This study examines how Large Language Models advise users during healthcare conflicts where power differences matter (e.g., patient vs. clinician, junior vs. senior staff). We introduce a Thomas–Kilmann–grounded benchmark of 150 short healthcare conflict vignettes constructed from author-defined role, hierarchy, and option-style rules, with LLM-assisted drafting followed by validation and revision. Each vignette is paired with five response options representing common conflict styles: accommodate, assert, compromise, collaborate, and avoid. We evaluate multiple leading LLM families in both single-turn (stateless) and history-conditioned settings, and compare their choices with a small human-judgment study. Across models, recommendations concentrate heavily on collaboration, with assertion a distant second and very little accommodation or avoidance. Models show hierarchy-sensitive shifts in assertiveness, while remaining strongly collaboration-dominant overall. Human choices are more varied, suggesting LLMs may over-default to one “best” conflict approach. These findings motivate conflict-aware evaluation and calibration to better support autonomy in healthcare advice.

The emergence of Large Language Models (LLMs) 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 LLM 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.

The effectiveness of Digital Therapeutics (DTx) relies on understanding and shaping human behavior, yet development is hindered by fragmented conversational scripts and repetitive chatbot implementation. Integrating advanced agentic components such as Large Language Models (LLMs) and multi-agent systems further complicates clinical safety, data provenance, and ethical risks. We present Flow Builder, a no-code platform within the MAESTRO (Multi-Agent Enhanced System for Therapy Routine Orchestration) framework, designed to standardize conversation design. Leveraging a modular block-based hierarchy with over seventeen specialized processors, from text delivery to semantic retrieval and multi-step LLM reasoning, Flow Builder enables clinicians to focus on therapeutic logic while ensuring safe, consistent, and scalable agentic execution.

Large Language Models are increasingly used in clinical tasks, yet their ability to reason over temporally dependent rules and generate audience-tailored explanations remains underexplored. We evaluate seven recent LLMs on five real-world patient scenarios, comparing structured (Turtle) and unstructured (natural language) representations of clinical knowledge. Performance is assessed in terms of guideline violation detection accuracy and explanation quality, based on expert evaluation of correctness, clarity, temporal traceability, and reasoning alignment. Results show that LLMs perform more reliably with unstructured prompts, while structured formats often lead to misinterpretations, particularly in temporally complex scenarios. Although LLMs generate coherent explanations, their consistency and audience alignment vary, underscoring the need for expert validation. These findings highlight key limitations and opportunities for temporally aware clinical decision support. GitHub: https://github.com/IDA-FBK/LLM-Benchmarking-Temporal-Clinical-Reasoning .

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. GitHub: https://github.com/IDA-FBK/LLM-Temporal-Clinical-Rea soning

The integration of Large Language Models (LLMs) into conversational systems has greatly enhanced their expressive capabilities but also introduced challenges in structure, control, and reproducibility. Current design approaches often fall between rigid, rule-based systems that are predictable but limited in scalability, and fully LLM-driven models that enable natural interaction but lack transparent logic. We present MAESTRO (Multi-Agent Enhanced System for Therapy Resources Orchestration), a framework for the trustworthy integration of LLMs in complex settings such as psychological digital interventions. Central to MAESTRO is a flow-based abstraction model that represents conversational logic as assemblies of modular, reusable blocks. This approach allows designers to construct explicit interaction flows in which LLMs can serve as bounded-reasoning components, preserving predictability while enabling controlled flexibility.

Length of hospital stay (LoS) is a key metric for healthcare quality and hospital resource management. This study investigates factors influencing LoS within the Italian healthcare system, using patient-level hospitalization records from standardized hospital discharge forms (Schede di Dimissione Ospedaliera, SDO) for the population served by a local health authority in Piedmont. The dataset included 37,526 patients across 66 facilities over four years. We analyzed patient characteristics, comorbidities, admission details, and hospital-specific factors. Significant associations were observed with age group, comorbidity burden, admission type, and month of admission. Machine learning models, CatBoost and Random Forest, were used to predict LoS, with CatBoost achieving the highest validation R ^2 of 0.49. Historical LoS and procedure were the most influential predictors. As diagnosis and procedure information is recorded at discharge, the models are intended for retrospective analysis rather than real-time prediction at admission. The results demonstrate the potential of administrative SDO data for understanding LoS patterns and supporting hospital planning.

