HC@AIxIA+HYDRA 2025 is over! We'll be working hard for the 2026 edition. Stay tuned!

Awards and Accepted Papers

HC@AIxIA + HYDRA 2025

International Joint Workshop of Artificial Intelligence for Healthcare (HC@AIxIA) and HYbrid Models for Coupling Deductive and Inductive ReAsoning (HYDRA)

Bologna, Italy, 25-26 October 2025

Awards

The Program Committee of HC@AIxIA + HYDRA 2025 is proud to present two distinct awards: one is granted to the best among all ORIGINAL works accepted to the workshop, while the other is awarded to the best among full that features a STUDENT as main author.

Best Paper AWARD

WINNER:

👑 Expert-Guided POMDP Learning for Data-Efficient Modeling in Healthcare - by Marco Locatelli; Arjen Hommersom; Roberto Clemens Cerioli; Daniela Besozzi; Fabio Antonio Stella

Other shortlisted works:

Automated Machine Learning to Enhance Knowledge Retrieval in Retrieval-Augmented Generation Pipelines - by Matteo Magnini; Gianluca Aguzzi; Leonardo Sanna; Simone Magnolini; Patrizio Bellan; Mauro Dragoni; Sara Montagna
Triage Discrimination: Myth or Reality? - by Justin Armanini; Fabio Antonio Stella

Best Student Paper AWARD

WINNER:

👑 LUME-DBN: Full Bayesian Learning of DBNs from Incomplete data in Intensive Care - by Federico Pirola; Fabio Antonio Stella; Marco Grzegorczyk

Other shortlisted works:

Triage Discrimination: Myth or Reality? - by Justin Armanini; Fabio Antonio Stella
Using Temporal Features to Improve Accuracy in Multivariate Pattern Analysis (MVPA) of M/EEG Data - by Fatemeh Barazesh Morgani; Erin Goddard; Gustavo Enrique Batista

List of Accepted Papers

Please find next the list of accepted papers along with authors, abstracts and some additional pieces of information. An accessible version of the same list can be found here.

Depth Sensor based AI-services for Nursing Homes

Authors: Rinu Elizabeth Paul; Lucas Deichsel; Tanja Schultz

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): YES
Summer School Track: NO

Abstract: Falls are a frequent and critical event among older adults, often leading to severe consequences such as injuries, loss of independence, fear of walking, and even death. Therefore, it is essential to detect and prevent them. AI services in homes and nursing facilities enable early fall detection and timely support. Another major concern in nursing homes and clinics is the frequency of bed-exit attempts by residents or patients who cannot stand independently. These individuals often overlook their limitations and attempt to get out of bed alone, increasing the risk of falling. Our approach focuses on bed-exit detection to notify caregivers in real time, potentially preventing these incidents. We present a system that uses depth video to detect falls and bed-exit attempts among older adults. A configuration interface is developed to provide seamless access to fall and bed-exit detection functionalities and their results. The interface is compatible with various operating systems and optimized for both CPU and GPU versions, making it suitable for home and nursing facility applications. Our system achieves 68% accuracy and a 69% F1-score for fall detection, and 82% accuracy and an 88% F1-score for bed-exit detection. Room occupancy detection, which determines whether an older adult is present in the room, achieved 92% accuracy and an 89% F1-score. This work demonstrates significant potential to enhance services that improve assisted living environments.

Generating patient cohorts from electronic health records using two-step retrieval-augmented text-to-SQL generation

Authors: Angelo Ziletti; Leonardo D'Ambrosi

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Clinical cohort definition is crucial for patient recruitment and observational studies, yet translating inclusion/exclusion criteria into SQL queries remains challenging and manual. We present an automated system utilizing large language models that combines criteria parsing, two-level retrieval augmented generation with specialized knowledge bases, medical concept standardization, and SQL generation to retrieve patient cohorts with patient funnels. The top-performing configuration achieves 0.75 F1-score in cohort identification on EHR data, effectively capturing complex temporal and logical relationships. These results demonstrate the feasibility of automated cohort generation for epidemiological research.

Multi-View Multi-Instance Learning for Health Recovery Monitoring in Older Adults

Authors: Veselka Boeva; Alexander James Orest Ojutkangas; Shahrooz Abghari

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: In this study, we propose a multi-view version of the BARTMIP algorithm, which can learn from data for which only coarse-grained labels are provided. The proposed MV-BARTMIP algorithm can deal with weakly annotated multi-modal data and is able to handle cases of missing data, including modalities and views. The performance of the MV-BARTMIP algorithm is evaluated in a scenario involving the monitoring of older adults' health recovery at home following hip replacement surgery. The performance of MV-BARTMIP and the traditional BARTMIP is benchmarked against several baseline solutions. The experimental results demonstrate that approaching the use case as a multi-view, multi-instance learning task results in more robust and interpretable models. MV-BARTMIP shows superior performance to the best baseline model in all but one scenario, where the results are comparable. Furthermore, its performance is comparable to that of the BARTMIP algorithm, which outperforms the best baseline model in all experimental scenarios.

SEDformer: Path Signatures and Transformers to Predict Newborns Movement Symmetry

Authors: Philippe Rambaud; Arpad Rimmel; Imen Trabelsi; Justine Zini; Alexandre Wodecki; Emmanuelle Motte Signoret; François Jouen; Joanna Tomasik; Jean Bergounioux

Accepted As: Short/position/discussion paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "In pediatric healthcare, the accurate prediction of movement symmetry in newborns is crucial for early detection of potential motor development issues, enabling timely interventions and personalized treatment plans. Traditional time series forecasting methods often struggle to capture the complex, non-linear dependencies present in real-world temporal data, particularly in the context of infantile movement patterns. To address these challenges, we introduce the Signature Enhanced Decomposition Transformer (SEDformer), a model that integrates path signatures from rough path theory into the Transformer architecture. Path signatures provide a robust mathematical framework for encoding higher-order dynamics in time series data, allowing SEDformer to model both linear and non-linear temporal dependencies effectively. Our experiments demonstrate that SEDformer achieves superior performance on a specialized dataset of infantile movement patterns, outperforming existing models. This highlights the model's potential for practical medical applications, particularly in the early detection and intervention of motor development issues in infants. The results underscore the value of path signatures in enhancing time series forecasting models and open avenues for further investigation into their application in complex clinical forecasting tasks."

