Machine learning-based prediction of postoperative delirium from intraoperative EEG signal alterations in brain functional connectivity.

OBJECTIVES: Postoperative delirium (POD) is a frequent complication following cardiovascular surgery and requires timely intervention. While early risk prediction holds clinical value, practical tools remain limited. This study aims to identify EEG-based features associated with POD and evaluate their predictive potential using machine learning. METHODS: 70 patients undergoing cardiovascular surgery were enrolled, including 35 with POD and 35 without. Functional connectivity was assessed using phase-locking values (PLV) derived from intraoperative EEG data. Network edges showing significant between-group differences in the delta, theta, alpha, and beta bands were selected as features for input into an XGBoost classifier. RESULTS: POD patients exhibited the most pronounced alterations in the theta band, with 12 significantly different PLV edges, followed by 1 in delta, 4 in alpha, and 5 in beta. All significant edges showed increased PLV values. Theta-band features demonstrated the highest predictive performance, achieving 90.7% accuracy and an Area Under the Receiver Operating Characteristic Curve of 0.91. CONCLUSIONS: Our findings reveal significantly enhanced theta-band connectivity and general abnormal cortical synchrony in POD patients. SIGNIFICANCE: These EEG-based features may serve as potential biomarkers and provide a foundation for machine learning models to enable early POD risk prediction.