Feature selection using optimization algorithms in heart disease detection

Heart disease continues to be among the leading causes of mortality worldwide, necessitating its early detection for improving patient prognosis. Although conventional statistics-based algorithms and feature selection approaches have been applied for heart disease prediction purposes, the complex nature of medical data may not necessarily be accounted for using such techniques. Removing unnecessary features improves how well machine learning algorithms perform. In this study, we present a step-by-step feature selection methodology leveraging Genetic Algorithms (GA) and the Gini index. Utilizing greedy search for parameter tuning and and a comprehensive suite of confusion matrix metrics for evaluation, we compared six architectures: MLP, TabNet, Random Forest, SVM, Logistic Regression, and XGBoost are applied to the heart disease dataset. Our results demonstrate that the Random Forest (RF) model delivers superior diagnostics. By reducing the feature space from 13 to 8, the RF model attained exceptional predictive power, charting 98.4% accuracy, 98.9% sensitivity, 98% specificity, and a 98.5% F1-score.