Preoperative ECG-based machine learning model for early prediction and risk stratification of postoperative delirium

Postoperative delirium (POD) is a serious neuropsychiatric complication that commonly occurs within days after surgery. Early identification of high-risk individuals is crucial for implementing effective preventive and management strategies. This paper developed a machine learning model for early POD prediction solely based on ECG as a single biomarker. A total of 596 surgical patients aged 40 years or older were enrolled at a tertiary hospital in 2023. POD occurrence within 3 days after surgery was assessed using the 3D-CAM-CN, yielding 136 POD cases and 460 non-POD cases. This cohort was randomly divided into a training and internal validation set at a 3:1 ratio. To evaluate model generalizability, an independent external validation cohort comprising 102 surgical patients was recruited from a separate tertiary hospital in 2026. A novel feature selection algorithm identified the top 19 informative features from 12-lead ECG signals. Six machine learning models were developed and evaluated on both the internal validation and external validation cohorts. Shapley Additive Explanations (SHAP) analysis was applied to interpret features at both the global and individual levels. The extreme gradient boosting model achieved superior performance, with an AUC as high as 0.97, an accuracy of 92.0%, sensitivity of 85.3%, precision of 80.6%, and F1-score of 82.9%. Furthermore, the model maintained good performance in the independent external validation cohort, achieving an AUC of 0.840 with an accuracy of 90.20%. Global SHAP analysis identified heart rate variability (HRV) features and inter-lead R-peak timing features as major contributors to the model’s predictions. Inter-lead R-wave delay features emerged as informative contributors to model prediction. Individual-level SHAP analysis further revealed patient-specific variability in feature contributions, supporting personalized risk assessment. A clinician-friendly online platform was developed to facilitate risk stratification using POD risk scores, thereby supporting preoperative decision-making and personalized perioperative management. This preoperative ECG-based machine learning model provides a highly accurate, non-invasive, low-cost, and accessible tool for early POD risk prediction. The incorporation of SHAP analysis increases clinical interpretability, and the developed visualized online platform offers patient-specific risk stratification to support individualized perioperative management.