ENHANCING HEART DISEASE PREDICTION PERFORMANCE WITH A SOFT VOTING ENSEMBLE AND ADDITIONAL TECHNIQUES

Abstract

Heart disease is a leading cause of mortality, affecting a significant number of people world wide. The pressing demand for diagnostic methods that deliver both superior effectiveness and accuracy is evident. Machine learning techniques, including deep tabular learning mod els, have been extensively applied to tabular healthcare data, demonstrating promising results in prediction and analysis. However, traditional machine learning models may suffer from limitations in accuracy, precision, and recall values. We propose a Soft voting meta-classifier comprising Catboost, Light-Gradient Boosting Machine, Gaussian Naive Bayes, Random Forest, and XGBoost to address these issues. Additionally, we explored deep tabular learning models TabNet and TabPFN. Our study was conducted on a fused dataset from UCI heart disease and Statlog sources. The proposed soft voting ensemble outperformed the individual models and achieved an accuracy of 91.85% and an AUC score of 0.9344, showcasing its potential for effective heart disease prediction.