supervised MLR trained on XHRO EEG+HR achieves accuracies of 0.626, 0.515, 0.559, and 0.629, while transfer learning increases them to 0.740, 0.736, 0.739, and 0.643. Overall, clinical contrastive representations can be transferred to a minimally invasive wearable despite non-standard EEG montages and real-world noise, providing a reproducible framework for scalable sleep staging.
Download

Paper Nr:	261
Title:	Playful Aging: A Systematic Review of Game Design Elements for Promoting Well-Being in Older Adults
Authors:	Eliana Silva, Marcos Pinto, Vânia Silva, Susana Pedras, Luís Paulo Reis and Sara A. Silva
Abstract:	Supporting the well-being of older adults has become increasingly important, and serious games offer a promising, accessible and engaging approach. However, designing games that effectively meet the specific usability, accessibility, and engagement requirements of this demographic remains challenging. This study aims to identify game design elements that enhance well-being, self-efficacy, and social interaction among older adults, while examining the impact of usability and accessibility on engagement and the overall game experience. A systematic review was conducted in accordance with PRISMA guidelines, involving searches of PubMed, IEEE Xplore, the ACM Digital Library, Scopus, and ScienceDirect. Out of the 123 records identified, five studies met the inclusion criteria and were evaluated. The systematic review revealed that usability, accessibility and social engagement are critical for the effective design of serious games for older adults. Games with simple navigation, clear feedback, customisable interfaces and social interaction features were found to most effectively enhance motivation, inclusivity and overall well-being. These findings emphasise the importance of designing user-friendly, socially engaging and personalised games based on evidence for this population and highlight the need for future research investigating long-term outcomes.
Download

Paper Nr:	262
Title:	TeleRehaSync: A Multimodal Platform Combining Wearable Physiological and Motion Capture Data for Telerehabilitation
Authors:	Mariana Silva, Carla Quintão and Cláudia Quaresma
Abstract:	Telerehabilitation (TR) has become a key complement to conventional therapies, enabling patients to perform rehabilitation exercises remotely. However, many existing TR solutions lack quantitative measures to objectively monitor performance and support clinical decision-making, limiting accurate assessment of patient progress. To address this gap, this work presents the development and technical validation of TeleRehaSync, a TR platform designed to integrate qualitative and quantitative data and support different user profiles, providing role-specific functionalities for administrators, healthcare professionals, and patients. It enables management of records, creation of personalized treatment plans with embedded exercises, and collection of performance metrics and patient feedback, according to each user’s role. This quantitative monitoring is incorporated via wearable sensors and motion capture systems, providing objective insights into patient performance. A controlled laboratory study was conducted to validate the platform’s data integration and functionality. TeleRehaSync demonstrates strong potential for future clinical translation, supporting personalized rehabilitation programs, enhancing patient monitoring, and enabling data-driven decision-making. By providing objective and comprehensive performance data, the platform aims to optimize recovery, improve adherence to rehabilitation plans, and support professionals in delivering more effective, individualized care.
Download

Paper Nr:	263
Title:	Relationships between the Commissioning Behavior of Statutory Health Insurance Funds and the Assessment Outcomes of the Medical Services in Quality Assurance within Health Insurance Using Shannon Entropy, Correlations and Partial Correlations
Authors:	Reinhard Schuster, Paul-Ulrich Menz and Mareike Burmester
Abstract:	Quality assurance within the Medical Services is defined by a guideline approved by the Federal Ministry and follows the general procedure of the PDCA quality assurance cycle. Territorial differentiation is predetermined by the regional Medical Service, while professional differentiation is defined by ‘occasion groups’ (incapacity for work, rehabilitation, pharmaceuticals, . . . ), which determine the responsible medical assessors. Differences in assessment outcomes may result from varying emphases in the commissioning behavior of health insurance funds. Some results are highly sensitive in internal discussions, and the use of synthetic data to address the issue is currently under consideration. This approach applies Shannon entropy, which is also used in subsequent analyses.Since this year, two variants of assessment have been employed: a longer version with higher requirements (Social Medical Assessment, SGA) and a shorter, time-saving version (Social Medical Expert Statement, SGS). The interactions between outcome codes and occasions within an occasion group, as well as the proportion of SGAs, are examined using Shannon entropy. Correlations and partial correlations are employed, with the modulus determining whether the absolute value increases or decreases when moving from correlations to partial correlations.
Download

