Multimodal Machine Learning Model Predicting Postoperative Delirium Based on Heart Rate Variability: A Prospective Observational Study.
BACKGROUND Postoperative delirium is a common and serious complication after general anesthesia; its accurate prediction remains a substantial challenge in perioperative medicine. Existing models primarily rely on clinical variables and may have limited predictive accuracy. This study aimed to evaluate the added value of heart rate variability parameters in predicting postoperative delirium and construct an interpretable multimodal predictive model. METHODS In this prospective observational study, 1418 patients undergoing general anesthesia were included. Seventy-three features, including electrocardiogram abnormalities and heart rate variability time-, frequency-, and nonlinear-domain indicators, were extracted from electrocardiogram data. Postoperative delirium was assessed using the Chinese version of the 3-Minute Diagnostic Interview for Delirium within 3 days postoperatively. Feature selection was conducted by combining least absolute shrinkage and selection operator (LASSO) regression, the Boruta algorithm, and random forests, and 10 machine learning models were developed. Model performance was evaluated through receiver operating characteristic curves and decision curve analysis, with interpretability assessed via Shapley additive explanations. Clinical prediction tools were derived from key features. We used an external validation set to further evaluate the generalization ability of the models. RESULTS Postoperative delirium occurred in 255 (18%) patients. Seventeen key predictors were identified in total. The combined clinical-electrocardiogram-heart rate variability model demonstrated the highest predictive performance (area under the curve = 0.728), outperforming clinical-only (area under the curve = 0.673) and electrocardiogram-only models (area under the curve = 0.679). Logistic regression showed the highest discrimination. In the external validation set, the model maintained robust performance with an area under the curve value of 0.836. Shapley additive explanations highlighted seven core predictors: atrial or ventricular arrhythmia, operative time, ST-segment abnormalities, age, American Society of Anesthesiologists classification, heart rate variability entropy, and overall electrocardiogram abnormalities. A nomogram and online platform enabled personalized risk assessment. CONCLUSIONS Our results indicate that integrating heart rate variability with clinical and electrocardiogram features significantly enhances the personalized predictive efficacy of postoperative delirium.