ABSTRACT. Early detection of cardiovascular diseases (CVDs) is crucial for minimizing their adverse impact on patients' health. Electrocardiograms (ECGs), which capture the heart's electrical activity, have been widely used to primarily evaluate heart conduction disorders. On the other hand, phonocardiograms (PCGs) recorded during cardiac auscultation, have been less explored, often being overlooked in favor of echocardiograms for detecting mechanical issues such as valvular diseases. However, due to their low cost and non-invasive nature, the analysis of both ECGs and PCGs can be easily integrated into preventive settings. Combining effectively the complementary information from these two modalities could significantly enhance the early detection of CVDs, where Machine Learning (ML) techniques can offer promising and cost-effective solutions. Progress in this area, however, has been limited by the lack of large enough datasets containing both ECG and PCG signals. One objective of this work is to analyze in-depth prior bimodal CVD detection research, identifying key issues to better address data collection and transfer learning limitations. We also propose a different approach to transfer learning for improving heart sound interpretation. Our findings confirm the effectiveness of using both signals to detect abnormal heart conditions. However, we also notice that even a refined transfer learning approach to enhance PCG interpretation is not enough to fully address the issues coming from the lack of bimodal data, indicating the need for further efforts in this direction. Ultimately, our bimodal approach achieved an overall AUROC of 96.4%, exceeding the performance of corresponding ECG-only and PCG-only models by approximately 3% and 10%, respectively. Compared to the other existing approaches, our method demonstrated superior AUROC performance while maintaining a relatively low false-negative rate, which is critical in CVD screening contexts.

ABSTRACT. In the last decades, the growing population of cancer survivors has shifted researchers’ focus from primary toward tertiary prevention. Particularly, adolescents and young adults (AYAs) breast cancer (BC) survivors may face long-term outcomes as a result of their treatments, among which cardiovascular diseases (CVDs) are the most life-threatening ones. To plan effective follow-up guidelines for preventing and treating these events, it is essential to disentangle the causal role of cancer treatments in these patients. In this work, we aim to extend the current state of BC treatment guidelines by leveraging on the estimate of the risk of CVDs in AYAs who underwent BC treatments, as provided by a causal Bayesian network. In these regards, we provide counterfactual explanations of a clinical query, using real-world data, algorithms and methods from the causal inference domain. We show that while ovarian suppression combined with tamoxifen may be a necessary cause for ischemic heart disease, it is not a sufficient one, i.e., this treatment alone is not enough to cause the disease, other factors must also be present. These findings can contribute to support clinicians in the treatment choice and help in refining treatment strategies and follow-up protocols for AYAs, advancing personalized healthcare in oncology.

ABSTRACT. Disease management, especially for chronic conditions or the elderly, involves continuous monitoring, lifestyle adjustments, and frequent healthcare interactions, necessitating effective home-care ICT solutions. To address these needs, chatbot technology has emerged as a promising tool for supporting patients in managing their health autonomously. In this context, chatbots must provide timely information and continuous support to maintain patient engagement. Moreover, key requirements include maintaining patient engagement through empathetic interactions and delivering accurate information without direct healthcare professional oversight. Additionally, data privacy concerns necessitate avoiding third-party Natural Language Processing and Generation services. To meet these needs, in this paper we propose the development of a chatbot to support patients in managing chronic conditions, focusing on hypertension. Particularly, we utilise open-source large language models to avoid proprietary systems due to privacy requirements. Given that their performance, based on state-of-the-art metrics, do not compete third-party services, we incorporate RAG techniques, building a knowledge base with input from medical professionals to enhance model performance. We evaluated seven open-source models, including two specifically trained in the medical domain. Our results indicate that RAG significantly improves performance, surpassing that of specialised medical-domain models without RAG. This approach offers a promising solution for managing chronic conditions independently and securely.