Paper Nr:	269
Title:	The Role of Synthetic Data on Cancer Classification and Bias Mitigation
Authors:	John Staunton, Fernando Perez-Tellez and Adrian Byrne
Abstract:	Artificial Intelligence (AI) in healthcare has the potential to transform the diagnosis and treatment of patients worldwide but is vulnerable to bias. This research work carried out bias mitigations for colorectal cancer classification on a tabular dataset of 2,093 real-world patients. The dataset contained irregular time-series data corresponding to each patient’s longitudinal medical history. Four models were evaluated: Logistic Regression, Naïve Bayes, Random Forest and LSTM, with LSTM selected as the baseline due to superior Sensitivity of 0.77. Four bias mitigation approaches were examined: (1) Class balancing (2) Synthetic data generation (SDG) (3) Loss function adjustment and (4) Feature engineering. Loss function adjustment achieved the highest Sensitivity (0.92) and reduced bias across gender and age groups. In contrast, SDG failed to improve Sensitivity or reduce bias. This highlights that current SDG tools have limitations in generating realistic synthetic versions of this complex medical time-series data – but should be revisited in future work as such SDG tools develop and improve. These findings underscore both the promise and current boundaries of bias mitigation strategies in healthcare AI.
Download

Paper Nr:	270
Title:	Integrating Feature Selection and Highway Networks for Early Alzheimer’s Disease Classification from Blood Transcriptomic Data
Authors:	Muhammad Hamza Rafique Bhatti, Asiya Khan, Shakil Awan and Dena Bazazian
Abstract:	Alzheimer's Disease (AD) is an incurable neurodegenerative disease where early diagnosis remains challenging for healthcare practitioners. Current methods like neuroimaging techniques are costly and invasive. Gene expression offers a less invasive and budget-friendly but exhibits limitations like high-dimensional features, small sample size, and class imbalance that led to overfitting and biased models in conventional approaches. In this study, a Deep Learning (DL)-based framework was proposed to address these challenges. The ProWSyn data-balancing technique was applied to overcome the class imbalance problem. The Elastic Net and Boruta feature selection techniques were employed to solve the curse of dimensionality. A highway network model was proposed to learn effectively from high-dimensional data without facing vanishing gradients for accurate AD prediction. A 10-fold cross-validation was used to overcome the risk of optimistic results. The proposed highway model achieved an accuracy of $0.941 \pm 0.014$, precision of $0.929 \pm 0.011$, recall of $0.929 \pm 0.034$, and an F1-score of $0.929 \pm 0.019$. The 10-fold cross-validation confirmed the results, achieving an accuracy of 0.9155, a precision of 0.9209, a recall of 0.9100, and an F1-score of 0.9119. The proposed framework demonstrates a non-invasive computational pathway for early AD detection, potentially reducing reliance on costly neuroimaging techniques.
Download

Paper Nr:	272
Title:	TikTok Restores the Supply-Side Signal for Allergic Rhinitis Nowcasting in England: A Benchmark against Google Trends and Legacy Twitter
Authors:	Krishnamoorthy Manohara, Hannah Béchara and Slava Jankin
Abstract:	Twitter’s post-2023 access changes weakened supply-side infodemiology. We test whether TikTok restores that pillar for allergic rhinitis (AR) nowcasting in England. Using weekly incident AR consultations from the RCGP Research and Surveillance Centre as the clinical target, with MASK-Air® diaries and aerobiological pollen as covariates, we assemble a UK-located TikTok activity series via the Academic Research API and benchmark it against Google Trends and a legacy Twitter archive. A fixed-origin 2020–2024 design evaluates horizons h=0–4 with SNaive, SARIMAX, Elastic Net, Random Forest, and XGBoost. TikTok aligns with incidence at lag 0–1, whereas birch pollen leads by about six weeks. Replacing Twitter with TikTok (GT+TF→ GT+TK) reduces mean absolute error by about 20% for Elastic Net, 12% for SARIMAX, and 3–6% for tree models; RMSE reductions are smaller for Elastic Net/SARIMAX and comparable for tree models. Gains concentrate at nowcasting and attenuate with horizon, and some models do not beat the seasonal naive baseline at longer horizons. Operationally, TikTok’s Academic Research API enables scheduled, zero-cost weekly ingestion, while live Twitter ingestion is currently impracticable. We recommend pairing Google Trends with TikTok for operational nowcasting of AR in England, treating non-clinical streams as UK/GB proxies for the English clinical target.
Download

