Answers are grounded in this paper and cite their sources.
Major Depressive Disorder (MDD) is a common mental health condition with significant societal and clinical impact, yet current diagnostic methods remain subjective and often unreliable. This study proposes a new pruned hybrid Convolutional Neural Network (CNN)-Transformer architecture for depression detection based on the Electroencephalography (EEG) signal. This hybrid system uses a pretrained CNN for the spatial feature extraction unit and a lightweight Transformer as a classification unit. Then, for reduced complexity and increased throughput, a new Center-based Separation Index (CSI) pruning scheme removes redundant Transformer blocks adaptively based on feature separability. Experiments on a clinical EEG dataset with 58 subjects (30 with MDD and 28 Healthy Controls) show that the EfficientNetV2 + CSI-Pruned Data-efficient Image Transformer (DeiT)-Tiny model achieves 95.35% accuracy with average sensitivity and specificity of 95%. Besides, model reduction saves parameters by 10.5%, FLOPs by 4.4%, and increases throughput by 7.6%. Consequently, the results confirm that this novel geometry-aware pruning methodology is highly effective for designing dense, accurate, and efficient hybrid neural networks, advancing their feasibility for EEG-based depressionrelated systems.
No takes yet. Share an insight, caveat, or question.
Shahsafdari et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: