Explainable Temporal Deep Learning for EEG-Based Depression Detection Using Resting-State Brain Dynamics

Depression is a major mental health disorder, and EEG-based automated detection is emerging as a potential objective diagnostic tool. However, achieving both high accuracy and interpretability remains challenging due to the complex spatiotemporal structure of EEG signals. This study proposes an explainable deep learning framework for depression detection using resting-state EEG data.

A retrospective computational study using deep learning was conducted using EEG data from 122 subjects in the OpenNeuro dataset (ds003478). Based on Beck Depression Inventory (BDI) scores, participants were classified into control (BDI ≤ 13; n = 76) and depressive (BDI ≥ 20; n = 30) groups; intermediate cases (BDI 14-19; n = 16) were excluded to reduce label ambiguity and construct a high-confidence binary classification framework, although this may reduce applicability to mild or subclinical depression, resulting in a final cohort of 106 subjects. Preprocessing included band-pass filtering (1-40 Hz), 50 Hz notch filtering, Independent Component Analysis, and average referencing. A subject-wise 5-fold cross-validation was applied. A CNN-BiLSTM architecture with an attention mechanism integrated with explainable AI techniques (Grad-CAM and SHAP) was developed.

The proposed model achieved an accuracy of 89.76%, an F1-score of 89.58%, and an AUC of 0.936. Ablation analysis confirmed the contribution of temporal modeling and attention mechanisms. Explainability analysis using Grad-CAM and SHAP showed that frontal EEG channels were the most influential in classification, consistent with neurophysiological findings.

The proposed framework provides an accurate and interpretable method for EEG-based depression detection, supporting applications in computational psychiatry and decision-support systems.