Neurophysiological Signal Analysis and Stress Classification System

Abstract

Stress-related problems have grown to be a major issue in contemporary settings for academia and workplace since it affects brain function, emotion control, and overall neurological health negatively. Stress measurement techniques used previously mainly involve subjective evaluation which makes the technique unreliable when it comes to real-time monitoring purposes. Electroencephalogram (EEG) represents a tool capable of direct observation of neuroactivity which makes it a promising candidate for objective detection of stress state. But at the same time, the use of EEG in stress detection faces difficulties due to noise in the EEG signals, inter-subject variance, and computational limitations. In this paper we introduce real-time stress classification based on EEG with signal preprocessing, interpretable machine learning and embedded feedback. The preprocessing includes band-pass filtering, Singular Spectrum Analysis (SSA) noise removal and Effect-Matched Spatial Filtering (EMS) feature enhancement. The feature extraction process is performed by using Random Forest, Support Vector Machine and XGBoost classifiers. SEED database was used for modelling, whereas WESAD for cross-dataset validation. It resulted in 82.7% multi-class classification accuracy with low inference latency.