Machine learning-based predictive model for postoperative delirium of elderly patients with coronary heart disease undergoing non-cardiac surgery: a retrospective cohort study.

Background: Patients with advanced age and coronary heart disease (CHD) are at significantly increased risk for postoperative delirium (POD). However, there is no method to predict POD in elderly patients with CHD. Methods: Date from elderly patients with CHD who underwent non-cardiac surgery was collected. The dataset is subdivided into training and validation sets at a ratio of 7:3. Boruta algorithm, least absolute shrinkage and selection operator (LASSO) regression and multiple logistic regression analysis were used to select features. Machine learning method was used to construct a model for predicting the occurrence of POD. Receiver operating characteristic (ROC) curve, decision curve, calibration curve, specificity, sensitivity, accuracy, F1 score and Brier score were used to compare the predictive performance of these machine learning models, and the interpretability of the models was evaluated by Shapley additive interpretation (SHAP). Results: A total of 861 patients were included in the study. The incidence of POD was 16.6% (143/861). Seven key features were identified. Ten machine learning models were constructed. Among the models, gradient boosting model (GBM) performed better. The area under the ROC curve (AUC) is 0.856 (95% confidence interval [CI]: 0.796-0.916). The decision curve, calibration curve, specificity, sensitivity, accuracy, F1 score and Brier score were also relatively good. SHAP plots of GBM showed that Clinical Frailty Scale (CFS) grade, Mini-mental State Examination (MMSE) score, and Athens Insomnia Scale (AIS) score were significant predictors of POD in elderly CHD patients, and an easy-to-use calculator for predicting the risk of POD was developed based on the GBM model. Conclusion: This study developed a reliable GBM model for predicting the occurrence of POD in elderly patients with CHD. Higher CFS grade, lower MMSE score and higher AIS score significantly enhanced the predictive ability of the model. External validation of our model is needed before it can be applied in a clinical setting. Trial registration: Registration number of the Chinese Clinical Trial Registry: ChiCTR2500097325, Registration Date: 17/02/2025.