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.

Automated persuasion systems (APS) are conversational agents that exchange arguments and counterarguments with users during dialogues to persuade them to believe in something. Such systems use strategies (or policies) to carefully select a sequence of arguments that are tailored to the user’s needs and will likely have a positive outcome, that is, changing the user’s belief in a certain argument. Biparty Decision Theory (BDT) is a framework that uses game theory to formalize a dialogue between an APS and a user, that is, an exchange of (counter) arguments during each turn of the APS or the user. During the APS turn, the BDT policy selects the best argument to maximize only the utility for the APS and neglects the utility of that argument for the user. This is a reasonable choice in games, but in a persuasive dialogue, it can result in arguments that have a high utility for the APS but a modest utility for the user. There the user may be less likely to be persuaded. This is crucial in settings where there are no arguments with good utilities for both the APS and the user and a compromise has to be found. To this extent, we define a new family of policies for BDT, called aggregated policies, that consider, during the decisions of the APS, an aggregation of the APS and user’s utilities. Such an aggregation considers both the APS and the user’s needs leading toward a sequence of arguments representing the best trade-off of utilities. We evaluate the approach using both a new synthetic dataset and a published dataset of utilities for dialogical argumentation. The results show the aggregated policies find better compromise arguments w.r.t. the classical policy of BDT.

In the domain of healthcare and well-being, the fusion of machine learning and data collection through medical examinations has propelled significant advancements in diabetes monitoring. Diabetes, a prevalent and intricate health condition, has garnered increasing attention due to its substantial impact on individuals’ mental and physical well-being. The use of medical examinations for real-time diabetes assessment has become pivotal, with various physiological monitoring capabilities aiding in this endeavor. Machine learning, as a subset of artificial intelligence, has further elevated the precision and effectiveness of diabetes monitoring by extracting meaningful insights from the extensive and intricate data collected through these examinations. This paper introduces a novel and interpretable computational model known as the Evolving Fuzzy Neural Network Uni-Nullneuron-Based Approach (EFNN-UniNull). Comprising three interconnected layers, this model collaboratively produces classification outcomes while concurrently providing insightful interpretations of relationships within the age group identification of diabetes patients dataset. The fuzzification method based on grid partition helps in obtaining adequate knowledge about the problem. The model underwent a comparative analysis against evolving neuro-fuzzy systems, demonstrating results approaching 85% accuracy. Notably, the model extracted knowledge based on fuzzy rules pertinent to diabetes identification.

Functional Status Information refers to a person’s overall mental and physical health. Collecting and analyzing Function Status Information data is crucial for addressing the needs of a growing elderly population, as well as for providing effective care to those with chronic diseases, multiple health issues, or disabilities. Knowledge Graphs provide an effective method for organizing and representing Functional Status Information data in a structured way. Furthermore, they can also allow reasoning over this data to create personalized health support solutions that assist people in maintaining a healthy lifestyle and improving daily living. In this paper, we describe the integration of our Functional Status Knowledge Graph, namely FuS-KG , into a real-world application run within a large-scale living lab involving more than 4,000 people. We provide the road map of this experience including the challenges, the platform’s architecture, the focus on the knowledge layer, the evaluation and the insights observed.

Automated persuasion systems (APS) aim to persuade a user to believe something by entering into a dialogue in which arguments and counterarguments are exchanged. To maximize the probability that an APS is successful in persuading a user, it can identify a global policy that will allow it to select the best arguments it presents at each stage of the dialogue whatever arguments the user presents. However, in real applications, such as for healthcare, it is unlikely the utility of the outcome of the dialogue will be the same, or the exact opposite, for the APS and user. In order to deal with this situation, games in extended form have been harnessed for argumentation in Bi-party Decision Theory. This opens new problems that we address in this paper: (1) How can we use Machine Learning (ML) methods to predict utility functions for different subpopulations of users? and (2) How can we identify for a new user the best utility function from amongst those that we have learned. To this extent, we develop two ML methods, EAI and EDS, that leverage information coming from the users to predict their utilities. EAI is restricted to a fixed amount of information, whereas EDS can choose the information that best detects the subpopulations of a user. We evaluate EAI and EDS in a simulation setting and in a realistic case study concerning healthy eating habits. Results are promising in both cases, but EDS is more effective at predicting useful utility functions.

