Interpretable deep learning for depression detection in neurological patients using EEG signals

Depression affects over 280 million people worldwide, with neurological patients particularly prone to medication-induced episodes. Conventional diagnostic approaches rely on subjective evaluations, limiting reproducibility and consistency in clinical settings. This study proposes an interpretable deep learning framework for objective depression detection using EEG signals. We hypothesize that combining EEG-based features with explainable artificial intelligence can provide both high accuracy and transparency in diagnosis. The model was trained on EEG data from 232 neurological patients, achieving 98 % classification accuracy. Interpretability was enhanced through SHAP (SHapley Additive exPlanations) analysis, which identified clinically meaningful EEG biomarkers such as the delta/alpha ratio and theta band power. This paper highlights the following contributions: Integration of EEG features with a lightweight deep learning model for depression detection High diagnostic accuracy achieved while maintaining interpretability for clinicians An objective tool that is compatible with existing EEG infrastructure, supporting clinical adoption These results show that our framework bridges predictive performance with interpretability, offering a transparent and scalable EEG-based diagnostic tool. We conclude that this approach can complement clinical decision-making, reducing dependence on subjective evaluation and enabling more consistent, data-driven mental health care.