ABSTRACT. The convergence of Artificial Intelligence and healthcare is not only a trend; it represents a transformative shift that holds the potential to redefine how we understand and improve patient outcomes. As machine learning evolves, leveraging vast datasets to uncover patterns and predictive insights, we find ourselves at a pivotal moment where the challenges of trustworthiness and explainability demand our attention.

In this talk, I will explore the Informed Machine Learning paradigm, specifically discussing the framework of symbolic knowledge integration—where domain expertise is integrated into machine learning frameworks—and symbolic knowledge extraction, which empowers us with interpretable insights from complex models. The goal is to create predictive models that integrate empirical data with established medical guidelines. By fostering this synergy, we can enhance diagnostic accuracy and reduce the prevalence of false negatives, ultimately leading to improved patient care. Accordingly, these approaches are not merely technical enhancements; they signify a paradigm shift toward creating AI systems that are both intelligent and transparent.

ABSTRACT. The medical field is constantly evolving, integrating the latest technologies to enhance patient care and treatment efficacy. While various methodologies are available to evaluate the quality of research studies, checklists are often favored for their efficiency and ease of use. In this study, we contribute to this area of research by 1) analyzing the components of the most widely used checklists, and 2) proposing a more comprehensive checklist, CLARITY AI, which synthesizes the strengths of existing tools. This study analyzed several established checklists—CLAIM, CONSORT, DECIDE, FUTURE, IJMEDI, PRISMA, SPIRIT, STARD, STARE-HI, and TRIPOD—with the goal of developing a comprehensive checklist for evaluating research studies. Each item in these checklists was carefully cataloged, labeled, and assessed. The analysis aimed to identify the most critical items for inclusion in a definitive checklist for research study evaluation. The final version of the checklist is a coherent integration of structural elements—such as Title, Abstract, and Introduction—and essential parameters like Study Identification and Data Handling. This synthesis results in a comprehensive tool for thorough study and research evaluation. By integrating the strengths of multiple established checklists, CLARITY offers a robust, systematic, and user-friendly framework for assessing research quality. This tool not only elevates research standards but also enhances transparency, reproducibility, and overall credibility in the field of medical AI studies. Its application has the potential to produce more reliable and effective healthcare solutions, ultimately improving patient outcomes and advancing medical research.

ABSTRACT. Zero-shot anomaly detection (ZSAD) offers potential for identifying anomalies in medical imaging without task-specific training. In this paper, we evaluate CLIP-based models, originally developed for industrial tasks, on brain tumor detection using the BraTS-MET dataset. Our analysis examines their ability to detect medical-specific anomalies with no or minimal supervision, addressing the challenges posed by limited data annotation. While these models show promise in transferring general knowledge to medical tasks, their performance falls short of the precision required for clinical use. Our findings highlight the need for further adaptation before CLIP-based models can be reliably applied to medical anomaly detection.

ABSTRACT. Accurate cardiovascular (CV) risk assessment is relevant for asymptomatic individuals, in particular for those at risk for cardiovascular diseases (CVD). Metabolic-associated fatty liver disease (MASLD), previuosly known as non-alcoholic fatty liver disease (NAFLD), is recognized as a critical independent risk factor for increased CV morbidity and mortality. With a prevalence of 25% in the general population, MASLD is the leading cause of chronic liver diseases and is strongly associated with the development of coronary artery disease (CAD). Coronary CT is widely used to detect CAD, and it can also assess liver steatosis, providing valuable prognostic information for at-risk patients. In this paper, we present a study on methods for performing prognostic stratification of CAD risk in asymptomatic MASLD patients using Machine Learning (ML) approaches. In particular, we conducted a retrospective analysis of clinical data from 60 patients who underwent Coronary CT at L'Aquila Hospital (Italy) between 2017 and 2021. Dataset includes significant features, such as radiodensity (Hounsfield Unit), calcium score (Agatston score), and liver fibrosis (Fib-4 score and APRI). We compared several ML algorithms (Logistic Regression, Support Vector Classifier (SVC), Random Forest, Extreme Gradient Boosting, K-Nearest Neighbors (KNN), Naive Bayes), with the main goal of performing binary classification tasks and creating a model able to differentiate between healthy patients and those affected by both MASLD and CVD. SVCs emerged as the best-performing models, achieving an AUC of 94%, an accuracy of 95%, and a recall of 94%. Our approach offers a robust and accurate tool for predicting CAD risk in MASLD patients, providing a valuable contribution to clinical practice for early CV risk stratification and management.