Paper Nr:	296
Title:	Bridging Clinical Practice and Research: A Web Platform for Ramadan Diabetes Assessment
Authors:	Fatma Zohra Rennane, Mouaad Boucherit, Abdelkrim Meziane and Amar Tebaibia
Abstract:	During Ramadan, the Islamic holy month, all healthy adult Muslims from the age of puberty are obliged to observe fasting, as an act of worship and self-control. Ramadan fasting means abstaining from drink and food, and some medication all day long from sunset to sunrise during a period of 29-30 consecutive days. Fasting for a long time may lead to significant changes in daily routines, including eating habits, sleeping patterns, physi-cal activity levels, as well as medication intake schedules. These changes can affect the metabolic balance of diabetic patients and expose them to an increased risk of hypoglycemia, hyperglycemia, dehydration and other metabolic complications especially for individuals with poorly controlled diabetes and comorbidities. For that reason, physicians need to check and evaluate their patients’ ability to fast without risk to give the right advice before Ramadan, and then to follow up and analyze the effects of fasting afterwards. However, physicians face challenges in assessing fasting risks, documenting clinical data, and following up after Ramadan due to reliance on manual calculation and paper-based tracking. To address this challenges, a web-based platform has been developed, to help them make better decisions about whether a patient can fast or not, and making data collection easier for clinical studies and research, as well as providing useful data analysis tool to understand how fasting impacts patients’ health.
Download

Paper Nr:	297
Title:	High Users Management with AI: Predicting Emergency Department Readmission for Resource Optimization
Authors:	Inês Gonçalves, Mariana Dias, Inês Metelo, Marta B. Lopes, Luís Silva, Ana Januário, Andreia Mesquita, Carla Menino, Carla Vidinha, Fausto Silva, Mafalda Gonçalves, Maria José Guimarães, Pedro Casimiro, Rui Malha, Ana Londral and Federico Guede-Fernández
Abstract:	Patients who visit Emergency Department (ED) Services more than 10 times in 12 months are designated as High Users (HU). These patients have complex care needs, at times linked to multimorbidity or socioeconomic vulnerability. These complex care needs require effective interventions in the health services in order to mitigate the impact in the quality of life of these patients, as well as the increased strain on health systems associated with high frequency of readmissions and consequent resource allocation. Case Management strategies have shown potential, but their effectiveness depends on accurate patient identification for intervention. As such, the present study aimed to develop Machine Learning (ML) and Deep Learning (DL) models for predicting future ED readmissions using retrospective clinical data from 972 patients from 2016 to 2020. The preliminary results indicate that ML models outperformed DL models, particularly for shorter prediction horizons. Specifically, the random forest model using 1.5 years of historical data achieved an AUC of 0.69 for predicting whether a patient would have at least three readmissions within 3 months, while the Extreme Gradient Boosting model reached an AUC of 0.70 for predicting at least five readmissions within 6 months.
Download

