An ACS-Stacking Prediction Model Based on Interpretable Machine Learning

Abstract

Background: Acute coronary syndrome (ACS) is an important disease threatening human health, and the rapid differential diagnosis of acute myocardial infarction still requires further studies. Purpose: This study aims to establish an interpretable machine learning (ML) model and perform visual and interpretable analysis to the prediction results using SHAP (SHapley Additive exPlanation). Then significant correlation indicators are determined to assist clinicians in providing rapid and effective identification for ACS patients. Method: This study involves the clinical data of 813 patients from the Shanxi Cardiovascular Hospital , which is described by 24 predictor variables in relation to demography/comorbidity characteristics and in-hospital complications. Taking the binary variables of “Acute Myocardial Infarction (AMI) and Unstable Angina (UA)” as target variables, we have trained and evaluated the performance of seven ML models in this study and fused Adaboost, Xgboost and Randomforest with better performance in the test set into the best interpretable Stacking fusion model (named as: ACS-Stacking prediction model). Results: The ACS fusion prediction model achieves an AUC value of 0.96562 in the test set and an accuracy of 89% under 10-fold cross-validation. This study interprets the model using SHAP. Among the related continuous variables, neutrophil and admission heart rate have a positive effect on the mode while LVEF, BMI, systolic pressure and diastolic pressure have a negative effect on the model. However, age is not significantly correlated with target variables. For the classified variables, the patients with smoking history are predisposed to myocardial infarction; sex and history of hypertension are not significantly correlated with target variables. Conclusion: This study shows that the interpretable ACS-Stacking prediction model has a good differential prediction effect on myocardial infarction and angina pectoris, and the Summary Plot shows the specific effect of ten significant correlation indicators on the output of the model. This conclusion helps clinicians to rapidly identify ACS patients in clinical diagnosis based on the model prediction result, model visualization and clinical experience.