Artificial Intelligence is increasingly used in healthcare for clinical decision-making and personalized recommendations. However, most systems rely on an implicitly anthropocentric ethical framework that prioritizes short-term, patient-centered outcomes while overlooking long-term, systemic, and multi-agent consequences of clinical decisions. This paper introduces a healthcare-oriented instantiation of NAEL (Non-Anthropocentric Ethical Logic), a computational framework that models ethics as an adaptive inferential process rather than static rules or external constraints. Grounded in active inference, NAEL frames ethical decision-making as the minimization of global expected free energy, extending agent-centric formulations to account for uncertainty induced across other agents and the environment. Ethical considerations are embedded directly within the agent’s generative model and policy evaluation. Through a conceptual ICU resource allocation case study, we show how NAEL-driven policies differ from conventional short-term optimization approaches, highlighting their potential for clinically informed and ethically coherent decision-making in complex healthcare settings.

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 reasoning 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.

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.

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.

Hospital readmissions are a major burden on healthcare systems and result from complex interactions among heterogeneous clinical factors. While machine learning can model such interactions, interpretability is essential for clinical use. We propose a logic-guided learning framework for readmission prediction based on Logic Tensor Networks (LTN), integrating first-order logical rules into supervised learning using Italian administrative healthcare data. The rules encode clinically motivated relationships among chronic conditions, prior utilization, and patient stability. Training combines classification loss with rule-satisfaction constraints, enabling interpretable predicate representations. Results across cross-validation runs show competitive performance compared to baseline models, while the learned predicates provide clinically meaningful and interpretable structure for readmission risk.

Large Language Models (LLMs) have been widely investigated for conversational support in mental health. However, their therapeutic reliability in empathy and perspective-taking remains uncertain. We aim to fill this gap by introducing a comprehensive evaluation of the Pool of Experts framework, a multi-agent approach that enables role-specific identities without retraining the underlying model. Beyond its effectiveness, this framework provides a controlled test-bed to study whether personality framing induces measurable behavioral variation across roles and tasks. We systematically assess the Pool of Experts capability in empathy and theory-of-mind-oriented benchmarks through question-answering tasks, evaluating accuracy via strict and relaxed match against gold answers, and compare structured multi-agent orchestration with less structured conditions. Results demonstrate three main findings. First, architectural orchestration with deliberative aggregation consistently improves performance: a Final Decision Maker agent improves accuracy by up to 4.1

This paper presents DigiTher, a flexible AI-based platform for clinical research on digital therapeutics, enabling personalized treatment plans and real-time monitoring of patient progress. The DigiTher platform has been designed to deliver three specific kinds of Digital Therapeutics, which are mainly based on conducting motivational dialogues with the patients or providing multimedia material to accompany and support patients. Conversational solutions and sensing technology are used to (i) support the management of chronic conditions (e.g., nutritional ones), (ii) run Cognitive Behavioral Therapy (e.g., within mental health management scenarios), and (iii) support Behavioral Enhancement Interventions (e.g., improve lifestyle).

