<![CDATA[Heart disease remains a major health problem and has encouraged the use of machine learning for data-driven prediction. This study investigates the effect of ANOVA F-score feature selection on the predictive performance and interpretability of XGBoost using the UCI Heart Disease Cleveland subset with 303 observations and 13 predictor features. Two scenarios were evaluated: XGBoost using all features and XGBoost using features selected by ANOVA F-score. Model evaluation employed nested stratified cross-validation with 10 outer folds and 5 inner folds, while hyperparameter optimization was performed using Optuna with 100 trials. XGBoost using all features achieved an accuracy of 0.8346 ± 0.0684, an F1-score of 0.8149 ± 0.0759, and a ROC-AUC of 0.9020 ± 0.0447. XGBoost with ANOVA F-score achieved an accuracy of 0.8149 ± 0.0532, an F1-score of 0.7948 ± 0.0527, and a ROC-AUC of 0.9033 ± 0.0381. Feature selection reduced the transformed input space from 22 to an average of 17.8 features, equivalent to a 19.09% reduction. The Wilcoxon signed-rank test showed no statistically significant performance differences at the 0.05 significance level. SHAP analysis identified cp, ca, thal, slope, and oldpeak as the dominant features in both models. Overall, ANOVA F-score produced a more compact feature representation while preserving the main interpretability pattern, but did not significantly improve XGBoost predictive performance.]]>
Copyrights © 2026
