Development and validation of an explainable machine learning model for predicting myocardial injury after non-cardiac surgery: multicenter, retrospective study (Preprint)

Abstract

Myocardial injury after non-cardiac surgery (MINS) is an easily overlooked complication but closely related to postoperative cardiovascular adverse outcomes, so the improved risk prediction tools are critically needed.

To develop and validate an explainable machine learning model for predicting MINS in older patients undergoing non-cardiac surgery.

The retrospective cohort study assessed operations performed on non-cardiac surgical older patients at center 1 in the training set. The least absolute shrinkage and selection operator (LASSO) method and recursive feature elimination (RFE) methods were used to select key features. Prediction performance was measured by the area under the receiver operating characteristic curve (AUC) as the main evaluation metric to select the best algorithms. Validation data from the two datasets were explored to validate the performance of the model and the developed model was compared with the RCRI model. 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.

A total of 12424 patients were included in training set, 4754 in center 1, 2245 in center 2 were included as validating sets. The best-performing model for prediction was CatBoost algorithm, achieving an AUC of 0.805 (95% confidence interval, 0.778–0.831) in the training set, and validating with an AUC of 0.780 in center 1, 0.70 in center 2, with superior performance compared to RCRI (AUC:0.636, P<0.01). The SHAP values indicated the ranking of the level of importance of each variable, and preoperative serum creatinine concentration, red blood cell distribution width, and age accounted for the top three. The results from SHAP method can make predictions towards event with a positive values or non-event with negative value and an explicit explanation of individualized risk prediction.

The Catboost model demonstrated superior capability of predicting individual-level risk of MINS, and the explainable perspective can allow identification of potentially modifiable sources of risk on patient level.