Caglar UYULAN
The Journal of Neurobehavioral Sciences - 2026;13(2):49-61
Aims: Major depressive disorder (MDD) is a prevalent psychiatric disorder, and objective neurophysiological measures may complement clinical assessment. This study assessed whether short resting-state EEG segments can distinguish MDD from healthy controls (HC) using two compact deep learning architectures and whether channel ablation provides interpretable spatial information. Materials and Methods: The retrospective dataset included 112 eyes-closed resting-state EEG recordings from 56 individuals with MDD and 56 HC. Ninety-two recordings were used for model development and internal evaluation, while 20 independent recordings (10 MDD, 10 HC) were reserved for external validation. Nineteen channels were standardised to 125 Hz and segmented into non-overlapping 1-s windows. Lightweight 1D Convolutional Neural Network (Light 1D-CNN) and CNN-MiniTransformer models were trained using identical inputs. Channel ablation quantified global and class-specific spatial importance. Results: Internally, Light 1D-CNN achieved 96.43% accuracy, 96.39% F1-score, and 0.9941 AUROC; CNN-MiniTransformer achieved 95.26%, 95.31%, and 0.9906, respectively. On external validation, Light 1D-CNN achieved 93.72% accuracy and 0.9785 AUROC, while CNN-MiniTransformer achieved 94.65% and 0.9872. P4 was the most influential MDD-related channel in both models, followed by Fz and F4, with parietal and frontal regions showing highest importance. Attention-based modelling showed greater external stability while preserving efficient temporal feature extraction. Conclusion: Compact deep learning models maintained strong segment-level discrimination on independent EEG recordings and converged on similar parietal-frontal importance patterns, supporting further participant-level, multicentre validation of interpretable EEG-based MDD classification.
