Interpretable Machine Learning for 30-Day Mortality in PIVSR: A Cohort Study

Abstract

Post-infarction ventricular septal rupture (PIVSR) is a fatal mechanical complication of acute myocardial infarction. We aimed to develop and interpret a machine learning (ML) model to predict 30-day mortality using routinely available clinical variables in PIVSR patients.

This retrospective cohort study included consecutive PIVSR patients treated at Fuwai Central-China Cardiovascular Hospital from 2018 to 2024. Candidate predictors were screened by three complementary procedures including bootstrap resampling with LASSO, stepwise logistic regression, and the Boruta algorithm. Nine supervised ML algorithms were trained and compared. Model performance was comprehensively evaluated using multiple discrimination, calibration, and clinical utility metrics. Model interpretability was examined using Shapley Additive Explanations (SHAP). We additionally implemented the final model as a web-based calculator to support individualized risk estimation.

A total of 237 PIVSR patients were analyzed. Nine clinical predictors were identified. The CatBoost model demonstrated favorable overall performance in both the training set (AUC = 0.92, 95%CI: 0.89–0.94) and testing set (AUC = 0.88, 95%CI: 0.79–0.95). In the testing set, CatBoost achieved an accuracy of 0.79, sensitivity of 0.85, specificity of 0.71, F1-score of 0.82, and Brier score of 0.15. SHAP analysis identified operation, hemodynamic status, inflammatory markers, and renal function as key contributors.

We developed, evaluated, and interpreted ML models for predicting 30-day mortality in patients with PIVSR. The CatBoost model demonstrated favorable predictive performance and transparent interpretability. Hemodynamic compromise, inflammation, and renal dysfunction were identified as key predictors of adverse outcomes. An exploratory web-based calculator was developed to support dynamic in-hospital prognostication.