Paper Nr:	298
Title:	Digital Twins in Healthcare: Conceptualization and Use Case Definition
Authors:	Mariana Dias, Federico Guede-Fernández, Ana Januário, Anabela Silva, Carolina Capitão, Hugo Gamboa, Joel Ribeiro, Julian Perelman, Luís Silva, Mariana Carvalho, Marta Moreira Marques, Rui Malha and Ana Londral
Abstract:	Digital Twin (DT) technology offers a novel opportunity to address inefficiencies in current healthcare work-flows by creating dynamic, AI-driven digital replicas of patients that integrate real-world multimodal data. DTs can predict patient trajectories and simulate interventions before implementation, therefore enhancing outcome precision and supporting resource-efficient care. In this work, a methodology for DT conceptualization and use case definition is proposed. The methodology is grounded in participatory action research to guarantee that the resulting solution effectively addresses real-world needs and delivers practical value. Through a co-creation process, involving healthcare professionals, administrators, patients, and health policy experts, diverse stakeholder perspectives are integrated. The methodology is illustrated through two case studies demonstrating the relevance of DT development and implementation: (i) management of high users in Primary Care Services and (ii) optimization of care pathways for patients with hip and knee osteoarthritis. Beyond these case studies, the broader goal is to establish a comprehensive, adaptable methodology for designing and defining use cases of DT technologies applicable across multiple healthcare contexts.
Download

Paper Nr:	315
Title:	GeoCrecheAI: An Intelligent System for Georeferenced Prediction of Daycare Slots to Support Public Policies
Authors:	João Pedro Barbosa Santana Costa, Pedro Almir Martins de Oliveira, Rossana Maria Castro Andrade, Evilasio Costa Junior, Wilson Castro, Victória Tomé Oliveira and Pedro A. Santos Neto
Abstract:	Early childhood is a crucial moment in the personal development of the citizen. In this sense, education emerges as one of the primary pillars for ensuring the proper development of children in their first years of life, with public administration responsible for allocating resources to ensure the effectiveness of this process. Given this context, this work aimed to integrate Artificial Intelligence and georeferencing tools to optimize public administration decision-making regarding the allocation of daycare slots. To achieve this objective, a methodology based on Technical Action Research (TAR) was chosen, dividing the actions into four phases: investigation and design, development and validation, empirical evaluation, and technology transfer. As a partial result, the architecture of an application was developed to assist the government of the State of Maranhão in allocating daycare slots, along with the identification of the ARIMA model as the most suitable algorithm for the system, showing superior performance with R\textsuperscript{2} = 0.998 and RMSE = 32.112.
Download

Paper Nr:	323
Title:	Anxiety Map: Longitudinal Monitoring of Anxiety High School Using Mobile Health
Authors:	Dáryo G. S. Arruda, Glória M. Alexandre, Pedro Almir M. de Oliveira, Rossana Maria C. Andrade, Evilasio Costa Junior, Wilson Castro, Victória Tomé Oliveira and Pedro A. Santos Neto
Abstract:	Anxiety among high school students is a growing challenge that significantly affects well-being and academic performance. Furthermore, reactive interventions often fail to account for the cyclical and longitudinal nature. The Anxiety Map project aims to address this limitation by presenting its longitudinal, continuous monitoring system, supported by digital health technologies, along with the initial development and prototyping stages. The proposed framework combines a mobile application (developed with MIT App Inventor) for collecting weekly anxiety data using the Beck Anxiety Inventory (BAI) with a key analysis phase. In this phase, machine learning methods are used to correlate anxiety spikes with key events on the academic calendar. The goal is to develop a robust predictive model that converts anxiety signals into actionable insights. This will enable the Anxiety Map to assist school administrators in transitioning from a reactive stance during crises to proactive and preventive planning, serving as a key resource in supporting student mental health.
Download

Paper Nr:	329
Title:	In-a-Container: Sustainable Digital Health Solutions for Resource-Challenged Communities
Authors:	Sylvia Delpratt, John Chelsom, Rupert Gammon, Michael Odeh, Naveed Dogar, Ngonidzashe Daniel Sakupwanya, Jennifer Ramirez, Adam Retter and Rachael Brooker-Langston
Abstract:	Access to comprehensive healthcare, education, and enterprise services in remote and resource-challenged communities remains a significant global challenge. This paper describes the In-A-Container initiative, a four-stage approach to delivering technology services that support telemedicine, community hospitals, education and community enterprise using secure, battery powered containers, photovoltaic panels and satellite communications. The first stage, for telemedicine, is in a suitcase-sized container small enough to transport in the cabin of a commercial airliner and delivered on-site using an electric bike; the container supporting a community hospital is a quarter-sized shipping container, delivered on an electric tuk-tuk; the largest containers for community education and enterprise are half-sized shipping containers, delivered on an electric truck. The final stage Enterprise-In-A-Container includes a battery bank that provides power for local community enterprises and a rental income that supports the ongoing operations. The first prototype of the Telemedicine-In-A-Container has been developed and tested in the laboratory; the second prototype is due for completion by the end of 2025. First pilots are scheduled for deployment in the first half of 2026 at the Darul Ikram Orphanage in Porto Novo, Benin and the El Roi Inland Missions Hospital in Jos, Nigeria.
Download

