XGBoost machine learning algorithm for predicting unplanned readmission in elderly patients with coronary heart disease

Most studies have focused on 30-day rather than 1-year unplanned readmissions in elderly patients with coronary heart disease (CHD). The extreme gradient boosting (XGBoost)-based model demonstrates good predictive performance and explainability. This study aimed to establish an XGBoost model to predict 1-year unplanned readmission in Chinese elderly CHD patients. The clinical data of elderly CHD patients were collected retrospectively. The stepwise forward method was used for feature selection. The area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC) and calibration curve were used to evaluate the performance of the ML models. SHapley Additive exPlanations (SHAP) analysis was used to evaluate the importance of features. A total of 2137 patients were enrolled. The AUROC of the XGBoost model was 0.704, and the AUPRC was 0.392. SHAP analysis showed that length of stay (LOS), age-adjusted Charlson comorbidity index (ACCI), monocyte count, blood glucose level and red blood cell (RBC) count were the most important predictors. XGBoost can predict 1-year unplanned readmissions in elderly patients with CHD and identify the risk factors.