ABSTRACT. Causal networks go beyond the purely correlative approach that most machine learning models pursue. Indeed, they explicitly represent cause-effect relationships and explain how human actions can ramify towards different outcomes. For these reasons, causal networks are attracting increasing interest in the healthcare domain, where physicians commonly need a mechanistic portrait of the system under study and support for effective decision-making. Developing a causal network for the problem at hand is a complex task known as causal discovery, which combines prior knowledge and available data. Extensive literature covers theoretical aspects of causal discovery. However, the task is still challenging in settings characterized by low sample size and limited prior knowledge, a typical scenario when trying to disentangle rare diseases functioning over time. This paper tackles the challenge by developing a novel and original pre-processing algorithm for survival data, i.e., data measuring whether and when an event of interest occurred, and a highly structured workflow for learning causal networks related to different time windows. Comparing the structure of these causal networks enables domain experts to study the evolution over time of the causal mechanisms ruling the system. The proposed methodology is unique in interacting with experts and refines the generalizability and reproducibility of causal discovery studies in similar settings. Moreover, the case of soft tissue sarcoma, a class of rare cancers, is presented. The obtained results demonstrate the effectiveness of our approach in the rare disease domain and provide the first cause-effect representation of soft tissue sarcoma natural history.

ABSTRACT. South Africa faces a critical shortage of blood donors, leading to substantial deficits in the national blood supply. Blood donations are vital for treating life-threatening conditions, making it crucial to develop efficient models for managing blood stocks. This paper presents a mathematical model to optimize blood donation and ensure sufficient supply to meet fluctuating demands. The model captures the complex interactions within the blood banking system, focusing on minimizing costs, reducing waste, and efficiently distributing blood units. Specifically, it addresses daily supply challenges by minimizing the need for emergency imports and reducing blood wastage due to expiration while meeting all demand requirements. The core objective is to minimize blood wastage and reduce the reliance on imported blood banks during emergencies. The proposed objective function incorporates variables such as emergency importation and expiration rates, and robust optimization techniques are applied to identify optimal solutions while satisfying operational constraints. Symbiotic Organism Search (SOS), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO) methods are utilized for optimization. Among these, SOS demonstrated superior performance, achieving the lowest levels of importation and wastage. However, the algorithms could not significantly reduce supply levels due to the accumulation of excess stock from the previous day, which carried over into the next day. This paper provides valuable information on blood supply management and highlights the potential for optimization techniques to improve efficiency and sustainability in blood banking.

ABSTRACT. The objective is to construct a prognostic index that incorporates radiomic information with the validated prognostic index (Sarculator) provided by the Fondazione IRCCS Istituto Nazionale dei Tumori di Milano. A Bayesian approach was employed, utilising a Weibull model. Vague prior distributions were elicited for the shape parameter, the intercept, and the Sarculator. A multivariate Gaussian prior was elicited for the 2,144 radiomic parameters, incorporating a penalty factor, λ. A total of 100 penalty values were considered. A new, ad hoc adaptive version of the pre-conditioned Metropolis adjusted Langevin algorithm (A-MALA) was proposed for sampling. Bayesian Model Averaging (BMA) was employed to yield a composite of the 100 models. A Bayesian hypothesis test was constructed to evaluate the superiority of the BMA prognostic index relative to the Sarculator. The five-year AUC posterior mean was 0.809, with a 95% credible interval (CI) of (0.768, 0.851). The posterior mean of the C-index was 0.804 (95% CI, 0.764, 0.845) for the BMA, 0.743 (95% CI, 0.713, 0.771) for the best model log λ = 10.39 and 0.735 (95% CI, 0.674, 0.761) for the Sarculator. The results suggest that radiomic variables should be included in the model.

ABSTRACT. In this paper, I propose a novel approach to improve the safety of classification models by incorporating adversarial training into the model training process. I demonstrate that adversarial training enhances the safety of models without significantly affecting their accuracy. my methodology involves generating adversarial examples from the original dataset and creating a mixed training dataset that includes both original and adversarial examples. By evaluating the model’s performance on both types of datasets, I can achieve a balance between accuracy and safety. The proposed approach offers a more resource- efficient alternative to traditional model tuning processes, which often require multiple iterations to meet performance standards. My results indicate that including at least 20% adversarial examples in the training data effectively improves safety while allowing adjustments based on the relative importance of accuracy and safety. The new method simplifies the process of achieving safety requirements and reduces the resources needed for training and evaluation. Future research will focus on validating this approach with diverse safety-related datasets, various model structures, and different adversarial attack methods, as well as exploring its applicability to other non- functional requirements in AI.

