Application of nonlinear indices and metrics for emotions recognition by means of a dynamical study of electroencephalographic signals

Resumen

Affective computing has become one of the most relevant research fields in recent times. Its main purpose is to endow machines with emotional intelligence, with the aim of humanizing the communication and interaction processes between people and machines. To achieve this, it is necessary to develop mathematical models for the automatic recognition of emotions based on the analysis of the information that can be extracted from a subject under different emotional conditions. Concretely, in the last years the interest in the analysis of physiological variables for emotions identification has notably grown, with special emphasis on the assessment of the brain activity through electroencephalographic (EEG) recordings. Brain signals represent the first reaction of the body against any emotional stimulus, hence providing more information than other physiological variables resulting from the spread of the brain signal to the rest of systems by means of the central nervous system. Traditional methods of study of brain signals are focused on the computation of linear parameters in time domain (such as mean, standard deviation, and other statistics) and frequency domain (like spectral power, or asymmetry among brain hemispheres in the different frequency bands). Nevertheless, it is known that the brain follows a nonlinear and nonstationary behavior, thus nonlinear methodologies of analysis could report more information than linear techniques, providing a more complete description of the brain dynamics. This doctoral thesis is based on the recognition of emotions by means of the study and application of nonlinear metrics for the analysis of EEG recordings. In particular, the works carried out during this predoctoral stage are focused on the computation of entropy-based indices for the estimation of regularity, predictability and functional connectivity of the brain under different emotional conditions. These metrics have been mainly applied for the detection of negative stress, given the strong influence of this emotional state on the current society and its negative consequences for the health of people who suffer from it. The results obtained demonstrate that these entropy-based indices present a good performance when identifying negative stress, calm and other emotional states from brain signals. In addition, it has been possible to discover unrevealed insights about the nonlinear behavior of the brain in diverse emotional situations and which brain regions are the most involved in cognitive and emotional processes.