ABSTRACT. The Nuclear Medicine Scheduling problem consists of assigning patients to a day, in which the patient will undergo the medical check, the preparation, and the actual image detection process. The schedule of the patients should consider their different requirements and the available resources, e.g., varying time required for different diseases and radiopharmaceuticals used, number of injection chairs, and tomographs available. Recently, this problem has been solved using a logic-based approach utilizing the Answer Set Programming (ASP) methodology. However, it may be the case that a computed schedule can not be implemented due to a sudden emergency and/or unavailability of resources, thus rescheduling is needed. In this paper we present an ASP-based approach to solve such situation, that we call Nuclear Medicine Rescheduling problem. Experiments employing real data from a medium size hospital in Italy show that our rescheduling solution provides satisfying results even when the concurrent number of emergencies and unavailability is significant.

ABSTRACT. This preliminary study explores the use of diffusion models for brain imaging generation to address the limitations of small datasets in rare neurodegenerative conditions. Our goal is to improve model robustness by generating realistic variations in medical images. Data scarcity is the main issue for the application of deep learning techniques in neurodegeneration. In the last decade, diffusion models have tried to address this problem as a novel generative technique widely applied for image and video generation. A diffusion model, known for capturing complex data distributions, was trained on a multi-center dataset of Structural Magnetic Resonance Images of healthy subjects to generate a high-quality synthetic dataset. Our results show that the Maximum Mean Discrepancy between two distributions is 0.036, thus indicating that the two distributions are quite similar. However, other metrics such as the Frechet Inception Distance and the Multiscale Structural Similarity Index Measure achieve suboptimal results. Although far from model optimization, these preliminary results demonstrate that diffusion models can be a valid tool to generate high-quality brain imaging data.

ABSTRACT. Despite their good characteristics, tools based on Bayesian networks are not yet widespread used in medical decision support. Meeting the needs of the intended users of these tools is crucial for their acceptance and clinical involvement in the development of these tools is thus required. As one of such needs, a measure of query response uncertainty was put forward during the development of a Bayesian network-based tool for the prediction of lymph node metastases in patients with endometrial cancer.In this paper we sustain this need by exploring options for the presentation of query response uncertainty. We consider level of detail, one-sided versus two-sided intervals and two different 'look-ahead' options. The different options are illustrated by means of a small test network.

ABSTRACT. Deep learning methods have become a powerful tool in medical imaging, with great potential to improve diagnostic accuracy and support early disease detection. This is especially crucial for breast cancer, one of the most common cancers among women, where early detection of abnormal tissue is key to improving survival rates. AI-based methods show great promise in detecting this pathology. In this study, we explore the application of deep learning techniques to classify breast masses as malignant or benign using ultrasound images, aiming to support breast cancer diagnosis. We propose a workflow that integrates two neural networks: a U-Net for image segmentation and a SegNet for classification. An ablation study was conducted to determine the optimal configuration of parameters. Our approach was tested on 780 ultrasound images. The results show promising improvements in diagnostic accuracy, demonstrating the potential of AI-powered tools to significantly enhance the early detection of breast cancer.

ABSTRACT. Sepsis is a life-threatening condition with complex and dynamic progression, often requiring timely and personalized treatment strategies. In this paper, we propose a multivariate longitudinal clustering, an advanced data analysis technique, as a powerful approach to understanding the diverse trajectories of sepsis by grouping patients based on multiple clinical variables measured over time. Dynamic Time Warping (DTW) is integrated into the longitudinal clustering as a distance measure to identify subgroups of patients with similar temporal patterns in multivariate data. By leveraging sepsis-related electronic health records (EHRs), which provide rich time-series data on laboratory results along with patient demographics and underlying health conditions, the proposed method reveals distinct sepsis phenotypes that reflect variations in disease progression. Our results confirm the critical role of the Thrombin-Antigen complex and the International Normalized Ratio as predictors of poor outcomes for septic patients. Despite challenges like missing data and interpretability, multivariate longitudinal clustering in sepsis offers significant potential to enhance clinical decision-making and improve patient outcomes.