ABSTRACT. The treatment of patients suffering from chronic diseases is a difficult problem to be tackled. Its complexity mainly originates from the following sources: the patient-specific response to the prescribed therapy, the impact of the interplay between disease and therapy on the quality of life of the patient and relatives, and the economic costs incurred by the healthcare system. Recently, there has been considerable interest in developing, studying, and applying artificial intelligence methods to diagnosis, prognosis and treatment personalization. This paper combines two techniques from artificial intelligence, namely fuzzy logic and reinforcement learning, to develop optimal dynamic treatment for patients suffering from a chronic disease. In this paper, we focus on cancer as a chronic disease and leverage a biologically validated fuzzy logic model from the literature. Different problem settings, of increasing complexity, are taken into account, presented and analyzed. Results of an extensive numerical experimental plan confirm the potential of non-myopic decision-making when treating chronic disease patients.

ABSTRACT. COVID-19 vaccination was adopted worldwide due to the advent of the COVID-19 pandemic in 2019. However, many post-vaccination adverse events, such as death and severe illness, were reported. So far, the specific case of post-vaccination adverse events pertaining to persons with cardiovascular risk factors and comorbidities has not been explored empirically, which limits the understanding of the underlying causes of adverse reactions to vaccination by this category of persons. This paper explored Explainable AI (XAI) to identify the critical determinants of post-vaccination mortality in persons with cardiovascular risk factors. To do this, we extracted 16657 records of persons with cardiovascular risk factors from the VAERS open dataset (from 2020 to May 2024). We then employed predictive modelling using a process that involved four stages. The first stage involved extracting relevant data from VAERS, data preprocessing, and handling class imbalance. In the second stage, we conducted a comparative performance evaluation of seven machine learning (ML) algorithms (Logistic Regression — LR, K-Nearest Neighbour — KNN, Deep Multilayer Perceptron — Deep MLP, Support Vector Machines — SVM, Random Forest — RF, Extreme Gradient Boosting — XGBoost, and Categorical Boost — CatBoost). In the third stage, we compared the performance of two stacked ensemble models composed of six base models, using Catboost and XGBoost as the meta-learners in each case. The fourth stage involved using SHAPley Additive Explanations (SHAP) to interpret the predictions of the best-performing model. The result showed that CatBoost has the best performance among the base ML models (Acc = 0.96, F1=0.96, AUC = 0.96), while Stacked ensemble - XGBoost had the best overall performance (Acc = 0.96, F1=0.96, AUC = 0.99). Also, we found the important predictors of post-vaccination mortality in persons with cardiovascular comorbidity. Generally, older age, a higher number of days spent in the hospital increases the risk of mortality, while the absence of current illness, life-threatening condition, hospitalization, prolonged hospitalization, disability, birth defect, doctor visit, and emergency care; and vaccination dose completion will enhance the probability of survival. However, the presence of diabetes, high cholesterol, high blood pressure, and other illnesses increases the risk of mortality. This study's findings contribute to a better understanding of critical factors that could enable better handling of adverse events related to post-vaccination in persons with cardiovascular disease comorbidity.

ABSTRACT. Stress is a prevalent and growing phenomenon in the modern world that could lead to significant physical issues, both physical and mental health. Analyzing physiological signals collected from wearable sensors using artificial intelligence methods has emerged as a promising approach to predicting and managing stress. However, conventional models for time series analysis are RNN architectures and encounter challenges like high computational costs and issues with vanishing or exploding gradients. Inspired by the success of deep learning methods in computer vision, several studies have proposed transforming time series into images by applying encoding time series algorithms. This work intends to compare three time-series encoding methods: Gramian Angular Field (GAF), both summation and difference, Markovian Transition Field (MTF) and Recurrent Plot (RP) in the stress detection scenario. We employ two architectures, VGG-16 and ResNet, based on Convolutional Neural Network (CNN), to evaluate the performance of these methods on a public dataset, WESAD. Our results demonstrate that the GAF encoding method proves to be the most effective for classifying physiological signals related to stress.

ABSTRACT. Functional Status Information describes physical and mental wellness at the whole-person level. Collecting and analyzing this information is critical to address the needs for caring for an aging global population, and to provide effective care for individuals with chronic conditions, multi-morbidity, and disability. Knowledge Graphs represent a suitable way for meaning in a complete and structured way all information related to people’s Functional Status Information and reasoning over them to build tailored coaching solutions supporting them in daily life for conducting a healthy 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, and the evaluation and insights observed.

