An Interpretable Machine Learning Framework for Multilevel EEG-Based Stress Recognition

Stress is a hallmark of the majority mental health disorders, yet objective measures for its quantification remain limited. Electroencephalography (EEG) provides a scalable window into brain dynamics, but existing approaches often reduce this multiscale condition to binary states and rely on narrow spectral bands. This study presents a subject-specific framework to classify four stress levels, labeled using the Depression Anxiety Stress Scales (DASS), from the resting-state EEG without requiring task performance, a key advantage for clinical applicability. EEG data from 75 subjects were preprocessed with Independent Component Analysis (ICA) and AutoReject, followed by segmentation into overlapping epochs. From these segments, statistical, spectral, entropy-based, Hjorth, and fractal features were extracted. Correlation filtering and ANOVA were applied before subject-level cross-validation with imbalance correction. Models using 3s epochs achieved 92.8% accuracy and 92.6% macro-F1, outperforming 5s windows and deep learning baselines. SHAP confirmed physiological relevance, highlighting EEG-based features as objective neural markers of stress.