Adaptive Compliance Control for User Centric Resistance in Upper Limb Rehabilitation Robots

Authors: Kiwon Yeom

Accepted As: Short/position/discussion paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Upper limb rehabilitation roborts provide consistent training to support the recovery from injuries caused by accidents, illness, or aging to this end. In general, rehabilitation robots employ impedance control strategy to produce mechanical resistance with respect to external force. However, conventional impedance control uses fixed spring coefficient and damping value, resulting in uniform mechanical resistance regardless of individual differences in strength or rehabilitation progress. As a result, patients may experience unintended excessive motion under high external forces, increasing the risk of injury. To address this issue, this paper proposes a variable impedance control strategy that dynamically adjusts the mechanical resistance of robots based on real-time user force input. By considering the patient’s physical condition and rehabilitation stage for providing suitable resistance, the proposed method enhances user safety, delivers more personalized resistance, and contributes to more effective and consistent rehabilitation training.

An approach leveraging Deep Learning and Stream Reasoning for dynamic task assignments balancing productivity and well being

Authors: Luca Laboccetta; Giorgio Terracina; Francesco Calimeri; Simona Perri; Massimiliano Ruffolo; Davide Iacopino; Marta Maria; Salvatore Iiritano

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Employee well-being is an increasingly crucial factor for sustainability and efficiency in modern industry, especially within the context of Industry 5.0, which places humans at the center of production processes. This article presents some innovative aspects of the InCoP platform, which is part of a more general project developed under the NRRP MUR initiative FAIR (Future AI Research): Green-aware AI, and integrates Stream Reasoning (SR) and Machine Learning (ML). In particular, we focus on the real-time monitoring of operators’ psychophysical well-being and optimization of task assignment in a real industrial bakery environment. Here, an Assignment Module integrates results from ML-based models, used to predict stress, with the deductive capabilities of the SR engine, used to dynamically reassign operators to departments and suggest breaks in cases of persistent or dangerous stress levels. The goal is to minimize worker stress and prevent overload, while maintaining adherence to the production plan. Notably, the dynamic assignment module is fully integrated within the InCoP platform and leverages also sensors data coming from smartwatches and environmental monitoring devices which are fed into the system through Kafka, Elasticsearch, and MongoDB based technologies.

A boxology-based analysis of design patterns for neuro-symbolic medical decision making systems

Authors: Chi Him Ng; Annette ten Teije; Frank van Harmelen

Accepted As: Non-original communication
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "This paper presents a structured analysis of neuro-symbolic design patterns for medical decision-making systems through a graphical notation (the ""boxology"") for neuro-symbolic architectures. We formalize and validate five archetypal neuro-symbolic architectures initially defined through textual descriptions and informal diagrams by Kierner et al. We systematically define and refine these archetypes across 68 systems from the literature. Our contributions include: (i) a formalization of these archetypes, (ii) empirical validation of these archetypes via system refinements, (iii) enhanced understanding of neuro-symbolic integration in clinical applications, and (iv) establishing the boxology as a robust tool for comparative architectural analysis. The findings indicate that the elementary patterns within the boxology framework remain consistent across clinical applications, offering new avenues for systematic development and comparison in neuro-symbolic AI for healthcare. This paper has been published in the proceedings of the 23rd International Conference on Artificial Intelligence in Medicine (AIME 2025, https://link.springer.com/chapter/10.1007/978-3-031-95838-0_33). The presentation at HC@AIxIA + HYDRA will place more emphasis on the neurosymbolic aspects of the work."

MultiH-EU: An AI-Driven Decision Support System for Multi-Disease Diagnosis

Authors: Diogen Babuc; Teodor-Florin Fortiş

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Recent advancements in artificial intelligence (AI) have substantially impacted healthcare, particularly in disease diagnosis and prediction. This paper introduces an AI-driven decision support system (DSS) designed for multi-disease diagnosis. The DSS leverages custom deep learning architectures: convolutional neural networks, vision transformers, and hybrid models to perform precise and efficient medical image classification and predictive analytics. Image preprocessing is enhanced through masked and variational autoencoders. To enhance model performance and adaptability, the system employs neural architecture search methods guided by stochastic principles. The proposed DSS addresses significant challenges in deploying AI in healthcare, including cross-domain generalization and secure integration within medical settings. Federated learning techniques are incorporated to enable decentralized model training across multiple medical institutions. This integrated DSS showcases the transformative potential of AI in healthcare.

Application of Nonlinear Autoregressive Neural Network for Prediction of HIV Infections in Poland

Authors: Marek Landowski; Anna Landowska

Accepted As: Short/position/discussion paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: The article presents the application of a nonlinear autoregressive neural network to predict the HIV infection rate in Poland. The prediction was made using neural network models with three training functions: Levenberg-Marquardt, Bayesian Regularization, and BFGS Quasi-Newton. Real data was used to predict the HIV infection rate. Due to lower testability during the COVID-19 pandemic, these data indicated a lower infection rate during this period. This fact did not significantly affect the prediction using the NAR neural network. The prediction results obtained using the NAR neural network, linear regression and exponential smoothing methods were compared. The presented research is important due to the increasing number of HIV infections worldwide. The use of artificial neural networks to predict the infection rate can help locate places that require intensified preventive measures.

RAG-Enhanced LLMs for Interactive Explainability in Clinical Decision Support Systems

Authors: Lorenzo Mannocci; Francesca Naretto; Lucia Passaro; Anna Monreale

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Artificial Intelligence (AI) is increasingly embedded in high-risks decision-making systems across domains such as healthcare, finance, and the judicial system. As reliance on predictive models grows, so does the demand for Explainable AI (XAI) to ensure transparency, trust, and interpretability. However, existing XAI approaches often fail to present explanations in a format accessible to domain experts, limiting their practical utility. Large Language Models (LLMs) have recently emerged as a bridge between complex models and end-users, offering natural language explanations. Yet, their use introduces new risks, including hallucinated outputs and lack of source traceability. To address these challenges, we propose a Clinical Decision Support System (CDSS) that not only combines standard predictive models with XAI techniques but also incorporates a Retrieval-Augmented Generation module as an interactive enhancer. This module grounds explanations in verified medical knowledge, reformulates technical outputs into a human-like format, and enables users—such as clinicians—to actively query the system. By fostering an interactive, human-in-the-loop environment, our approach empowers domain experts to explore model decisions, contextualize explanations, and build trust in AI-assisted diagnostics. The study concludes with preliminary experiments validating the proposed methodology.

