• affective computing
• autonomic nervous system
• electrodermal activity
• emotion recognition
• heart rate variability
• machine learning
• physiological signal processing

One of the most common and noninvasive ways to infer human psycho-physiological states is investigating the Autonomic Nervous System (ANS) dynamics by monitoring physiological signals. Several previous works aimed at exploring the regulatory mechanisms of ANS, quantifying the activity of ANS by exploring its sub-branches and finally devising computational tools to infer the psycho-physiological states. Particularly, Heart Rate Variability (HRV) and Electrodermal Activity (EDA) signals have been studied thoroughly providing robust correlates of ANS dynamics. This thesis introduces innovative computational tools for ANS characterization for the final task of inferring human psycho-physiological states. One of the main challenges in ANS studies is the study of the multiple autonomic regulatory mechanisms involved in the autonomic response to concurrent stressors/emotional stimuli. In this context, I propose a novel model for disentangling the contribution of multidimensional autonomic control mechanisms on cardiovascular dynamics. The proposed model provides new perspectives
for ANS activity on cardiac control dynamics, likely highlighting new biomarkers in the psycho physiology and physio-pathology fields. Another important aspect is the quantification of ANS activity to study the behavior of the interactive mechanisms involved. Particularity, the characterization of the balance between the sympathetic and parasympathetic nervous systems (i.e., the sympatho-vagal balance) has some major limitations in the literature. State-of-the-art has demonstrated that the spectral analysis of HRV and EDA signals contains reliable biomarkers about parasympathetic and sympathetic activity, respectively. Here, using feature extraction methods, I propose a novel index correlated with sympatho-vagal activity through bi-variate spectral analysis of these signals. Moreover, based on the nonlinear time-varying nature of the ANS, I use the instantaneous bispectral estimates computed using a nonlinear point-process model for a nonlinear characterization of the sympatho-vagal activity. Finally, in the common processing chain for detecting psycho-physiological states, the information coming from ANS raw signals and the features extracted from them through signal analysis or physiological based models are often used as the input of a data-driven pattern
recognition system (machine learning or deep learning algorithms). In this dissertation, I propose a novel hybrid approach, that I refer to as the Physiologically-informed Gaussian Process (PhGP) model, that allows a principled, probabilistic combination of the information obtained from automatically-discovered patterns in the physiological data and the information coming from expert knowledge (e.g., physiological theories, physiologically inspired models describing HRV and EDA generation and extracted features correlated with ANS dynamics) to detect psycho-physiological states. Building
upon Bayesian statistics, the Gaussian Process (GP) models allow the injection of problem-specific domain knowledge in the form of an a-priori distribution on the GP latent function. I also propose qualitative and quantitative computational tools to get interpretability measures over the predictions made by the proposed PhGP model. I evaluate the discussed methodologies on different experimental protocols aiming at recognizing different physiological states from human subjects when these are altered by a stimulus or a group of affective/stressful stimuli. The outcomes of this thesis have
been published in the scientific literature.