Comparative analysis of explainable machine learning models for cardiovascular risk stratification using clinical data and shapley additive explanations
Komal Kumar Napa, Rajkumar Govindarajan, S. Siva Sathya, Jeff Murugan, Bindu Kolappa Pillai Vijayammal
Abstract
Heart disease remains a leading cause of mortality globally, demanding timely and reliable diagnostic support in clinical settings. This study proposes an interpretable machine learning framework that leverages Random Forest (RF) models integrated with SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP) to predict heart disease risk using clinical data. By addressing the limitations of “black-box” predictive systems, the framework ensures transparency and trust in decision-making. Multiple machine learning classifiers were benchmarked, with RF demonstrating high performance among the classifiers under analysis in terms of accuracy and interpretability. A streamlined, user-friendly graphical interface was developed using Streamlit to facilitate real-time risk assessment, feature-level explanations, and actionable clinical insights. The system incorporates electronic health records, utilising preprocessing and imputation strategies to enhance model robustness. Experimental evaluations demonstrated that the proposed method strikes a balance between good predictive accuracy and interpretability, making it suitable for integration into clinical workflows. This work contributes to the advancement of explainable AI in healthcare engineering, supporting clinicians in early diagnosis and preventive care for cardiovascular conditions. • An interpretable machine learning framework using Random Forest is proposed for heart disease prediction. • SHAP and Partial Dependence Plots are integrated for transparent clinical feature explanations. • KNN imputation improves data quality in Electronic Health Records for robust model training. • A real-time Streamlit GUI enables interactive risk prediction and visual explanation for clinicians. • The model achieves strong predictive performance (81.3 % accuracy) while ensuring explainability.