ABSTRACT. The brain network damage provoked by a neurological disease can be modeled as the result of the action of an operator, $K$, acting on the brain, inspired by physics. Here, we explore the matrix formulation of $K$, analyzing eigenvalues and eigenvectors, with heuristic considerations on different techniques to approximate it. The primary objective of this paper is to lay the foundational groundwork for an innovative framework aimed at the development of predictive models regarding the progression of neurodegenerative diseases. This endeavor will leverage the potential of integrating these novel representations of brain damage with advanced machine-learning techniques. A case study based on real-world data is here presented to support the proposed modeling.

ABSTRACT. Accurate identification of patient emergency states in emergency rooms is vital for delivering timely and appropriate medical intervention. This paper presents a comprehensive approach using a dataset from a Northern Italian hospital to predict patient urgency levels with machine learning algorithms. We processed and analyzed the data through feature selection techniques, feature importance analysis, and interpretability methods, focusing on the dataset dimensions. Our tests resulted in accuracy exceeding 95% in three machine learning algorithms, demonstrating the feasibility of developing an intelligent computerized system capable of predicting emergency states in an emergency room setting. These findings suggest that integrating advanced data analytics can significantly enhance patient triage and hospital resource planning.

ABSTRACT. Medical imaging is a cornerstone of modern healthcare, facilitating early diagnosis and the development of efficient treatment plans. Breast imaging includes different imaging modalities, including mammography and MRI, each encompassing unique information. Unfortunately, improving diagnostic performance can be accompanied by an increase in patient-related risks. Specifically, Contrast-enhanced mammography (CEM) offers better performance while exposing women to the risk of adverse reactions from the contrast agents used for it. To reduce these risks, deep learning solutions have become one of the promising research frontiers in recent years. In image-to-image translation, a mapping function is learned to transform a given image from a source domain to a target domain. In medical imaging, the most common solutions are based on GANs, such as pix2pix. When applied to CEM, we found that pix2pix encounters specific challenges due to low data quality, insufficient model capacity, and domain-derived requirements. Thus, these models have low performance out-of-the-box. In this paper, we highlight these specific challenges, propose tailored evaluation strategies, and present preliminary results on a novel dataset, showcasing the need for specialized approaches in medical imaging translation.

ABSTRACT. Anxiety and depression are the most common mental health issues worldwide, affecting a non-negligible part of the population. Accordingly, stakeholders, including governments' health systems, are developing new strategies to promote early detection and prevention from a holistic perspective (i.e., addressing several disorders simultaneously). In this work, an entirely novel system for the multi-label classification of anxiety and depression is proposed. The input data consists of dialogues from user interactions with an assistant chatbot. Another relevant contribution lies in using Large Language Models (LLMs) for feature extraction, provided the complexity and variability of language. The combination of LLMs, given their high capability for language understanding, and Machine Learning (ML) models, provided their contextual knowledge about the classification problem thanks to the labeled data, constitute a promising approach towards mental health assessment. To promote the solution's trustworthiness, reliability, and accountability, explainability descriptions of the model's decision are provided in a graphical dashboard. Experimental results on a real dataset attain 90 % accuracy, improving those in the prior literature. The ultimate objective is to contribute in an accessible and scalable way before formal treatment occurs in the healthcare systems.

ABSTRACT. Risk communication represents a nuanced discourse within the healthcare sector, characterized by the topics' sensitivity and the potential for misunderstandings between healthcare providers and patients. This delicacy stems from the complexity of effectively conveying information about risks. Consequently, a primary obstacle lies in fostering healthcare providers' understanding of implicit communication nuances inherent in pre-operative risk discussions. This study aims to address this gap in the literature by examining the topic through the lens of the philosophy of language, specifically utilizing pragmatic analysis tools to elucidate implicit understandings in doctor-patient interactions. We employ this approach to scrutinize instances of risk evaluation preceding cardiac surgery. Through empirical analysis of gathered data, we illustrate the inadequacies of current state-of-the-art models in accurately identifying misunderstandings within healthcare dialogues. In conclusion, we propose avenues for future research in this domain, underscoring the importance of further exploration into improving risk communication in healthcare settings.