Explainable Machine Learning for Cardiovascular Disease Detection: Addressing Interpretability Limitations and Evaluating Transparency for Clinical Decision Support

Abstract

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, 
accounting for approximately 17.9 million deaths annually. Although machine learning (ML) 
models have demonstrated high predictive performance for CVD detection, their 
inherent black-box nature limits clinical trust, interpretability, and adoption in 
healthcare practice. This study addresses three objectives: (1) to examine the limitations 
of conventional ML models for CVD prediction; (2) to develop an interpretable ML 
framework that enhances model transparency through the integration of SHapley 
Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations 
(LIME); and (3) to evaluate the predictive performance, transparency, and clinical 
applicability of the resulting models. Logistic Regression, Random Forest, and Extreme 
Gradient Boosting (XGBoost) classifiers were trained and evaluated using the Kaggle 
Cardiovascular Disease Dataset comprising 70,000 patient records and 11 predictor 
variables. Model performance was assessed using accuracy, precision, recall, F1-score, 
receiver operating characteristic area under the curve (ROC-AUC), calibration error, and 
a structured transparency scoring framework. XGBoost achieved the highest 
discriminative performance, with an ROC-AUC of 0.794, an accuracy of 0.729, and the 
lowest calibration error of 0.024, whereas Logistic Regression attained the highest 
transparency score (73.8/100). Global model interpretation using SHAP identified 
cholesterol, age, systolic blood pressure, and diastolic blood pressure as the most 
influential predictors, consistent with established clinical evidence. Similarly, LIME 
provided patient-specific explanations that highlighted the same dominant risk factors, 
thereby enhancing decision interpretability at the individual level. These findings 
demonstrate that high predictive accuracy and model explainability can coexist within a 
unified ML framework, providing a replicable blueprint for transparent cardiovascular 
risk assessment, particularly in resource-limited healthcare settings.