Interpretable machine learning models as an instrument for explaining predictive assessments in cardiology
Karina I. Shakhgeldyan, V. Yu. Rublev, Nikita S. Kuksin, Regina L. Pak, B. I. Geltser · The Bulletin of Contemporary Clinical Medicine · 2025
Abstract. Introduction. In-hospital mortality during coronary artery bypass grafting in patients with coronary artery disease varies from 1 to 6%. In recent years, machine learning methods have been increasingly used to predict in-hospital mortality after coronary artery bypass grafting. However, their widespread implementation in clinical practice is limited by the non-transparency of the predictive results. Aim. The aim of the study is to develop interpretable machine learning models for predicting in–hospital mortality in patients with coronary artery disease after coronary artery bypass grafting, based on the multi–level categorization of continuous predictors. Materials and Methods. A single-center retrospective study was conducted using data from 1504 case histories of patients with coronary artery disease after coronary artery bypass grafting. Two groups were identified: The first one included 79 (5.3%) patients who had died in the hospital, and the second one included 1425 (94.7%) patients with favorable coronary artery bypass grafting outcomes. Statistical analysis methods were used to select predictors, and models were developed using machine learning methods, such as multivariate logistic regression, random forest, and stochastic gradient boosting. Grid search optimization and the Shapley additive explanation method were used for the multi-level categorization of predictors. Results and Discussion. As a result of the multi-step analysis of the clinical and functional status indicators in patients with coronary artery disease, in-hospital mortality predictors were identified, validated, and categorized. Predictive models of in-hospital mortality with multi-level categorization of predictors surpassed the models with continuous predictors in terms of performance (AUC 0.882 vs 0.856) and provided clinical interpretability of the generated conclusions. Conclusions. Interpretable predictive machine learning models have been developed, with the main element being the multi-level categorization of in-hospital mortality predictors. This approach was characterized by high accuracy and interpretability of the prediction results, which allows using it for other cardiological conditions.