<![CDATA[Feature selection is often reported to improve clinical prediction, yet optimization and evaluation on the same data can produce optimistic estimates. This study provides a leakage-controlled and reproducible reassessment of genetic algorithm (GA) feature selection for Gaussian Naive Bayes prediction of heart disease. The open-Heart Failure Prediction dataset contains 918 observations, 11 predictors, and a binary target. Clinically implausible zero values in resting blood pressure and cholesterol were recoded as missing; imputation, scaling, one-hot encoding, GA selection, and model fitting were confined to training data. Performance was estimated using repeated nested stratified cross-validation with five outer folds repeated five times and four inner folds for GA fitness. The GA used an 11-bit chromosome, population size 12, 10 generations, 0.80 uniform-crossover probability, 0.08 bit-flip mutation probability, tournament selection of size three, and two elites. Across 25 held-out outer folds, baseline Naive Bayes achieved 84.66% accuracy (SD 2.01%), whereas GA-selected Naive Bayes achieved 84.31% (SD 2.40%). The paired mean difference was −0.35 percentage points, with a bootstrap 95% confidence interval of −0.72 to 0.05 percentage points and a Wilcoxon p-value of 0.120. AUC values were 0.912 and 0.907, respectively. The GA retained 8.44 of 11 predictors on average; ExerciseAngina, FastingBS, ST_Slope, and Sex were selected in every outer fold. These results do not support a material accuracy gain from GA selection, but demonstrate modest dimensionality reduction and reveal predictors that are stable under resampling. The study emphasizes nested evaluation, uncertainty reporting, and full parameter disclosure for credible optimization claims]]>
Copyrights © 2026