Paper Nr:	331
Title:	Integrating Structured and Unstructured EHR Data for Early Differentiation of Dementia with Lewy Bodies and Parkinson’s Disease Dementia Using Machine Learning
Authors:	Sophia Yang and Betina Idnay
Abstract:	Dementia with Lewy bodies (DLB) and Parkinson’s disease dementia (PDD) are frequently misdiagnosed due to overlapping clinical features, leading to delayed treatment outcomes for patients. This study leverages deidentified electronic health record (EHR) data from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) to develop machine learning models that distinguish between DLB and PDD. The XGBoost classifiers combining structured and unstructured features achieved the highest performance (AUC = 0.921), outperforming models trained on either data type alone. The results of the study can help healthcare providers improve differentiation between DLB and PDD and suggest avenues for scalable application to other dementia subtypes.
Download

Paper Nr:	332
Title:	Digital Interventions for Depression and Anxiety: A Systematic Synthesis of Chatbots, App-Based CBT, and Immersive VR/XR
Authors:	Zannatul Ferdousee, Varshini Bhavanam, Maggie Beach and Arshia Khan
Abstract:	This review synthesizes recent evidence on digital interventions for depression and anxiety, spanning chatbots, app-based cognitive behavioral therapy (CBT), and immersive VR/XR. Following PRISMA 2020, searches of PubMed and Web of Science (2021–2025) identified primary studies evaluating symptom change, feasibility, or engagement. Across modalities, randomized and quasi-experimental studies generally reported short-term reductions in depressive and anxiety symptoms relative to comparison conditions, with consistent signals for CBT-oriented chatbots and apps and additional improvements reported in VR/XR studies. Engagement varied widely and was supported by empathic tone, brief structured activities, light gamification, and contextual tailoring; repetitive or generic responses reduced adherence. Common limitations included modest samples, brief follow-up, reliance on self-report, variable control conditions, and inconsistently described safety and privacy practices. Overall, digital tools appear feasible, acceptable, and potentially effective short-term supports; priorities include larger and longer trials with diverse populations, clinician-verified outcomes, robust crisis-escalation and data-governance procedures, and evaluation of blended human–AI models.
Download

Paper Nr:	349
Title:	Forensic Medicine and Radiology in the Age of Deepfakes: Prompt Robustness and Reasoning Advantage in Multimodal LLM Detection of Radiological Image Manipulation
Authors:	Eren Çamur, Turay Cesur and Yasin Celal Güneş
Abstract:	High-fidelity image manipulation techniques increasingly threaten the integrity of radiological images, posing significant risks for both clinical decision-making and forensic medicine. This study investigated whether multimodal large language models (LLMs) with explicit reasoning capabilities outperform non-reasoning models and a human radiologist reference in detecting altered radiological images, and whether model performance is sensitive to prompt design. In a prospective, image-only benchmarking study, 100 radiological images (50 original and 50 altered) were evaluated by six multimodal LLMs across three prompting strategies, including zero-shot, minimally role-based, and structured forensic prompts. Model performance was compared with a blinded board-certified radiologist serving as a human reference baseline. Reasoning-capable models consistently achieved higher accuracy than non-reasoning models and demonstrated reduced prompt sensitivity, with peak performance reaching 74%. In contrast, non-reasoning models and the human reference showed limited ability to detect subtle manipulations. These findings suggest that reasoning-oriented multimodal LLMs may serve as supportive integrity-screening tools in radiology and forensic workflows, although further validation in larger, multi-reader studies is required.