An automatic alignment strategy for medical ontologies: the MedPix 2.0 case study

Authors: Luigi Colucci Cante; Mariangela Graziano; Irene Siragusa; Beniamino Di Martino; Roberto Pirrone

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: In this paper, we present an automatic alignment strategy for the medical domain. Standard medical thesauri as the 11th Revision of the International Classification of Diseases (ICD-11) and a vocabulary of radiological terms (RadLex) are aligned with a multimodal medical data set (MedPix 2.0). The alignment process was conducted over their ontological representation and using natural language processing techniques and large language models for annotation purposes across both corpora. The obtained automatic pipeline ensures the terminology alignment between the terms in MedPix 2.0 and the standard terminology in ICD-11 for terms such as body parts and diseases. Data and developed pipeline are available on https://github.com/CHILab1/MedOntoAlignment

Mining Scientific Abstracts for Pesticide-Related Health Risks

Authors: Roushnaty Ali Yamani; Naira Abdou Mohamed; Aouane El Mahjoub

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: The widespread use of pesticides poses significant public health risks, necessitating effective analytical tools. We present an AI-driven platform that analyzes environmental and pesticide-related texts to extract key information such as chemical compounds, diseases, and symptoms. Using the manually annotated MACCROBAT dataset, we fine-tuned RoBERTa and compared it to BlueBERT, which outperformed with 98.10% accuracy and a 95.32% F1 score. The optimized BlueBERT model was applied to a dataset from the Beyond Pesticides website to validate our approach on real-world data. Furthermore, Llama3 was used to extract additional entities, such as countries, affected insects, and other contextual data. This integrated AI and environmental health approach aims to support informed pesticide regulation.

KaiMed: A Hybrid Multi-Agent AI Platform for Clinical Reasoning and Cross-Specialty Discovery

Authors: Lorenzo Megliola; Francesco Barbato; Emanuele Ielo; Filippo Sorichetti

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "KaiMed is a modular AI platform designed to support adaptive clinical reasoning in complex decision-making scenarios. Unlike traditional Clinical Decision Support Systems (CDSS)—which are typically static, narrow in scope, and confined to individual specialties—KaiMed integrates symbolic and sub-symbolic approaches within a unified architecture to enable scalable, explainable, and cross-specialty medical reasoning. The platform incorporates: (1) a clinically grounded knowledge graph built from curated clinical trials and peer-reviewed literature; (2) a semantic retrieval engine leveraging vector embeddings to access unstructured scientific content; and (3) a multi-agent reasoning layer that coordinates specialized agents responsible for diagnosis, treatment, literature validation, and referral. Each agent operates within a transparent workflow supervised by a central coordinator, dynamically interacting with both the knowledge graph and the semantic index to enable hybrid, context-aware reasoning that mirrors expert clinical thought processes. The system is evaluated in the context of Inflammatory Bowel Disease (IBD), where it demonstrates high scores in clarity, relevance, and perceived usefulness. Its performance across structured metrics (QAMAI, TDS, ACCS) consistently exceeds baseline results from standalone LLMs, underscoring the value of coordinated agents and knowledge-aware reasoning. Structured source attribution is already implemented and undergoing further refinement to enhance traceability. KaiMed is actively being extended to additional domains—including urology and chronic rhinosinusitis—leveraging shared ontologies and a unified framework to support cross-domain discovery. Rather than replacing clinical expertise, KaiMed amplifies it—bridging fragmented knowledge, surfacing non-obvious connections, and enabling modular, transparent decision-making across medical specialties."

Hidden Rhythms: an Embedding Framework for Downstream Tasks Validated on ECG Data

Authors: Simone Bartucci; Edoardo De Rose; Francesca Filice; Francesco Calimeri; Simona Perri

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: YES

Abstract: "Foundation models (FMs) are large-scale deep neural networks trained on vast and diverse datasets, capable of learning transferable, domain-agnostic representations that enhance performance across a wide range of downstream tasks. These models have attracted significant attention across multiple fields, potentially reshaping standard paradigms of model design. However, adapting FMs to specific tasks typically requires computationally intensive fine-tuning and large labeled datasets, which can be significant limiting factors, especially in resource-constrained settings. In this work, we explore an alternative strategy: leveraging pre-trained FM embeddings (also called vector embeddings) as inputs to downstream supervised models. By utilizing these rich, general-purpose representations, the approach retains the expressive power of FMs while significantly reducing both computational costs and data requirements. We present a general framework, namely CURE-FM(Context‑aware Use of data REpresentations with Foundation Models), designed to extract embeddings from self-supervised FMs trained on raw data, and to leverage downstream models for generic tasks. We present an application of CURE-FM relying on raw time series data; in particular, the application is geared towards binary classification tasks over multi-lead electrocardiogram (ECG) signals. We report a comprehensive evaluation, including ablation studies on FM and classifier architectures, structural and explainability analyses of the embeddings, and a comparison between our approach and common fine-tuned based approaches, with the aim of investigating how different architecture and approaches affect performance and generalizability. Experimental results show that embedding-based approaches can provide a scalable, robust, and efficient solution for downstream tasks, and that CURE-FM holds significant promise for advancing ECG analysis."

Which Metadata Matters? Evaluating Predictive Features for Environmental Time Series

Authors: Ann-Kathrin Dörr; Tina Schmidt; Jens Schoth; Ivana Kraiselburd; Folker Meyer

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: Forecasting complex, multivariate time series using machine learning presents significant challenges, particularly when dealing with high-dimensional biological data. One such application is the prediction of microbial community dynamics in wastewater treatment plants (WWTPs), which is directly relevant to public and environmental health. Previously, we addressed this problem by training different model architectures and comparing their performance. In this study, an attempt was made to enhance model performance by adding operational and environmental metadata, such as temperature, inflow rates, and nutrient levels. The parameters were incorporated into the time series, and their influence was systematically evaluated. The findings indicate that while metadata enhances predictive accuracy, the significance of individual features varies across WWTPs. Training a model on one dataset and retraining that model on site-specific data has been demonstrated to enhance performance, underscoring the necessity for adaptive, localized modeling strategies. The findings demonstrate the value of machine learning approaches for extracting predictive insights from complex, high-dimensional biological time series data in wastewater systems. The implementation of such a system as an additional wastewater monitoring technique has the potential to facilitate more timely public health decision-making in the future.

When ‘Scarce’ Becomes ‘Too Scarce’: The Rare Diseases Drama

Authors: Alice Bernasconi; Federico Pirola; Alessio Zanga; Antonio Balordi

Accepted As: Short/position/discussion paper
Student as main author: NO
Student(s) as author(s): YES
Summer School Track: NO

Abstract: "Rare diseases pose unique methodological challenges for causal questions, where data scarcity is not incidental but inherent. We present a case study on the external validation of a causal Bayesian network (BN) developed to investigate cardiovascular diseases (CVDs) in adolescent and young adult female breast cancer survivors. Despite strong internal performance, external validation across multiple Italian regions revealed scenarios where the model’s predictions collapsed into complete uncertainty. These included: (i) circumstances in which shifts in missingness and selection mechanisms made the original model parameters inappropriate (ii) unobserved infrequent treatments in training data; and (iii) novel treatment combinations emerging from evolving clinical protocols. We discuss the implications of these findings, emphasizing the need for including informative priors, making timely model updates, and explicit uncertainty quantification. Our experience underscores that in rare disease research, robust causal modeling demands critical scrutiny, adaptive learning, and the indispensable role of human revision. These lessons could be broadly applicable to other domains where data are scarce, emphasizing that methodological rigor and adaptability are as crucial as the models themselves."