eHealth Literacy (eHL) is a set of competencies and skills encompassing the knowledge, comfort and perceived ability to identify, evaluate and apply electronic health information to health problems. Given its role in the appropriate use of health technologies, ensuring equitable access to health information and improving patient outcomes, this study aims to systematically retrieve, qualitatively and quantitative pool and critically appraise available experimental evidence on the effectiveness of eHL interventions across different population groups. Following the PRISMA guidelines, we conducted a systematic review in PubMed/Medline, Scopus, Web of Science, Embase, Cochrane Library and ClinicalTrials.gov, including original experimental studies quantifying the effectiveness of interventions aimed at increasing eHL, as assessed by the eHealth Literacy Scale (eHEALS) or other validated scales. We performed a random-effects model meta-analysis comparing changes in eHL levels before and after the interventions, and between the intervention and control groups. Heterogeneity was assessed using I2 statistics. Out of the 504 studies retrieved, 15 studies conducted between 2011 and 2023 met the inclusion criteria. Target populations of eHL interventions included adults in 7 studies, older people in 5 and young people in 4. The meta-analysis included 10 studies that used the eHEALS. Participants showed a mean increase in eHEALS scores of 5.81 points (95

Large Language Models generate answers to any questions provided by users. Even though this is a positive characteristic, when they are integrated into a real-world domain-specific solution, the generation of an answer to a question that is not related to the domain is a weakness. This work addresses the challenge of classifying users' questions as in-topic or out-of-topic to limit the capabilities of a Large Language Model. We propose a multi-agent approach, called "Pool of Experts", which leverages a structured hierarchy of specialized agents to synthesize expert contributions into a final decision. To evaluate the effectiveness of the proposed approach, we tested our methodology by integrating two description-based frameworks for agent profile creation: User Design Persona and Cognitive Load Theory. We compared our approach against traditional Transformer-based Natural Language Inference models as a baseline. Experimental results, observed in a real-world scenario concerning a question-answering system supporting pregnant women, demonstrate the superiority of the proposed methodology.

The past years have been characterized by an upsurge in opaque automatic decision support systems, such as Deep Neural Networks (DNNs). Although DNNs have great generalization and prediction abilities, it is difficult to obtain detailed explanations for their behavior. As opaque Machine Learning models are increasingly being employed to make important predictions in critical domains, there is a danger of creating and using decisions that are not justifiable or legitimate. Therefore, there is a general agreement on the importance of endowing DNNs with explainability. EXplainable Artificial Intelligence (XAI) techniques can serve to verify and certify model outputs and enhance them with desirable notions such as trustworthiness, accountability, transparency, and fairness. This guide is intended to be the go-to handbook for anyone with a computer science background aiming to obtain an intuitive insight from Machine Learning models accompanied by explanations out-of-the-box. The article aims to rectify the lack of a practical XAI guide by applying XAI techniques, in particular, day-to-day models, datasets and use-cases. In each chapter, the reader will find a description of the proposed method as well as one or several examples of use with Python notebooks. These can be easily modified to be applied to specific applications. We also explain what the prerequisites are for using each technique, what the user will learn about them, and which tasks they are aimed at.

This study investigates Large Language Models' capabilities in reasoning with structured clinical knowledge to identify and justify violations of health guidelines and to generate clear, audience-specific explanations for patients and physicians. Using a structured prompt and a use case involving dietary monitoring for a cardiovascular patient, we evaluate the performance of different Large Language Models. Preliminary findings highlight the potential of Large Language Models to effectively reason with structured health data and deliver tailored explanations, offering valuable insights into their role in personalized healthcare and decision support systems. GitHub: https://github.com/IDA-FBK/LLMReasoningPersonalHealthData

Digital health interventions often require structured, protocol-driven dialogues delivered with high fidelity. Evaluating whether an agent employing a Large Language Model (LLM) can meet these requirements remains challenging, especially in early development stages. In this work, we present VALISE (Virtual Agent Laboratory for Instruction-Following Simulation and Evaluation), a modular framework for simulating and evaluating LLM agent behavior in delivering structured health interventions. VALISE enables configurable agent-patient simulations using synthetic personas and evaluates protocol adherence through a customizable, automated grid assessed by ensembles of LLM-based judges. We demonstrate its use with Brief Action Planning (BAP), a short intervention promoting behavior change in sedentary individuals. Our results strongly align LLM-based and expert annotations, supporting VALISE's effectiveness for early-stage evaluations. VALISE offers a reproducible, extensible platform for testing instruction-following capabilities of LLM agents in digital health.