Paper Nr:	353
Title:	Identifying a Critical Knowledge Infrastructure Gap: Cross-Database Bibliometric Analysis of Public Health Emergency Preparedness and Response Informatics
Authors:	Nikolay Lipskiy and Gora Datta
Abstract:	Public Health Emergency Preparedness and Response Informatics (PHEPRI) was defined in 2017 by N. Lipskiy and J. Tyson as the interdisciplinary science managing data and information across emergency preparedness and response lifecycles, supporting situational awareness and decision-making. COVID-19 exposed core PHEPRI challenges in interoperability, multi-jurisdictional data integration, and federated system governance. In 2023, ISO 5477:2023—developed through three-year collaboration among 34 nations—established technical specifications for PHEPR information systems interoperability, formally incorporating the 2017 PHEPRI definition. This bibliometric analysis quantified PHEPRI's visibility in indexed literature (PubMed, Scopus; 2015–2025), assessed knowledge infrastructure gaps, examined COVID-19 temporal trends, and identified opportunities to strengthen emergency preparedness capacity. Parallel title/abstract searches targeted PHEPRI terminology, related informatics fields, data science, professional roles, and emergency preparedness contexts. Cross-database correlation was very strong (r=0.94, p<0.001). PHEPRI terminology appeared in zero (PubMed) to one (Scopus) publication over the decade. Emergency preparedness literature showed ~50:1 "data" versus "informatics" framing; data science publications outnumbered informatician publications ~105:1. During COVID-19, data science grew 2.3-fold while informatician publications increased 1.5-fold; emergency preparedness informatics remained stable. Despite COVID-19 challenges and ISO 5477:2023 standardization, PHEPRI literature remains minimal. The 105:1 ratio risk conflating complementary roles: data scientists extract insights from existing data; PHEPRI informaticians architect federated systems enabling data collection, integration, and cross-jurisdictional use. Enhanced scholarly discourse would clarify competencies, guide workforce planning, and strengthen emergency preparedness.

Paper Nr:	360
Title:	A Methodology for Assessing Data Quality in the Analysis of Process Mining in Healthcare: A Case Study in a Pediatric Hospital
Authors:	Evelyn Salas, Esteban Chiu, Juan Diego Rodríguez, Matias Cornejo, Michael Arias, Macarena Silva, Arturo Solís and Eric Rojas
Abstract:	Process mining (PM) is increasingly used to analyze and improve healthcare processes; however, its results strongly depend on the quality of event data extracted from heterogeneous clinical information systems. Generic data analytics frameworks such as CRISP-DM provide high-level guidance, but they do not sufficiently address the domain-specific challenges of data quality assessment required for healthcare-oriented PM studies, particularly regarding event log construction and clinical validation. This paper proposes a structured methodology for assessing and improving data quality in healthcare process mining. The methodology refines generic frameworks by incorporating PM-specific data quality indicators, explicit validation steps, and iterative feedback loops between analysts and clinical domain experts, with a particular focus on the construction of reliable event logs. The proposed approach is instantiated through a real-world case study conducted at a tertiary pediatric hospital, using anonymized outpatient ophthalmology data extracted from an electronic health record system. The application of the methodology involved iterative data cleaning, quality assessment, and expert validation, leading to a substantial reduction in inconsistencies and missing values and resulting in a validated event log suitable for PM analysis. The results show that a systematic and transparent data quality assessment process improves the reliability and interpretability of PM outcomes in healthcare settings. The study also discusses limitations related to manual data extraction and highlights opportunities for automation and validation in other clinical contexts.
Download

Paper Nr:	369
Title:	SUGAR: Surgical sUturing GrAdeR
Authors:	Laura Manso
Abstract:	Suturing is a fundamental surgical skill evaluated by surgical trainers with methods that lack precision and reproducibility, while being time-consuming, resource-intensive, and expensive. This work aims to deliver a proof-of-concept for an objective, reproducible, and cost-effective method to evaluate suturing skills. The Surgical sUturing GrAdeR (SUGAR) tool applies Computer Vision (CV) techniques and an expert system to assess 13 hard-coded suture quality metrics. This work produces a novel CV-based system that evaluates key suture quality parameters with a precision of 85.68% and a recall factor of 61.75%, exhibiting promising potential to be used as an educational tool for developing basic suturing proficiency using self-directed learning.
Download