Relevance-Aware Thresholding in Online Conformal Prediction for Time Series

Authors: Théo Dupuy; Binbin Xu; Stéphane Perrey; Jacky Montmain; Abdelhak Imoussaten

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: Uncertainty quantification has received considerable interest in recent works in Machine Learning. In particular, Conformal Prediction (CP) gains ground in this field. For the case of time series, Online Conformal Prediction (OCP) becomes an option to address the problem of data distribution shift over time. Indeed, the idea of OCP is to update a threshold of some quantity (whether the miscoverage level or the quantile) based on the distribution observation. To evaluate the performance of OCP methods, two key aspects are typically considered: the coverage validity and the prediction interval width minimization. Recently, new OCP methods have emerged, offering long-run coverage guarantees and producing more informative intervals. However, during the threshold update step, most of these methods focus solely on the validity of the prediction intervals – that is, whether the ground truth falls inside or outside the interval – without accounting for their relevance. In this paper, we aim to leverage this overlooked aspect. Specifically, we propose enhancing the threshold update step by replacing the binary evaluation (inside/outside) with a broader class of functions that quantify the relevance of the prediction interval using the ground truth. This approach helps prevent abrupt threshold changes, potentially resulting in narrower prediction intervals. Indeed, experimental results on real-world datasets suggest that these functions can produce tighter intervals compared to existing OCP methods while maintaining coverage validity.

Query Carefully: Detecting the Unanswerables in Text-to-SQL Tasks

Authors: Jasmin Saxer; Isabella Maria Aigner; Luise Linzmeier; Andreas Weiler; Kurt Stockinger

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: Text-to-SQL systems allow non-SQL experts to interact with relational databases using natural language. However, their tendency to generate executable SQL for ambiguous, out-of-scope, or unanswerable queries introduces a hidden risk, as outputs may be misinterpreted as correct. This risk is especially serious in biomedical contexts, where precision is critical. We therefore present Query Carefully, a pipeline that integrates LLM-based SQL generation with explicit detection and handling of unanswerable inputs. Building on the OncoMX component of ScienceBenchmark, we construct OncoMX-NAQ (No-Answer Questions), a set of 80 no-answer questions spanning 10 categories (non-SQL, out-of-schema/domain, and multiple ambiguity types). Our approach employs llama3.3:70b with schema-aware prompts, explicit No-Answer Rules (NAR), and few-shot exemplars drawn from both answerable and unanswerable questions. We evaluate SQL exact match, result accuracy, and unanswerable-detection accuracy. On the OncoMX dev split, few-shot prompting with answerable examples increases result accuracy, and adding unanswerable examples does not degrade performance. On OncoMX-NAQ, balanced prompting achieves the highest unanswerable-detection accuracy (0.8), with near-perfect results for structurally defined categories (non-SQL, missing columns, out-of-domain) but persistent challenges for missing-value queries (0.5) and column ambiguity (0.3). A lightweight user interface surfaces interim SQL, execution results, and abstentions, supporting transparent and reliable text-to-SQL in biomedical applications.

Patient’s Privacy vs. Public Health: Can AI Solve the Hippocratic Dilemma?

Authors: Antonio Balordi; Alice Bernasconi; Alessio Merlo

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: "To address the medical dilemma between data privacy (regulated in Europe by the GDPR) and public health, we propose a machine-unlearning procedure that uses Hessian-diagonal scores to identify parameters most influenced by the data to be forgotten, resets only those weights, and performs a brief, focused fine-tuning on retained data. Using a German cohort of breast cancer (BC) patients derived from a clinical trial, we analyze recurrence-free survival, categorized into three clinically meaningful classes, and targeted unlearning requests concentrated among Adolescents and Young Adults (AYA, < 40 years), a peculiar population most likely to withdraw. Across 10 tests, our approach was able to preserve high utility (Normalized Test Accuracy, NTA = 0.96) while achieving strong forgetting (NFS = 0.98) and near-chance vulnerability to membership inference (MIA ≈ 0.5), closely matching a retrained-from-scratch benchmark. Thus, our findings show that the proposed unlearning procedure effectively restores privacy with minimal loss of model utility. However, attention must be directed to specific challenges, such as the non-IID nature of requests, the localized impact of homogeneous forgetting on model calibration, and broader implications, including potential fairness drifts or adversarial misuse. These aspects underscore the need for further methodological improvements and subgroup-aware auditing, as well as cautious deployment in sensitive healthcare settings."

Designing a Multimodal Dataset for AI-Based Prognosis in STEMI

Authors: Simone Bartucci; Edoardo De Rose; Alessandro Quarta; Rossella Quarta; Alessia Donata Camarda; Alberto Polimeni; Francesco Calimeri

Accepted As: Short/position/discussion paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: YES

Abstract: "Despite advances in reperfusion therapy, ST-segment elevation myocardial infarction (STEMI) remains a leading cause of mortality in the general population. In this work, we present a multicenter multimodal dataset designed for AI-based 12-month prognostic stratification after STEMI. The cohort includes 822 patients with predischarge post-PCI (Percutaneous Coronary Intervention) 12-lead ECGs and comprehensive clinical, laboratory, echocardiographic, angiographic and pharmacological data, along with standardized acquisition and 12-month follow-up for major adverse cardiovascular events (MACE). Data were anonymized and harmonized, and a computer-vision pipeline has been designed to detect lead regions and extract analyzable signals. To our knowledge, unlike other existing public datasets, this is the first multicenter dataset coupling predischarge ECGs with rich multimodal context and longitudinal outcomes in STEMI, enabling robust AI models for personalized risk prediction. This approach helps to bridge the gap between acute diagnosis and long-term risk stratification, with significant potential to improve clinical decision-making and patient outcomes."

LLM-Powered Conformance Checking for Medical Processes

Authors: Stefania Montani; Giorgio Leonardi; Manuel Striani

Accepted As: Short/position/discussion paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "Clinical guidelines serve as normative process models for healthcare organizations but are often presented in unstructured, textual formats. This lack of formalization hinders the use of traditional conformance checking algorithms, which require structured, machine-readable process descriptions. In this study, we address this challenge by: (i) employing a Large Language Model (LLM) to extract normative rules from textual guidelines; (ii) assessing and quantifying the conformance of patient event logs to these rules; and (iii) using this framework to evaluate the quality of process models generated by various process discovery algorithms, based on their conformance to the extracted rules. In the paper, we present the results that we obtained applying our approach in stroke care. Notably, however, the ease of rule extraction with our LLM-powered method suggests its potential applicability across diverse domains."

