EEG-Based Multi-Level Mental State Classification Using Partial Directed Coherence and Graph Convolutional Networks: Impact of Binaural Beats on Stress Mitigation

Abstract

This study addresses limitations in EEG-based stress detection research by developing a novel approach to differentiate multiple mental states in different stress baseline population samples. Utilizing EEG signals, graph convolutional neural networks (GCNs), and binaural beats stimulation (BBs), the research investigates stress detection and reduction in two population sample groups with distinct baselines (group 1: low daily baseline, and group 2: stressed daily baseline). The experiment comprises four phases: rest state, control alertness, stress induction, and stress mitigation. Mental states were assessed using behavioral data: reaction time to stimuli (RT) and target detection accuracy, subjective reports: Perceived Stress Scale scores (PSS-10), biochemical indicators: salivary cortisol levels, and neurophysiological measure: EEG effective connectivity via Partial Directed Coherence (PDC). BBs significantly improved target detection accuracy by 31.6% and 22.8% for low and high-stress groups, respectively. PDC connectivity showed a shift to the temporal region during mitigation, indicating a return to a more balanced state. GCN classification achieved accuracies of 76.43±9.01 % and 76.32±7.79 % for each group, and 76.37±8.40 % for a common baseline. While 16-Hz BBs enhanced focusing abilities they did not significantly reduce subjective stress scores. This study highlights the complex relationship between cognitive performance, perceived stress, and neurophysiological measures, emphasizing the need for multifaceted stress research and management approaches.