Dialogue systems for the health domain are expected to be reliable and to reason in accordance to medical experts' reasoning. Given the complexities of the health domain, these systems are frequently aided by semantic-aware approaches implementing technologies such as ontologies. However, the automated generation of such systems is still a challenging task. In this work, we propose an approach that integrates automated planning and information management with the aim of automating the generation of efficient dialogue managers. Resulting dialogue managers are capable of handling multi-turn goal-oriented dialogue sessions within the healthcare domain. By evaluating a prototype on the asthma domain, our results reveal the suitability of our approach to generate dialogue policies on real-time scenarios.

In the last years, the state of the art of NLP research has made a huge step forward. Since the release of ELMo (Peters et al., 2018), a new race for the leading scoreboards of all the main linguistic tasks has begun. Several models have been published achieving promising results in all the major NLP applications, from question answering to text classification, passing through named entity recognition. These great research discoveries coincide with an increasing trend for voice-based technologies in the customer care market. One of the next biggest challenges in this scenario will be the handling of multi-turn conversations, a type of conversations that differs from single-turn by the presence of multiple related interactions. The proposed work is an attempt to exploit one of these new milestones to handle multi-turn conversations. MTSI-BERT is a BERT-based model achieving promising results in intent classification, knowledge base action prediction and end of dialogue session detection, to determine the right moment to fulfill the user request. The study about the realization of PuffBot, an intelligent chatbot to support and monitor people suffering from asthma, shows how this type of technique could be an important piece in the development of future chatbots.

The use of knowledge resources in the digital health domain is a trending activity significantly grown in the last decade. In this paper, we present HeLiS : an ontology aiming to provide in tandem a representation of both the food and physical activity domains and the definition of concepts enabling the monitoring of users’ actions and of their unhealthy behaviors. We describe the construction process, the plan for its maintenance, and how this ontology has been used into a real-world system with a focus on “Key to Health”: a project for promoting healthy lifestyles on workplaces.

Multi-domain sentiment analysis consists in estimating the polarity of a given text by exploiting domain-specific information. One of the main issues common to the approaches discussed in the literature is their poor capabilities of being applied on domains which are different from those used for building the opinion model. In this paper, we will present an approach exploiting the linguistic overlap between domains to build sentiment models supporting polarity inference for documents belonging to every domain. Word embeddings together with a deep learning architecture have been implemented into the NeuroSent tool for enabling the building of multi-domain sentiment model. The proposed technique is validated by following the Dranziera protocol in order to ease the repeatability of the experiments and the comparison of the results. The outcomes demonstrate the effectiveness of the proposed approach and also set a plausible starting point for future work.

This work presents SimBa-2, an improved version of a novel crossover specifically adapted to the evolutionary optimization of neural network designs that aims at overcoming one of the major problems of recombination, known as the permutation problem. The crossover is based on a so-called `local similarity' between two individuals selected for the recombination process from the population, and it is applied according to a similarity threshold. An approach exploiting this operator has been implemented and applied to five benchmark classification problems in machine learning, chosen among some of the well known classification problems provided by the UCI Machine Learning Repository. The application of different similarity threshold values has been investigated and the experimental results show how the behavior of the operator changes with respect to this parameter.

We propose a supervised approach to word sense disambiguation based on neural networks combined with evolutionary algorithms. Large tagged datasets for every sense of a polysemous word are considered, and used to evolve an optimized neural network that correctly disambiguates the sense of the given word considering the context in which it occurs. The viability of the approach has been demonstrated through experiments carried out on a representative set of polysemous words.

The task of reasoning with fuzzy description logics with fuzzy quantification is approached by means of an evolutionary algorithm. An essential ingredient of the proposed method is a heuristic, implemented as an intelligent mutation operator, which observes the evolutionary process and uses the information gathered to guess at the mutations most likely to bring about an improvement of the solutions. The viability of the method is demonstrated by applying it to reasoning on a resource sheduling problem.