Expert-Guided POMDP Learning for Data-Efficient Modeling in Healthcare

Authors: Marco Locatelli; Arjen Hommersom; Roberto Clemens Cerioli; Daniela Besozzi; Fabio Antonio Stella

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Learning the parameters of Partially Observable Markov Decision Processes (POMDPs) from limited data is a significant challenge. We introduce the Fuzzy-MAP EM algorithm, a novel approach that incorporates expert knowledge into the parameter estimation process by enriching the Expectation–Maximization (EM) framework with fuzzy pseudo-counts derived from an expert-defined fuzzy model. This integration naturally reformulates the problem as a Maximum A Posteriori (MAP) estimation, effectively guiding learning in environments with limited data. In synthetic medical simulations, our method consistently outperforms the standard EM algorithm under both low-data and high-noise conditions. Furthermore, a case study on Myasthenia Gravis illustrates the ability of the Fuzzy-MAP EM algorithm to recover a clinically coherent POMDP, demonstrating its potential as a practical tool for data-efficient modeling in healthcare.

Learning a Causal Model for Intracranial Pressure in Patients with Traumatic Brain Injury

Authors: Alessio Zanga; Francesca Graziano; Giuseppe Citerio; Paola Rebora; Stefania Galimberti; Shubhayu Bhattacharyay; David K. Menon; Ewout W. Steyerberg; Fabio Antonio Stella

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Traumatic brain injury is a sequence of pathophysiological events that originates from an acute biomechanical insult. One of the major challenges physician face routinely in traumatic brain injury patients is the management of intracranial pressure. Elevated intracranial pressure may lead to herniation, causing injury through compression of brain tissue. Studying the underlying mechanism of intracranial pressure is crucial to develop personalized therapy planning. In this paper, we build a causal model from clinical experts knowledge and partially-observed event-based data to represent the trajectory of patients over time. We show how to derive insights on the effectiveness of multiple treatments allocations from the model parameters and evaluate the model against treatment policies reported in clinical guidelines.

Comparing different positional encodings for the interpretation of medical images

Authors: Andrea Santomauro; Giorgio Leonardi; Luigi Portinale

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "Traditional positional encoding (PE) methods in Vision Transformers (ViT) focus primarily on spatial information, but they may not adequately capture the complex geometric patterns intrinsic to medical images. To address this limitation, we have previously proposed a similarity-based positional encoding combining convolution operations and standard cosine similarity between image patches. In the present work we compare similarity-based PE with two traditional alternatives in ViT such as standard learned PE and rotatory PE. The goal is to show that, in addition to provide better classification accuracy of 2D images in different medical domains, the attention maps generated by similarity-based PE appears to be more meaningful than those generated by alternative encodings, focusing on the medical relevant part of the images. Finally, we also show the benefits of the proposed approach in dealing with 3D medical images, again in terms of classification performance. We validate our method on a set of six medical imaging datasets from MedMNIST which are benchmark datasets of medical images of various kinds, such as X-rays, histological samples, dermoscopic, ultrasounds and microscope images."

Assessing Temporal Reasoning of Large Language Models on Structured Temporal Clinical Data

Authors: Gianluca Apriceno; Tania Bailoni; Mauro Dragoni

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "This study explores the capabilities of Large Language Models in reasoning over structured clinical knowledge with a focus on temporally dependent monitoring rules. We assess the models’ performance across five real-world patient scenarios derived from a virtual coaching platform in the Trentino Salute 4.0 project. Using a structured prompt and diverse use cases, we examine whether these models can accurately detect and justify rule violations and generate clear, audience-specific explanations for patients and healthcare professionals. The results reveal that while Large Language Models demonstrate potential in interpreting complex, time-sensitive clinical data and adapting communication to different stakeholders, limitations persist in temporal reasoning, explanation consistency, and audience alignment. These findings offer insights into the strengths and challenges of deploying LLMs in personalized and temporally aware healthcare decision support systems."

A Hybrid Evolutionary Fuzzy Neural Approach for Maternal Health Risk Prediction

Authors: Gianluca Apriceno; Marina Segala; Giovanni Valer; Nicola Muraro; Vincenzo Netti; Paulo Viktor Campos Sousa; Mauro Dragoni

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: The demand for transparent and explainable AI solutions in healthcare has highlighted the limitations of purely data-driven models and emphasized the value of symbolic methods such as Fuzzy Inference Systems, which provide understandable rules and facilitate collaboration with human experts. To overcome the limitations and merge the strenghts of individual approaches, hybrid models that integrate sub-symbolic and symbolic reasonig have emerged as a promising direction. For example, Fuzzy Neural Networks merge the predictive strength of neural networks with the clarity of fuzzy systems, and their performance can be further enhanced through evolutionary techniques that optimize parameters and improve adaptability. In this study, we propose an evolutionary Fuzzy Neural Network framework that use a genetic algorithm to strengthen classification capabilities while maintaining interpretability. By incorporating the evolutionary optimization into the network’s parameter update process, the model achieves both robustness and transparency. Validation on the Maternal Health Risk dataset demonstrates that the proposed approach effectively balances predictive accuracy with explainability.

LUME-DBN: Full Bayesian Learning of DBNs from Incomplete data in Intensive Care

Authors: Federico Pirola; Fabio Antonio Stella; Marco Grzegorczyk

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: YES

Abstract: Dynamic Bayesian networks (DBNs) are increasingly used in healthcare due to their ability to model complex temporal relationships in patient data while maintaining interpretability, an essential feature for clinical decision-making. However, existing approaches to handling missing data in longitudinal clinical datasets are largely derived from static Bayesian networks literature, failing to properly account for the temporal nature of the data. This gap limits the ability to quantify uncertainty over time, which is particularly critical in settings such as intensive care, where understanding the temporal dynamics is fundamental for model trustworthiness and applicability across diverse patient groups. Despite the potential of DBNs, a full Bayesian framework that integrates missing data handling remains underdeveloped. In this work, we propose a novel Gibbs sampling-based method for learning DBNs from incomplete data. Our method treats each missing value as an unknown parameter following a Gaussian distribution. At each iteration, the unobserved values are sampled from their full conditional distributions, allowing for principled imputation and uncertainty estimation. We evaluate our method on both simulated datasets and real-world intensive care data from critically ill patients. Compared to standard model-agnostic techniques such as MICE, our Bayesian approach demonstrates superior reconstruction accuracy and convergence properties. These results highlight the clinical relevance of incorporating full Bayesian inference in temporal models, providing more reliable imputations and offering deeper insight into model behavior. Our approach supports safer and more informed clinical decision-making, particularly in settings where missing data are frequent and potentially impactful.