Paper Nr:	407
Title:	GlucoPatrol: A Multi-Stakeholder Dashboard for Contextual Exploration of Continuous Glucose Monitoring Data
Authors:	Thilini Savindya Karunarathna, Nhung Huyen Hoang and Zilu Liang
Abstract:	Continuous glucose monitoring (CGM) is increasingly used to support diabetes management and research. However, many existing dashboards provide limited integration of behavioral and other streams of physiological data, restricting contextual interpretation of glucose dynamics. Moreover, patients, clinicians, and researchers engage with CGM data for different purposes and possess varying levels of domain expertise and numerical literacy, leading to distinct requirements for data abstraction, interaction, and visual explanation. In this position paper, we present GlucoPatrol, a web-based dashboard that supports contextual visual exploration of longitudinal CGM data in free-living conditions. GlucoPatrol integrates CGM data with user-logged events and wearable physiological signals within a unified data backbone, and provides role-specific visual interfaces that adapt interaction depth and visual abstraction to stakeholder needs. The Patient Interface emphasizes intuitive daily glucose profiles and summary indicators to support awareness and reflection. The Clinician Interface focuses on compact, low-overhead visualizations of glucose trends and key metrics to support efficient remote review. The Researcher Interface supports flexible individual and cohort-level exploration of multimodal data, combining interactive visualization and analytical tools for glucose dynamics, physiological signals, and event annotations. By foregrounding contextual visualization and role-aware interaction, GlucoPatrol illustrates an approach to supporting multi-stakeholder exploration of CGM data in real-world settings.
Download

Paper Nr:	439
Title:	Digitalizing Mindfulness: A Systematic Review of Virtual Reality Applications for Chronic Pain Management
Authors:	Moisés Moreira, Estela Vilhena, Anabela Marques, Vitor Carvalho and Duarte Duque
Abstract:	Chronic pain, defined as pain persisting for more than three months, affects over 20% of U.S. adults. Mindfulness-based therapies have shown promise in pain management. Virtual reality (VR) is a potential platform to deliver such interventions, using immersion and gamification to deliver these therapies. This systematic review evaluates the current state of VR applications for chronic pain management incorporating mindfulness-based components. A total of 63 records were screened, and 6 studies were eligible for review. The mean age of the participants was around 50 years old, and their chronic pain related to different factors including fibromyalgia, back pain, cancer-related pain, and military-related trauma. Different headsets were used (Quest 2, Pico 4, Oculus Go and HTC Vive), a range in duration and frequency, from single sessions to 8-week daily interventions were found. We noticed that the apparent female dominance in the population (68%) is an artifact of one large study; without it, the field is gender balanced. The results reported better reductions in pain intensity and stress, high adherence, and good acceptability but it didn’t yield better results than control conditions. Adverse effects were generally low, but we found that studies using active checklists reported much higher rates that were sometimes downplayed. With such a small sample, we will need a larger body of work and more RCTs to understand where we can improve in this field.
Download

Paper Nr:	440
Title:	Towards Synthetic Generation of Daily Routines and Anomaly Detection with Wearable Sensors for Dementia Care
Authors:	Mariana Carvalho, Inês C. Rocha, Marcelo Arantes, Ana Rita Freitas, José Soares, Demétrio Matos, Pedro Morais and Vítor Carvalho
Abstract:	This paper presents an approach to generating synthetic daily routines and systematically injecting behavioural anomalies to support AI-based monitoring in dementia care. Given the ethical and practical challenges of acquiring labelled data from individuals with dementia, a probabilistic routine generator was developed grounded in observational literature, simulating realistic activity patterns using time-of-day segmentation, activity-specific probability distributions, typical duration ranges, and individualized variability in timing and sequencing, alongside controlled context-sensitive deviations. Using this dataset, we trained a Gated Recurrent Unit (GRU)-based sequential anomaly detection model that achieved high classification performance, including an F1-score of 0.959 for anomalous activity steps. The results demonstrate the feasibility and value of synthetic data in developing personalized monitoring systems for early detection of behavioural changes. This approach enables reproducible model benchmarking and lays the groundwork for real-world deployment in wearable-based dementia care systems.
Download