Comparative Analysis of EEG Signal Processing Methods for Emotion Recognition: Towards a Standardized Procedure

Emotion recognition has acquired growing interest in many scientific and applied fields, proving to be crucial for the effectiveness of the human-machine interaction, the development of virtual reality environments, and the improvement of various clinical applications, such as supporting therapies for emotional disorders or assessing stress in work or educational contexts. One of the most widespread and reliable technologies for identifying mental and emotional states in a non-invasive way is the use of electroencephalographic (EEG) signals, which allow monitoring brain activity in real time and with high precision. Starting from EEG signals, it is possible to calculate specific indicators, such as Valence, Arousal, Dominance, Engagement and Stress, which allow to identify a person's emotional state with greater precision. These indicators are based on the measurement of the activity in different frequency bands of some specific channels, and are closely linked to brain processes. The combination of some indicators on two or three-dimensional spaces, depending on the theories, allows to determine with high accuracy the emotion felt by the subject at any moment. However, the use of EEG for emotion recognition requires standardized approaches and strong methodologies for data analysis, since the accuracy of the interpretation depends both on the quality of the signal and on the algorithms used to process it. Currently, several software and toolboxes for the management and analysis of EEG data are widespread, each with specific methodologies for signal processing. Since there is no standardized procedure for EEG signal processing to date, the aim of this thesis is to compare the effectiveness and efficiency of three widely used toolboxes: EEGLAB, Brainstorm and Letswave, implemented in Matlab, in order to identify the most practical tool that guarantees reliable results. The dataset obtained from an experiment conducted at the 3D Lab of Politecnico di Torino is used, where 39 participants navigate in two virtual reality (VR) environments on a desktop monitor, wearing the Emotiv EPOC X device for the acquisition of EEG signals. The subjects have to perform a job interview in two scenarios, each characterized by a different level of stress: one in which it is elicited, and another in which it is alleviated. This work aims to evaluate the performances in terms of accuracy, absolute and relative errors, computational costs, usability and graphic rendering of each toolbox, developing, on the basis of a common pipeline, a specific script for each of the three cases that, starting from the raw EEG signals, calculates the powers for each band in each single channel, using exclusively the functions included in the specific toolbox, and then extracts the indicators useful for the emotion recognition. The results obtained highlight how two of these toolboxes guarantee an accuracy greater than 90% compared to the reference data provided by the Emotiv software, although with significant differences in execution times, while the last toolbox has not proven to be usable in a direct and automatic mode, requiring manual operations and a considerable loss of time. The work is therefore proposed as a contribution to the standardization of EEG signal processing techniques for future applications in the field of emotion recognition, seeking an approach that favors accuracy and ease of use.