ABSTRACT. This paper presents a work in progress to enhance a Retrieval-Augmented Generation (RAG) pipeline for a medical chatbot designed to address evaluative questions related to patient concerns about "normalcy". The chatbot uses a novel approach called Hypothetical Document Embeddings (HyDoc) to augment queries and improve the retrieval of certified medical information. In the first evaluation of the chatbot, it emerged that evaluative queries often fail to retrieve relevant documents as well as to produce appropriately framed responses. We, therefore, experiment with the impact of an additional naive-RAG module to improve the retrieval and a Chain-of-Thought (CoT) inspired prompting strategy to contextualize the queries better and advance response generation. Results demonstrate that this method enhances document retrieval and the framing of generated replies, improving the chatbot's ability to generate responses that consider emotional and communicative aspects.

ABSTRACT. In many fields, such as functional genomics or finance, data analysis, and predictive modeling are always challenging for the course of dimensionality and noisy data. In these cases, effective feature selection algorithms, based on Machine and Deep Learning, can perform and improve the identification of important features, leading to more treatable problems in terms of dimensionality. The paper proposes a novel algorithm to perform Feature Selection on highly dimensional data, which exploits the reconstruction capabilities of autoencoders and an ad-hoc defined Explainable Artificial Intelligence-based score to select the most informative feature for predictions. We benchmark such an approach on several state-of-the-art datasets and against the previously proposed algorithm in the literature, showcasing its effectiveness.

ABSTRACT. This paper introduces a novel similarity-based positional encoding method aimed at improving the classification of medical images using Vision Transformers (ViTs). Traditional positional encoding methods focus primarily on spatial information, but they may not adequately capture the complex geometric patterns characteristic of medical images. To address this, we propose a method that utilizes convolution operations to extract geometric features, followed by a similarity matrix based on cosine similarity between image patches. This encoding is then incorporated into the ViT model, enabling it to learn more meaningful relationships beyond basic spatial positioning. The effectiveness of this method is demonstrated through experiments on six medical imaging datasets from MedMNIST, where our approach consistently outperforms the conventional learned positional encoding. This is particularly true in datasets with prominent geometric structures like PneumoniaMNIST and BloodMNIST. The results indicate that similarity-based encoding can significantly enhance medical image classification accuracy.

ABSTRACT. This article explores the potential of ChatGPT as a support tool for individuals with disabilities, focusing on how it facilitates communication, learning, and social integration. Practical examples are presented where ChatGPT supports individuals with Specific Learning Disorders (SLD), autism, and dyscalculia, adapting responses to their needs. The versatility of ChatGPT as an accessible assistant is highlighted, while also raising reflections on the balance between automation and human assistance.

ABSTRACT. This paper presents a work in progress on the application of Blueprint Personas as a foundational tool for advancing the digital transformation of health and care services in an aging society. The paper present how artificial intelligence (AI) and intelligent agents can support patients, caregivers, and healthcare professionals through customized, patient-centered care. By synthesizing detailed patient profiles, including medical, social, and personal factors, the aim is to enhance the interaction between healthcare technologies and users. Additionally,the work introduces the use of ontologies to structure knowledge in e-health systems, with a particular emphasis on integrating a Reference Ontology of Trust to ensure the reliability and transparency of AI-driven care solutions.This ongoing research aims to contribute to a more empathetic and effective digital health ecosystem.hor mayuse in the preparation of the documentation of their work.

ABSTRACT. Establishing causal dependencies is crucial in applied domains, such as medicine and healthcare, where decision-making must be explainable. In these settings, small sample sizes and missing data call for federated approaches to maximise the amount of information we can use. We propose a novel federated causal discovery algorithm capable of pooling information from multiple sources with heterogeneous missing data to learn a graph representing cause-effect relationships. In particular, we learn a causal graph on a centralised server while taking into account both prior knowledge and missingness mechanism specific to each client. We applied the proposed algorithm to a real-world, multicentric study on endometrial cancer and validated the resulting causal graph through quantitative analyses and a clinical literature review. Our approach learns an accurate model despite the presence of data missing not-at-random.

