Performance Evaluation of SHAP and LIME for Cardiovascular Disease Prediction Using an XAI Framework

Abstract

Cardiovascular Diseases (CVD) is a leading cause of death globally, highlighting the need for accurate and interpretable prediction models. This study evaluates two Explainable AI (XAI) methods SHAP and LIME for interpreting machine learning models in CVD prediction. Using a public dataset, several models were tested, with AdaBoost achieving the best performance (AUC 91%, accuracy 85%). Feature importance was analyzed using SHAP and LIME, and their consistency was validated with Spearman Rank and Kendall’s Tau correlations. The results show that SHAP and LIME identify similar key features, but SHAP is more aligned with statistical patterns, demonstrating better reliability. SHAP also offered clearer, global interpretability, while LIME provided more varied, instance-specific insights. The study's main contribution lies in demonstrating SHAP’s superiority in delivering consistent, data-aligned explanations, making it a more effective tool for transparent and trustworthy CVD predictions. These findings support the integration of interpretable models in healthcare to enhance clinical decision-making and trust in AI systems.

Ähnliche Arbeiten