ADVANCES IN NONLINEAR HEART RATE VARIABILITY ANALYSIS FOR MOOD STATE ASSESSMENT AND EMOTION RECOGNITION

In this dissertation some advanced methodologies for the analysis of Heart Rate Variability (HRV) are proposed. This study is motivated by the necessity of improving the investigation of the autonomic control influence in the physiological aspects of mood assessment and affective computing. Several previous works pointed out the crucial role of the autonomic signals, in the assessment of health status. However many issues have arisen, due to the extraction of information from HRV in order to discern psychological states and emotions. At the same time, many evidences in the literature showed that physiological processes involve nonlinear frequency modulation or multi-feedback interactions associated to long-range correlations. For this reason many nonlinear parameters, extracted from autonomic signals, have already been used as markers of aging and presence of diseases. Throughout this thesis, existing nonlinear methodologies of analysis are improved and adapted to the study of long-term HRV recorded from bipolar patients, and novel approaches are proposed for the nonlinear analysis of the prompt response to emotion elicitation. In the first Chapter the literature about the control of Sympathetic and Parasympathetic Nervous Systems on the cardiovascular regulation is described, pointing out the difference between classical and nonlinear models. The second Chapter proposes a detailed review of standard and nonlinear methodologies for the analysis of HRV, describing time and frequency domains analysis, higher order spectra, time-frequency approaches and nonlinear methods derived from the phase-space theory. All the most powerful methods for the assessment of entropy in physiological signals are critically analyzed according to the changes between different algorithms. Statistical analysis and pattern recognition methods, used in to investigate the experimental data, are also reported in Chapter 2. The large majority of nonlinear techniques has to be applied to long-term recordings, in order to produce performing indexes. To overcome this limitation, which is especially relevant in the study of the immediate response to emotional stimuli, I discuss, in the Chapter 3, two main methods derived from the surface section of Poincaré, the Symbolic Analysis and the Lagged Poincaré Plot (LPP), proposing their use in the analysis of ultra-short time series (less than 5 minutes), through the study of novel parameters. The subjects recruitment and the experimental protocols are studied ad-hoc for each investigation and are described in Chapter 4. Three main objectives, according to the participants, can be identified: the discrimination of mood states in bipolar subjects, the characterization of psychological dimensions in non-pathological subjects and the emotion recognition in healthy subjects during protocols of affective computing. Wearable devices for the autonomic signals acquisition are used for the long-monitoring of patients affected by bipolar disease, in order to study the evolution of heartbeat dynamics while the subjects are involved in their every-day activities. In the emotion elicitation protocols, a multimodal stimulation is explored, through the research of standardized database of acoustic, tactile, olfactory and visual stimuli. In all the studies the Russel’s Circumplex Model of Affect (CMA) is used, detecting two dimensions in each emotions: the arousal (the level of intensity of the evoked emotion) and the valence (the pleasantness/unpleasantness). For visual and acoustic stimulation protocols, two existing databases of standardized stimuli, the International Affective Pictures System (IAPS) and International Affective Digitized Sound (IADS), are considered. The ratings of each sound and picture in this case is predetermined. In the other two cases, tactile and olfactory stimulation, the subjects are asked to assess their sensations through two scores of valence and arousal. In the affective touch protocol, caress-like stimuli, produced through an ad-hoc haptic device are elicited. The results of nonlinear methodologies application are presented in Chapter 5. A preliminary study about the reliability of LPP measures in ultra-short windows is described, through the analysis on synthetic series. The outcomes regarding bipolar patients are obtained through adapted versions of Multiscale Entropy (MSE) and Detrended Fluctuation Analysis (DFA). Ad-hoc algorithms of pattern recognition and prediction of the future mood state are developed using Markov model. A bivariate study of HRV and respiration signal of healthy subjects in resting state, through Multivariate Multiscale Entropy (MMSE), is applied to define the correlation between the complexity of autonomic dynamics and the psychological dimensions assessed through traditional questionnaires. Then the results of the validation of Symbolic Analysis and the LPP approaches to affective computing protocols are shown, using segments of signals of duration from thirty-five seconds to one minute and half.