A Comparative Analysis of Feature Selection Methods on the Accuracy of Heart Disease Prediction Models
Sultanus Salehin, Akib Jayed Islam, Azmain Mahpara Iqbal, Labannya Barua, M. Kamrul Islam, Arafat Uddin
Abstract
Heart disease is a leading cause of death globally, making accurate diagnostic methods crucial. This study explores the use of machine learning, particularly artificial intelligence, for heart disease prediction, addressing challenges posed by the high dimensionality of medical datasets. By comparing feature selection techniques, including Wrapper and Filter methods and expert-driven approaches, the study identifies the most relevant features for prediction. It uses two widely recognized datasets: the Cleveland Heart Disease dataset and the Framingham Heart Study dataset. These datasets, while both aimed at heart disease prediction, differ in their feature sets and target populations. The goal is to reduce the number of features while maintaining prediction accuracy. Various machine learning classifiers—SVM, Naïve Bayes (NB), Decision Tree (DT), K-Nearest Neighbor (KNN), Random Forest (RF), and Adaptive Boosting (AdaBoost)—were applied to both the full and reduced feature sets. Feature selection methods such as Sequential Forward and Backward Selection, Chi-Square, and Mutual Information were tested for effectiveness. The study shows that feature reduction improves classification accuracy and reduces training time. The comparison of the Cleveland and Framingham datasets reveals how each dataset's characteristics impact the performance of feature selection methods and heart disease prediction, highlighting the potential for generalizing across different datasets.