ABSTRACT. As deep learning models and datasets expand, the demand for computational resources and memory storage intensifies; at the same time, data privacy concerns hinder data and model sharing. Hence, accessibility of model training is significantly challenged. Vector embeddings, as compact representations of medical images, offer a solution to the challenges of computational resource demands and data privacy concerns in AI-based medical imaging. In this study we investigate the suitability of vector embeddings as substitutes for original medical images in disease prediction tasks, focusing on performance and fairness. Using datasets like MIMIC-CXR and CheXpert, we find that vector embedding-based models generally improve disease detection performance and mitigate unfairness in diagnosis rates. The reduced demographic signals in these embeddings may contribute to fairer outcomes without compromising performance. Our findings suggest that vector embeddings can enable more accessible and equitable medical computer vision, particularly in low-resource settings.

ABSTRACT. The application of Artificial Intelligence (AI) techniques for analyzing medical images and omics data is revolutionizing the healthcare industry by offering profound insights into various diseases. Achieving precise diagnoses and formulating effective treatment plans, however, demands intricate and multimodal analysis of complex, sensitive, and diverse medical datasets. Recent advancements in Machine Learning and Deep Learning have proven to be formidable in identifying and classifying specific diseases. This paper outlines the current projects undertaken by our research group in this innovative domain.

ABSTRACT. Introduction: Multi-view data offer advantages over single-view data for characterizing individuals, which is crucial in precision medicine toward personalized prevention, diagnosis, or treatment follow-up. Methods: Here, we develop a network-guided multi-view clustering framework named netMUG to identify actionable subgroups of individuals. This pipeline first adopts sparse multiple canonical correlation analysis to select multi-view features possibly informed by extraneous data, which are then used to construct individual-specific networks (ISNs). Finally, the individual subtypes are automatically derived by hierarchical clustering on these network representations. Results: We applied netMUG to a dataset containing genomic data and facial images to obtain BMI-informed multi-view strata and showed how it could be used for a refined obesity characterization. Benchmark analysis of netMUG on synthetic data with known strata of individuals indicated its superior performance compared with both baseline and benchmark methods for multi-view clustering. The clustering derived from netMUG achieved an adjusted Rand index of 1 with respect to the synthesized true labels. In addition, the real-data analysis revealed subgroups strongly linked to BMI and genetic and facial determinants of these subgroups. Discussion: netMUG provides a powerful strategy, exploiting individual-specific networks to identify meaningful and actionable strata. Moreover, the implementation is easy to generalize to accommodate heterogeneous data sources or highlight data structures.

ABSTRACT. Longitudinal analysis of multivariate individual-specific microbiome profiles over time or across conditions remains dauntin. Most statistical tools and methods that are available to study microbiomes are based on cross-sectional data. Over the past few years, several attempts have been made to model the dynamics of bacterial species over time or across conditions. However, the field needs novel views on handling microbial interactions in temporal analyses. This study proposes a novel data analysis framework, MNDA, that combines representation learning and individual-specific microbial co-occurrence networks to uncover taxon neighborhood dynamics. As a use case, we consider a cohort of newborns with microbiomes available at 6 and 9 months after birth, and extraneous data available on the mode of delivery and diet changes between the considered time points. Our results show that prediction models for these extraneous outcomes based on an MNDA measure of local neighborhood dynamics for each taxon outperform traditional prediction models solely based on individual-specific microbial abundances. Furthermore, our results show that unsupervised similarity analysis of newborns in the study, again using the notion of a taxon’s dynamic neighborhood derived from time-matched individual-specific microbial networks, can reveal dierent subpopulations of individuals, compared to standard microbiome-based clustering, with potential relevance to clinical practice. This study highlights the complementarity of microbial interactions and abundances in downstream analyses and opens new avenues to personalized prediction or stratified medicine with temporal microbiome data.

ABSTRACT. Gestational diabetes mellitus (GDM) is characterized by remarkable heterogeneity, making personalized treatment challenging. We applied unsupervised machine learning techniques to identify clinically relevant GDM clusters using data available in the routine clinical practice from 2,682 women diagnosed with GDM (1,865 women from Charité University Hospital in Berlin and 817 women from the Medical University of Vienna). We employed different clustering algorithms including k-means, k-medoids, and hierarchical clustering. Model selection was based on evaluating solution stability with the Jaccard index, cluster compactness via the silhouette score, and robustness assessed through two-fold cross-validation. External validation on independent cohorts was implemented to evaluate the generalizability of the identified clusters. Our final model identified three distinct clusters using maternal age, pre-pregnancy body mass index, and glucose levels from the diagnostic oral glucose tolerance test. Each cluster was associated with different treatment needs and neonatal outcomes. These findings provide a valuable clinical decision support that could help clinicians define more personalized treatment approaches, improving both maternal and neonatal outcomes.

