A comparative study of time–frequency features based spatio-temporal analysis with varying multiscale kernels for emotion recognition from EEG

Abstract

EEG-based emotion recognition has become recognized as a crucial field of study, utilizing brainwave patterns for understanding human emotional states. To identify emotions from EEG data, this study examines several time–frequency features utilizing spatiotemporal analysis using the DEAP dataset. Down sampling, bandpass filtering, segmentation, trimming, labeling, and common reference averaging are all part of the data pre-processing pipeline. Continuous Wavelet Transform (CWT) with Morlet wavelets was used to extract features, followed by the computation of differential entropy, wavelet energy, cross-correlation, and phase locking value (PLV). The distribution of features, participant-specific changes, and associations were examined through an exploratory feature analysis. Subsequently, final representations functioning in a spatiotemporal manner were constructed. For classification, a 3D Convolutional Neural Network with different kernel sizes (3×3×3, 5×5×5, and 7×7×7) was employed. Training accuracies reached up to 98.86% for arousal and 98.97% for valence, demonstrating robust generalization of particular feature-kernel combinations. The results also show that the 5×5×5 kernel size achieved the highest test accuracy for arousal (96.13% for differential entropy) and valence (96.19% for wavelet energy). In order to show the model’s universality, it was also validated using the popular SEED dataset. With PLV (kernel 7×7×7), the arousal validation accuracy was 97.42%, and with WE (kernel 3×3×3), the valence validation accuracy was 97.50%. This work sheds light on the choice of features and kernels for emotion identification models based on EEG.