Exploring the neural mechanisms of mild cognitive impairment in elderly patients with coronary artery disease using machine learning and source-localized EEG

Objective This study seeks to investigate the electrophysiological mechanisms associated with mild cognitive impairment (MCI) in elderly patients with coronary artery disease (CAD) through the application of source-reconstructed EEG in conjunction with machine learning methodologies. Methods We retrospectively analyzed clinical data and resting-state 64-channel EEG recorded during hospitalization at The First Hospital of Changsha. Participants included primary hypertension without CAD ( n = 53) and CAD with primary hypertension ( n = 117), with CAD stratified by Montreal Cognitive Assessment (MoCA) into MCI ( n = 49) and cognitively normal ( n = 68). EEG sources were reconstructed using an ICBM152-based head model and BEM forward modeling, yielding 82 Brodmann-atlas ROIs; functional connectivity was quantified using lagged phase synchronization (LPS) in delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz) bands. Group comparisons applied false discovery rate correction. For MCI classification among patients with CAD, the dataset was randomly split into training and testing sets (7:3). Feature selection was performed in the training set using an independent-samples t -test followed by L1-penalized logistic regression. Subsequently, eight machine-learning classifiers were trained using the selected LPS features, with hyperparameters optimized by grid search under five-fold cross-validation. Model interpretability was assessed using SHAP. Results Baseline demographics and vascular comorbidities were comparable across groups; MoCA scores were lower in the MCI subgroup. Relative to hypertensive controls without CAD, cognitively normal CAD patients showed reduced frontal connectivity, including decreased alpha-band LPS (BA8L–46R) and beta-band LPS (BA44L–44R). Compared with cognitively normal CAD, CAD with MCI exhibited broader multi-band dysconnectivity across alpha, beta, theta, and delta bands, with mixed delta-band changes. In the test set, the Gradient Boosting model achieved the best performance for identifying MCI within CAD (AUC = 0.895). SHAP highlighted the most influential features, led by decreased alpha-band BA8L–46R connectivity, alongside delta- and beta-band alterations. Conclusion Coronary artery disease is associated with frontal network disruption, which becomes more extensive and frequency-diverse as MCI progresses. Interpretable machine learning further highlights a small set of connectivity abnormalities—particularly within premotor–prefrontal pathways—as candidate markers for MCI classification within a CAD cohort, supporting a vascular-relevant interpretation, which warrants further validation.