Automatic detection of depression using a CNN-Transformer model based on low-channel EEG data

Background: Accurate and portable diagnostic tools for Major Depressive Disorder (MDD) remain a critical unmet need in clinical practice. Electroencephalography (EEG), with its noninvasiveness and high sensitivity, presents an ideal modality, yet traditional multi-channel EEG systems limit clinical convenience and practical deployment. Methods: We developed a novel CNN-Transformer hybrid model using only eight EEG channels selected from frontal regions to detect MDD. Minimal preprocessing steps (basic filtering) were applied to EEG data before the model automatically learned both local and global signal features through convolutional and self-attention mechanisms. We validated this method through comprehensive experiments, including within-dataset and cross-dataset evaluations, and systematically analyzed the model’s robustness under different EEG conditions. Results: Our model achieved exceptional classification accuracy (94.27%), sensitivity (93.76%), specificity (94.79%), and AUC (0.9428) on the HUSM dataset under eyes closed conditions, outperforming state-of-the-art methods with significantly fewer EEG channels. Robust performance was retained under eyes-open conditions (accuracy 93.28%, AUC 0.9330), and high generalization was demonstrated with cross-dataset validation on the MODMA dataset (accuracy 93.14%, AUC 0.9285). Ablation studies confirmed that integrating CNN and Transformer modules contributed complementary advantages, markedly enhancing diagnostic performance. Conclusion: This study demonstrates that our CNN-Transformer model can reliably classify MDD patients using a greatly reduced EEG setup, highlighting its potential for developing portable, efficient, and clinically feasible EEG-based diagnostic systems. Future integration with portable EEG devices and multimodal assessments may further expand its utility in clinical and real-world settings.