Multidimensional EEG Feature Fusion and Interpretable Machine Learning for Identifying Cognitive Impairment in Depression

Cognitive impairment is a clinically important dimension of depression that remains insufficiently assessed using objective measures. Existing assessments largely rely on scale-based screening, whereas EEG-based computational studies are often limited by the use of single-feature representations, window-level data leakage, class imbalance, and limited interpretability. In this study, we developed an interpretable subject-level machine learning framework based on resting-state EEG to identify cognitive impairment in patients with depression and investigate potential electrophysiological markers. A total of 142 patients with depression were categorized as having depression without cognitive impairment (D-NCI) or depression with cognitive impairment (D-CI) based on their Montreal Cognitive Assessment (MoCA) scores. Using 16-channel resting-state EEG recordings, phase lag index, relative power, and fuzzy entropy features were extracted across five frequency bands to characterize functional connectivity, spectral activity, and signal complexity. Five-fold stratified cross-validation at the subject level, ensemble undersampling, and a two-stage SHAP-based feature selection strategy implemented within the training procedure were used to reduce the risk of data leakage, mitigate class imbalance, and improve interpretability. The final model achieved an accuracy of 80.39 ± 7.91%, sensitivity of 83.29 ± 8.30%, specificity of 73.89 ± 15.75%, F1-score of 85.67 ± 5.91%, and AUC of 83.74 ± 9.63%. The retained features were mainly distributed in the beta1 and beta2 bands and reflected alterations in interregional phase synchronization, local oscillatory activity, and signal complexity. These findings support the feasibility of an interpretable EEG-based framework with subject-level data partitioning to assist in identifying cognitive impairment in depression.