An Explainable Self-Supervised Temporal Transformer for Participant-Level Classification of Depressive Symptoms From...

Resting-state electroencephalography (EEG) has emerged as a promising non-invasive modality for investigating depressive symptoms; however, conventional deep-learning methods often require large labeled datasets and provide limited interpretability. We developed an Explainable Self-Supervised Temporal Transformer Network for participant-level classification of BDI-defined elevated depressive symptoms from resting-state EEG.

The framework employed masked temporal modeling for self-supervised pretraining followed by supervised fine-tuning. Model performance was evaluated using repeated nested participant-wise cross-validation and compared with conventional machine-learning and deep-learning baselines. Explainability was investigated through transformer attention maps and SHAP-based channel attribution.

The proposed model achieved a receiver operating characteristic area under the curve (ROC-AUC) of 0.82 (95% CI: 0.73-0.90), a precision-recall AUC of 0.69, balanced accuracy of 0.75, sensitivity of 0.72, and specificity of 0.78. The positive predictive value was 0.57, indicating a substantial false-positive burden in the evaluated cohort. The proposed model showed the highest point-estimate ROC-AUC among the evaluated baselines, with a 0.07 improvement over EEGNet; however, the overlapping uncertainty intervals and the small elevated-symptom cohort warrant cautious interpretation of comparative superiority. Performance decreased from an ROC-AUC of 0.82 in Run 1 to 0.78 in Run 2 and to 0.72-0.73 under cross-condition transfer, supporting interpretation of these analyses as internal reproducibility and condition-shift assessments rather than independent external validation. Removing self-supervised pretraining reduced the ROC-AUC to 0.74. Explainability analyses identified distributed temporally relevant intervals and frontal-midline to centro-parietal channel contributions, with attribution interpreted as model-relevance evidence rather than causal biomarker evidence.

Overall, the findings provide preliminary internal evidence that self-supervised temporal representation learning can support participant-level classification of BDI-defined elevated depressive symptoms from resting-state EEG under the evaluated experimental conditions. The findings do not establish clinical diagnostic validity, screening readiness, or independent external generalizability.