Heart Failure Prediction: A Comparative Study of SHAP, LIME, and ICE in Machine Learning Models

Heart disease remains a critical public health issue, prompting the need for effective predictive modeling. This study evaluates the performance of LightGBM, SVM, Random Forest, and Logistic Regression models on a heart disease dataset. Logistic Regression achieved the highest accuracy of 86.89%, demonstrating strong performance in classification with balanced precision and recall. LightGBM and Random Forest also performed competitively, with accuracies of 85.33% and 85.25%, respectively. Notably, Random Forest had the highest recall (96.97%) but lower precision (80%). SVM showed strong recall at 93.94% but had the lowest accuracy (83.61%). The findings underscore the importance of model interpretability, facilitated by SHAP, LIME, and ICE, which enhance understanding of model decisions in healthcare applications, ultimately supporting improved clinical outcomes.