Differentially Private Fine-Tuning of Self-Supervised Learning Models for Human Activity Recognition on Wearables

Authors: Oktay Ozan Güner; Sergio Consoli; Vicenç Gómez; Mario Ceresa

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: The surge in wearable technology for health has advanced personalized Human Activity Recognition (HAR), yet introduces privacy concerns under EU regulations such as the GDPR, the AI Act and the European Health Data Space (EHDS). This paper examines fine-tuning approaches for self-supervised models on wearable data, using Differentially Private Stochastic Gradient Descent (DP-SGD) to balance privacy and utility. Using the PAMAP2 dataset and the HarNet10 model, we compare classifier head and full model fine-tuning. Our results show that tuning only the classifier head (4.83% of parameters) preserves higher accuracy and F1-scores while significantly reducing vulnerability to membership inference attacks compared to non-private baselines. This strategy provides a lightweight and regulation-compliant framework for privacy-preserving HAR on wearables.

Explaining medical time series classification through boundary-aware feature analysis

Authors: Christel Sirocchi; Damiano Verda

Accepted As: Short/position/discussion paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Borderline clinical cases, characterised by highly similar patient profiles but divergent outcomes, pose a persistent challenge in medical decision-making. Laboratory data from electronic health records can help trace temporal trajectories of disease progression and uncover subtle differences between such cases. However, their sparse, irregular, and asynchronous nature complicates analysis and calls for tailored methods. Recent efforts in modelling such data have primarily relied on deep learning architectures, which typically prioritise overall performance and often lack interpretability, particularly in difficult boundary cases. This study introduces a methodology to identify features and criteria associated with divergent outcomes in borderline cases. Medical time series are summarised into interpretable statistics and pairwise distances are computed to identify similar admissions with either the same or different outcomes. Feature-wise differences within these pairs are then used to train monotonic gradient boosting models that highlight key discriminative factors. Applied to hospital mortality prediction in two ICU cohorts, the approach shows that features relevant in borderline cases differ from those emphasised by models trained on all admissions. Rule extraction from shallow monotonic trees yields interpretable patterns associated with outcome divergence. These findings suggest that the proposed framework can contribute to the refinement of clinical guidelines and strengthen decision support in uncertain scenarios. The method implementation is openly available at: https://github.com/ChristelSirocchi/TS-XAI.

A Hybrid Inductive-Deductive Approach for Resource-Efficient Suicide Risk Detection in Counseling Conversations

Authors: Ron Franco; Avi Segal; Kobi Gal

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: The growing reliance on online platforms for mental health support underscores the need for accurate, real-time suicide risk detection. Most existing datasets utilized in training AI models for suicide risk detection consist of conversation level training examples, missing critical turning points within interactions. Research is especially limited in low-resource languages like Hebrew, restricting broader applicability. We propose a hybrid inductive-deductive framework for crisis hotline chats: an inductive component uses large language models to generate message-level pseudo-labels, distilled into a compact classifier for resource-constrained settings; a deductive component links predictions to a clinically curated taxonomy of risk factors. Experiments on real-world data from a national support service show that pseudo-labeling with curriculum learning improves performance over strong baselines. Our results highlight the promise of combining inductive and deductive strategies for resource-efficient, clinically meaningful AI in suicide prevention.

Automatic Explainable Progress Prediction of Wet Age-Related Macular Degeneration

Authors: Riccardo Fidanza; Daniele Meli; Emilia Maggio

Accepted As: Short/position/discussion paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: "Wet age-related macular degeneration (wAMD) is an aggressive pathology representing a leading cause of central vision loss in the elderly population. Currently, diagnosis and treatment evaluation rely on clinical experts visually interpreting tomography scans to identify pathological features for decision-making, a process that is time-consuming and subject to inter-observer variability. Stemming from the success of artificial intelligence (AI) in medicine, this paper proposes two main contributions. First, we address the \textbf{novel task of automatic diagnosis and monitoring of the treatment outcome for wAMD}. We adopt a \textbf{unique and novel dataset} containing recordings from 275 patients over different years of treatment. We show that different \textbf{AI models significantly outperform the recall (measuring misclassified wAMD worsening) of human evaluation (+20\% at least)}. As a second contribution, we perform an \textbf{explainability study} on the trained AI models, evidencing that the \textbf{relevant features guiding the predictions are indeed a smaller subset and clinically relevant}. Our results pave the way towards \textbf{trustable automatic diagnosis and treatment evaluation for wAMD and related pathologies}, reducing significantly the effort required from clinicians."

Using Temporal Features to Improve Accuracy in Multivariate Pattern Analysis (MVPA) of M/EEG Data

Authors: Fatemeh Barazesh Morgani; Erin Goddard; Gustavo Enrique Batista

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: Multivariate Pattern Analysis (MVPA), or ‘brain decoding’ methods, have become a popular approach for analysing time series brain recordings such as Magnetoencephalography (MEG) and Electroencephalography (EEG) to address experimental questions in cognitive neuroscience. Currently, the most common approach is to epoch data relative to an event of interest (e.g., stimulus onset) and apply the classification analyses to relative time points individually. This approach reveals how neural information unfolds over time but neglects any information that may be present in the temporal dynamics of brain activity. In this work, we employ several time series methods (classifiers that incorporate temporal information) with varying time window durations to analyse MEG and EEG data and compare their utility for revealing the temporal dynamics of neural information. As the window duration increases, classifier accuracy improves, but the results become less precise regarding the timing of neural information. We found that a window of ~20ms provided an optimal balance for maximising classifier accuracy while maintaining the temporal precision of the results. We also tested time series classifiers based on convolutional kernels that often outperform simpler methods in benchmark datasets, but these did not yield better performance in our datasets. Our results suggest that using a brief window, rather than a single time point, could lead to improved sensitivity for MVPA methods to detect neural representations from MEG and/or EEG data.

Unified Peripartum Database with Natural-Language-to-SQL Capabilities at Udine University Hospital: Design and Prototype

Authors: Doriana Armenise; Ginevra Battello; Andrea Brunello; Lorenza Driul; Angelo Montanari; Elisa Rizzante; Nicola Saccomanno; Andrea Salvador; Serena Xodo; Silvia Zermano

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: "The fragmentation of obstetric information across electronic health record modules, device repositories, and laboratory systems, as it is common in hospitals, hinders both intrapartum care and reproducible research. In this work, we present a practical blueprint for transforming heterogeneous peripartum records into computable, queryable assets by designing and prototyping a unified peripartum relational database with natural-language-to-SQL (NL2SQL) capabilities at the Obstetrics Clinic of Udine University Hospital. Requirements were co-defined with clinicians and formalized as an Entity-Relationship diagram, from which the logical schema and SQL implementation of the database were then derived. The latter integrates heterogeneous sources to connect maternal anamnestic and longitudinal history, current-pregnancy findings, intrapartum course, and delivery and neonatal outcomes. The NL2SQL layer enables clinicians to pose natural-language queries to the system, lowering barriers to audit and exploratory analysis."

