An Information-Theoretic Framework for the Machine Learning Detection of Physiological States from Cardiovascular Signals

Cardiovascular variability analysis represents a powerful non-invasive tool for assessing autonomic regulation, while its interpretation remains challenged by the nonlinear, multi-scale, and highly interdependent nature of physiological signals. Traditional approaches often rely on univariate or pairwise descriptors and treat the stages of the machine-learning (ML) pipeline as loosely connected steps, limiting both interpretability and physiological insight.This thesis proposes an integrated analytical framework grounded in information theory (IT), where signal processing, feature extraction (FE), feature selection (FS), feature importance (FI), and classification are treated as interconnected stages of a unified pipeline for detecting, dissecting, and interpreting the information contained in cardiovascular variability time series. Within this framework, several methodological developments are introduced to organise and analyse this information in a systematic and interpretable way. Information-theoretic measures are employed at multiple levels: entropy, Mutual Information (MI), and Conditional Mutual Information (CMI) are used to characterise cardiovascular time series and physiological interactions; information-based selection criteria guide the identification of compact and non-redundant feature subsets; finally, high-order information decomposition frameworks are adopted to quantify unique, redundant, and synergistic contributions of features to predictive tasks.From a methodological standpoint, the thesis introduces and validates information-theoretic approaches for FS and FI analysis and classification, explicitly accounting for feature interactions beyond first-order effects and enabling the identification of compact and non-redundant feature subsets. In addition, high-order FI frameworks based on CMI enable the decomposition of feature relevance into cooperative, redundant, and synergistic contributions. These methods complement classical ML models by improving interpretability while maintaining competitive predictive performance.The proposed analytical framework is evaluated across multiple experimental and applied scenarios, including assessment of stress (postural, mental, driving, and work-related stress) and of sex-related differences in cardiovascular regulation. Results consistently show that postural stress elicits clearer autonomic signatures than cognitive stress, that ultra-short-term features can retain discriminative capability when carefully selected, and that sex-related differences emerge primarily through coordinated patterns of cardiac and vascular features rather than isolated markers. These findings highlight how different stressors engage distinct autonomic regulatory mechanisms and show that coordinated cardiovascular interactions provide more reliable indicators of physiological state than isolated variability indices. Applications in real-world and large-cohort settings further demonstrate the generalizability of the framework and highlight the critical role of preprocessing and inter-subject normalization.Overall, the results demonstrate that information theory provides a unifying and physiologically meaningful foundation for cardiovascular variability analysis and ML-based modelling. By integrating methodological rigor with physiological interpretability, the proposed framework may advance the understanding of autonomic regulation and support the development of transparent, interaction-aware models for biomedical and human-centered applications.