ABSTRACT. This work proposes a hybrid approach to Data Augmentation that blends inductive and deductive reasoning. In particular, the approach effectively utilizes a modest collection of labeled images while employing logic programs to declaratively define the structure of new images, allowing for flexible and dynamic image generation; the use of logic programming ensures adherence to both domain-specific constraints and given desiderata. The resulting structures are then used for guiding the generation of new realistic images based on a dedicated Deep-Learning process. The general approach can be particularly of use in biomedical and healthcare scenarios, where building extensive datasets of quality images is in general a hard prerequisite for many applications that is challenging to meet. The approach is specialized to two real-world case studies featuring laryngeal endoscopic and cataract images, respectively, and experiments conducted for assessing the method are discussed.

ABSTRACT. Hyperthermia (HT) in combination with radio- and/or chemotherapy has become an accepted cancer treatment for distinct solid tumour entities. In HT, tumour tissue is exogenously heated to temperatures between 39 and 43 ℃ for 60 minutes. Temperature monitoring can be performed non-invasively using dynamic magnetic resonance imaging (MRI). However, the slow nature of MRI leads to motion artefacts in the images due to the movements of patients during image acquisition. By discarding parts of the data, the speed of the acquisition can be increased - known as undersampling. However, due to the invalidation of the Nyquist criterion, the acquired images might be blurry and can also produce aliasing artefacts. The aim of this work was, therefore, to reconstruct highly undersampled MR thermometry acquisitions with better resolution and with fewer artefacts compared to conventional methods. The use of deep learning in the medical field has emerged in recent times, and various studies have shown that deep learning has the potential to solve inverse problems such as MR image reconstruction. However, most of the published work only focuses on the magnitude images, while the phase images are ignored, which are fundamental requirements for MR thermometry. This work, for the first time, presents deep learning-based solutions for reconstructing undersampled MR thermometry data. Two different deep learning models have been employed here, the Fourier Primal-Dual network and the Fourier Primal-Dual UNet, to reconstruct highly undersampled complex images of MR thermometry. MR images of 44 patients with different sarcoma types who received HT treatment in combination with radiotherapy and/or chemotherapy were used in this study. The method reduced the temperature difference between the undersampled MRIs and the fully sampled MRIs from 1.3 ℃ to 0.6 ℃ in full volume and 0.49 ℃ to 0.06 ℃ in the tumour region for an acceleration factor of 10.

ABSTRACT. Identification of vessel structures of different sizes in biomedical images is crucial in the diagnosis of many neurodegenerative diseases. However, the sparsity of good-quality annotations of such images makes the task of vessel segmentation challenging. Deep learning offers an efficient way to segment vessels of different sizes by learning their high-level feature representations and the spatial continuity of such features across dimensions. Semi-supervised patch-based approaches have been effective in identifying small vessels of one to two voxels in diameter. This study focuses on improving the segmentation quality by considering the spatial correlation of the features using the Maximum Intensity Projection~(MIP) as an additional loss criterion. Two methods are proposed with the incorporation of MIPs of label segmentation on the single~(z-axis) and multiple perceivable axes of the 3D volume. The proposed MIP-based methods produce segmentations with improved vessel continuity, which is evident in visual examinations of ROIs. In this study, a UNet MSS with ReLU activation replaced by LeakyReLU is trained on the Study Forrest dataset. Patch-based training is improved by introducing an additional loss term, MIP loss, to penalise the predicted discontinuity of vessels. A training set of 14 volumes is selected from the StudyForrest dataset comprising of 18 7-Tesla 3D Time-of-Flight~(ToF) Magnetic Resonance Angiography (MRA) images. The generalisation performance of the method is evaluated using the other unseen volumes in the dataset. It is observed that the proposed method with multi-axes MIP loss produces better quality segmentations with a median Dice of $80.245 \pm 0.129$. Also, the method with single-axis MIP loss produces segmentations with a median Dice of $79.749 \pm 0.109$. Furthermore, a visual comparison of the ROIs in the predicted segmentation reveals a significant improvement in the continuity of the vessels when MIP loss is incorporated into training.