A Logic-Based Framework for Inferring Medical Events (Extended Abstract)

Authors: Yvon Kokou Awuklu; Vianney Jouhet; Katsumi Inoue; Fleur Mougin; Meghyn Bienvenu

Accepted As: Short/position/discussion paper
Student as main author: YES
Student(s) as author(s): NO
Summer School Track: NO

Abstract: We give a brief overview of our recent work on developing a novel logic-based framework for inferring high-level temporally extended events from timestamped clinical data and background knowledge. The framework specifies existence and termination conditions for simple events and derives meta-events by combining them. To address data imperfections, we introduce confidence annotations, consistency constraints, and a repair mechanism that selects preferred consistent event sets. While reasoning in the general setting is intractable, we identify useful fragments with polynomial-time data complexity. A prototype system, CASPER, implements the approach using answer set programming. Applied to two healthcare use cases, CASPER achieved feasible runtimes and produced clinically plausible inferences, demonstrating both computational feasibility and medical relevance. Although developed for healthcare applications, the framework is generic and can be reused in other domains.

Multi-agent Dynamic Resource Allocation: A Reinforcement Learning Approach

Authors: Stefania Costantini; Giovanni De Gasperis; Pasquale De Meo; Francesco Gullo; Alessandro Provetti

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "We investigate the problem of cooperative resource allocation in multi-agent systems, focusing on dynamic scenarios such as hospital networks. In our model, agents (e.g., hospitals) aim to redistribute limited resources, such as medical personnel, in a way that satisfies both local constraints and global equity objectives. We devise a reinforcement learning approach to a dynamic scenario with time-varying resource needs. We empirically evaluate the proposed approach through extensive experiments. Our results demonstrate the effectiveness of our approach."

An LBBD approach for Solving the Chemotherapy Treatment Scheduling Problem

Authors: Simone Caruso; Carmine Dodaro; Marco Maratea; Cinzia Marte; Marco Mochi

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "The Chemotherapy Treatments Scheduling problem in oncology clinics is a complex problem, as the solution has to satisfy multiple requirements. These include adhering to the cyclic nature of chemotherapy treatment plans, ensuring a steady flow of patients, and effectively managing limited resources such as treatment time, nursing staff, and drug availability. Simultaneously, achieving an effective schedule is crucial for ensuring optimal health outcomes. A previous approach addressed this problem through a direct encoding via Answer Set Programming. In this paper, we propose an alternative solution based on a Logic-Based Benders Decomposition approach, implemented using multi-shot solving, and test it on real data, where the results demonstrate advantages for the LBBD method."

Are we ready for multi-agent systems to deliver structured mental health support? Lessons from adapting Self-Help+ intervention

Authors: Leonardo Sanna; Marco Bolpagni; Valentina Fietta; Simone De Carli; Mattia Franzin; Giorgia Gavioli; Lorenzo Gios; Susanna Pardini; Anna Elena Nicoletti; Silvia Rizzi; Silvia Gabrielli; Mauro Dragoni

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "Recent progress in natural language processing has created new possibilities for delivering personalized digital mental health support. Yet, combining structured, evidence-based therapeutic frameworks with the flexibility of natural conversation remains a challenge. This study examines the use of Large Language Models (LLMs) in structured chatbots to improve delivery of the World Health Organization’s Self-Help+ program. We compared a conventional state-machine chatbot with an LLM-enhanced version and finally with a multi-agent architecture, examining the strengths and limitations of the different approaches. Through simulation testing and expert focus group analysis, we found that a multi-agent architecture, while significantly improving personalization, struggles in maintaining protocol fidelity and therapeutic structure. Our findings suggest that current LLM-based architectures, while promising, might not yet be ready for unsupervised deployment in mental health contexts."

Exploring Large Language Models as Digital Health Coaches: A Role-Play Approach to Brief Action Planning

Authors: Marco Bolpagni; Simone De Carli; Leonardo Sanna; Silvia Gabrielli; Mauro Dragoni

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: Brief Action Planning (BAP) is a structured method rooted in motivational interviewing, designed to support individuals in adopting healthier behaviors. This paper explores the feasibility of using Large Language Models (LLMs) as conversational agents to deliver BAP without specialized training data or fine-tuning. By employing role-play prompting, we guided an LLM to simulate a health coach that facilitates goal-setting and action planning for sedentary lifestyles. The approach was tested in simulated conversations and user evaluations, examining adherence to BAP protocols and user experience. Results demonstrate both promise and limitations: while the LLM could replicate several key components of BAP, challenges remain in consistency and personalization. These findings highlight the potential of LLM-driven role-play as a complementary tool for scalable, time-efficient health interventions.

Integrating Multi-Domain Data and Knowledge for Personalized Health: The Salute+ Use Case

Authors: Mauro Dragoni; Gianluca Apriceno; Tania Bailoni

Accepted As: Non-original communication
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: The growing demand for Artificial Intelligence based systems in personalized health highlights the need to integrate data from multiple domains. This includes user-specific information (e.g., meal descriptions and activity records) and contextual knowledge (e.g., food composition and caloric expenditure). Such integration enables a holistic understanding of an individual’s mental and physical health, allowing Artificial Intelligence systems to leverage multi-modal data for generating accurate, context-aware recommendations tailored to individual needs. Multi-Modal Knowledge Graphs provide a robust framework for organizing, representing, and reasoning over this diverse data, supporting the development of personalized health support systems. These systems empower users to maintain healthy lifestyles and improve daily living. In this paper, we present how structuring multi-modal data within a Multi-Modal Knowledge Graph enhances the recommendations of a digital health platform, enabling users to make informed decisions and achieve better health outcomes. We demonstrate the suitability of our solution through the integration of our strategy within the Salute+ case study.

Using multi-modal generative models for creating speech therapy material

Authors: Mihir Mulye; Stefan Conrad; Stefan Knecht

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: NO

Abstract: "The process of collecting rehabilitation material can be resource intensive, consume a lot of time and effort, and can also constrain the possibilities of customization. Improvements in multimodal generative Machine Learning can help accelerate and enhance the efficiency of this process. In this paper, we explore the use of such models to create rehabilitation material for aphasia. We propose a novel pipeline that uses existing rehabilitation concepts to generate not only therapy tasks using Large Language Models (LLMs), and images using Text-to-Image (T2I) models but also generate prompts for generating images and the therapy tasks themselves. We observe that combinations of Qwen (LLM) with both Stable Diffusion Turbo and Flux (both T2Is) tend to generate images with higher T2I evaluation metrics. Furthermore, our investigation of using the Large Language and Vision Assistant (LLaVa) in conjunction with our generated rehabilitation material shows that the LLaVa answers have 75.1% and 78.2% alignment with rehabilitation material generated using Stable Diffusion Turbo+Qwen and Flux+Qwen, respectively."

