A Hybrid Deep Learning Architecture with Retrieval-Augmented Explanation Generation for Interpretable EEG-Based Stress Classification Across Multiple Paradigms

Abstract

This paper presents a hybrid deep learning architecture integrating convolutional neural networks, bidirectional long short-term memory networks, self-attention, and retrieval-augmented generation (RAG) for accurate and interpretable electroencephalography (EEG)-based stress detection. The compact EEG encoder contains 138,000 trainable parameters and combines three convolutional blocks, a two-layer bidirectional LSTM, and self-attention, while the RAG module grounds explanations in scientific literature using FAISS vector search and a frozen Sentence-BERT encoder. The system is evaluated on two datasets representing distinct stress paradigms: DEAP (32 subjects, emotional arousal as stress proxy) and SAM-40 (40 subjects, cognitive stress from arithmetic and Stroop tasks). Under leave-one-subject-out cross-validation, the model achieves accuracies of 94.7% on DEAP and 93.2% on SAM-40, including a 12.6 percentage-point improvement over the previous state of the art on SAM-40. Signal analysis identifies consistent biomarkers, including 31–33% alpha-band power suppression, an 8–14% reduction in the theta-to-beta ratio, and a shift in frontal alpha asymmetry towards right-hemisphere dominance. The RAG module achieves 89.8% agreement with domain experts on explanation quality, while ablation results indicate limited impact on classification accuracy. Cross-dataset transfer shows 21.8–26.5% accuracy drops between stress paradigms. Gradient-based analysis identifies frontal alpha power, theta-to-beta ratio, and frontal alpha asymmetry as key decision features. The proposed framework supports real-time, interpretable EEG stress monitoring for clinical decision support and related applications.