From EEG signals to quantitative assessment: predicting depression severity using a novel deep learning framework

The assessment of depression severity still relies primarily on subjective rating scales, with a lack of objective quantitative biomarkers. This study aimed to develop a deep learning model based on resting-state electroencephalography (EEG) for automated and quantitative prediction of depression severity. Resting-state EEG data were collected from 70 patients with depressive disorder (DD) and 30 healthy controls (HC). A novel deep learning framework—PLI_GE_gMLP—was proposed. By integrating phase lag index (PLI), graph embedding (GE), and gated multilayer perceptron (gMLP), the model effectively captures both the spatiotemporal dependencies and functional connectivity features of EEG signals for depression severity prediction. The PLI_GE_gMLP model achieved a mean absolute error (MAE) of 4.30 in predicting depression severity, significantly outperforming various traditional machine learning (e.g., RF, XGBoost, LightGBM) and deep learning models (e.g., ResNet, GENet). SHAP (Shapley Additive Explanations) interpretability analysis further revealed that functional connectivity in the frontal and temporal lobes, particularly PLI features in the Beta and Theta bands, contributed most significantly to the predictions, aligning closely with the neuropathological mechanisms of DD. Through the synergistic integration of GE and gMLP, the proposed PLI_GE_gMLP model enables accurate and interpretable prediction of depression severity.