Triage Discrimination: Myth or Reality?

Authors: Justin Armanini; Fabio Antonio Stella

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: YES

Abstract: Discrimination, understood as the unjust or prejudicial treatment of individuals based on characteristics such as ethnicity, age, gender, or disability, is a deeply contested issue and a significant challenge for contemporary society. This paper examines potential instances of discrimination within the healthcare sector, with particular focus on the triage process. We employ causal networks and causal mediation analysis to answer two key questions: (1) Does triage decision-making exhibit disparity based on demographic factors such as race, gender, or age? (2) If such disparities are identified, to what extent can they be attributed to clinically relevant factors rather than to demographic characteristics? The results reveal demographic disparities in triage assignments for age, gender, and race. However, causal mediation analysis shows that gender and race disparities are entirely explained by clinical mediators. In contrast, age exhibits a direct effect on acuity scores beyond its influence on vital signs, though this may reflect clinically appropriate age-based triage protocols rather than unjustified discrimination. These findings highlight the importance of causal analysis in distinguishing between statistical disparities and actual discrimination in healthcare settings.

Survival Hidden Markov Model

Authors: Alessandro Bregoli; Francesco Bellocchio; Luca Neri; Fabio Antonio Stella

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "Survival analysis is a statistical approach used to predict the time until the occurrence of a specific event. It has broad applications across various domains, including healthcare, manufacturing, and logistics. In this paper, we advance standard statistical approaches by introducing the Survival Hidden Markov Model (SHMM), which decouples the modeling of the failure event from the representation of the hidden state. This design choice enhances the model’s interpretability and allows the latent states to capture underlying dynamics independently of the event occurrence. We evaluate SHMM by comparing its performance to state-of-the-art survival analysis methods, including the Cox proportional hazards model and Random Survival Forests, on synthetic data as well as on chronic kidney disease (CKD) single session data."

Automated Machine Learning to Enhance Knowledge Retrieval in Retrieval-Augmented Generation Pipelines

Authors: Matteo Magnini; Gianluca Aguzzi; Leonardo Sanna; Simone Magnolini; Patrizio Bellan; Mauro Dragoni; Sara Montagna

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: The emergence of Large Language Models has introduced exciting possibilities for applications in the digital health domain. However, their unpredictable nature necessitates the development of trustworthy strategies to prevent the generation of hallucinations. A common approach to address this challenge is using Retrieval-Augmented Generation (RAG), where text generation is supported by controlled knowledge injected into the prompts. Even with RAG, ensuring reliable and authoritative information generation requires further research. In a previous work, we presented an enhanced approach to the classic RAG pipeline, introducing an initial step where the Large Language Model generates an enhanced query to support the retrieval step. Results showed that performances are highly sensitive to the techniques adopted for embedding queries and retrieving documents. Accordingly, in this paper, we experiment with a novel automated machine-learning approach to conduct extensive testing across various configurations and explore the retrieval module. Our findings highlight that the embedder and, especially, the retrieval strategies strongly impact the overall performance of the RAG pipeline.

Machine Learning and Statistical Insights into Hospital Stay Durations: The Italian EHR Case

Authors: Marina Andric; Mauro Dragoni

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "Length of hospital stay is a critical metric for assessing healthcare quality and optimizing hospital resource management. This study aims to identify factors influencing LoS within the Italian healthcare context, using a dataset of hospitalization records from over 60 healthcare facilities in the Piedmont region, spanning from 2020 to 2023. We explored a variety of features, including patient characteristics, comorbidities, admission details, and hospital-specific factors. Significant correlations were found between LoS and features such as age group, comorbidity score, admission type, and the month of admission. % Machine learning models, specifically CatBoost and Random Forest, were used to predict LoS. The highest R$^2$ score, 0.49, was achieved with CatBoost, demonstrating good predictive performance."

Data-Driven Patient Care Level Classification Using Machine Learning

Authors: Paulo Viktor Campos Sousa; Mauro Dragoni

Accepted As: Full paper
Student as main author: NO
Student(s) as author(s): NO
Summer School Track: NO

Abstract: "Accurate patient care level classification is essential for optimizing hospital resource allocation and improving patient outcomes. This study presents a data-driven approach to classifying patient care needs using machine learning models, based on a dataset collected from an Italian hospital. The dataset originally contained missing values, which were addressed through data preprocessing techniques, including feature selection and normalization. Various machine learning models were trained and evaluated using performance metrics such as Area Under the Curve (AUC), classification accuracy (CA), F1-score, precision, recall, and Matthews correlation coefficient (MCC). Among the tested models, the AdaBoost classifier achieved the highest accuracy (0.914) and AUC = 0.971, outperforming other methods in predictive performance. Neural Networks also demonstrated competitive results (CA = 0.905, AUC = 0.975), followed by Random Forest (CA = 0.904, AUC = 0.976). These findings highlight the potential of ML-based decision-support systems for medical professionals in determining patient care levels."

Interpolation-Based Data Augmentation for Automatic Persuasive System

Authors: Mauro Dragoni; Renan Lirio de Souza

Accepted As: Full paper
Student as main author: YES
Student(s) as author(s): YES
Summer School Track: YES

Abstract: "Data scarcity presents a significant challenge in developing effective Automatic Persuasive Systems, particularly in domains where collecting high-quality conversational data is difficult due to privacy concerns or resource constraints. In this paper, we propose a novel approach to data augmentation for persuasive dialogues that uses the decision tree structure of argumentative conversations. By combining decision theory for dialogical argumentation with large language models, we generate contextually appropriate user arguments through an interpolation process between system arguments. Our method considers the tree structure of the dialogue, ensuring that the generated arguments maintain logical flow and coherence. We evaluated our approach in three domains - meat persuasion, bike commuting, and COVID-19 vaccination - comparing synthetic arguments with both original dataset arguments and human-created ones. The results demonstrate that our LLM-interpolated arguments achieve higher similarity to human-created arguments than to the original dataset arguments, suggesting that our approach generates more realistic dialogue content. Furthermore, the consistency of our results across different domains and models indicates the robustness of our strategy. This work contributes to addressing data scarcity in persuasive dialogue systems while maintaining coherence and quality of the arguments."

Page updated

Report abuse