COMPARATIVE ANALYSIS OF PERFORMANCE OF MACHINE LEARNING FEATURE SELECTION (GINI DECREASE AND RELIEF-F) IN HEART DISEASE DATASET

Abstract

Heart disease remains one of the leading causes of mortality worldwide and presents a major challenge in healthcare systems. Early detection plays an essential role in improving survival rates and minimizing complications through timely intervention. Recent advances in Machine Learning (ML) have provided new opportunities for developing accurate and efficient prediction systems for heart disease detection. However, one of the major challenges in ML-based prediction is identifying the most relevant features to improve classification performance while reducing computational complexity and noise. This study aims to evaluate the effectiveness of two feature selection techniques, namely Gini Decrease (GD) and ReliefF, combined with several ML models, including Support Vector Machine (SVM), Tree, Naïve Bayes, and Random Forest, for heart disease classification. The study employed the UCI Heart Disease Dataset consisting of 303 records and 14 attributes. Data preprocessing included handling missing values using mean imputation, followed by feature selection and classification using 10-fold cross-validation with an 80:20 training-testing ratio. Experimental results showed that ReliefF outperformed GD, achieving the highest average accuracy of 0.796, compared to GD with 0.767 and all features with 0.771. The SVM model achieved the highest accuracy using GD (0.833), while Random Forest demonstrated optimal performance with ReliefF (0.817). Furthermore, the Tree model exhibited the fastest computational time among all evaluated models. These findings indicate that integrating suitable feature selection methods with ML models significantly enhances heart disease classification performance, particularly in improving predictive accuracy and computational efficiency for early medical diagnosis applications.