Development and Validation of an Explainable Machine Learning Model for Predicting Myocardial Injury After Noncardiac Surgery in Two Centers in China: Retrospective Study

Abstract

Abstract Background Myocardial injury after noncardiac surgery (MINS) is an easily overlooked complication but closely related to postoperative cardiovascular adverse outcomes; therefore, the early diagnosis and prediction are particularly important. Objective We aimed to develop and validate an explainable machine learning (ML) model for predicting MINS among older patients undergoing noncardiac surgery. Methods The retrospective cohort study included older patients who had noncardiac surgery from 1 northern center and 1 southern center in China. The data sets from center 1 were divided into a training set and an internal validation set. The data set from center 2 was used as an external validation set. Before modeling, the least absolute shrinkage and selection operator and recursive feature elimination methods were used to reduce dimensions of data and select key features from all variables. Prediction models were developed based on the extracted features using several ML algorithms, including category boosting, random forest, logistic regression, naïve Bayes, light gradient boosting machine, extreme gradient boosting, support vector machine, and decision tree. Prediction performance was assessed by the area under the receiver operating characteristic (AUROC) curve as the main evaluation metric to select the best algorithms. The model performance was verified by internal and external validation data sets with the best algorithm and compared to the Revised Cardiac Risk Index. The Shapley Additive Explanations (SHAP) method was applied to calculate values for each feature, representing the contribution to the predicted risk of complication, and generate personalized explanations. Results A total of 19,463 eligible patients were included; among those, 12,464 patients in center 1 were included as the training set; 4754 patients in center 1 were included as the internal validation set; and 2245 in center 2 were included as the external validation set. The best-performing model for prediction was the CatBoost algorithm, achieving the highest AUROC of 0.805 (95% CI 0.778‐0.831) in the training set, validating with an AUROC of 0.780 in the internal validation set and 0.70 in external validation set. Additionally, CatBoost demonstrated superior performance compared to the Revised Cardiac Risk Index (AUROC 0.636; P<.001). The SHAP values indicated the ranking of the level of importance of each variable, with preoperative serum creatinine concentration, red blood cell distribution width, and age accounting for the top three. The results from the SHAP method can predict events with positive values or nonevents with negative values, providing an explicit explanation of individualized risk predictions. Conclusions The ML models can provide a personalized and fairly accurate risk prediction of MINS, and the explainable perspective can help identify potentially modifiable sources of risk at the patient level.