Answers are grounded in this paper and cite their sources.
Heart disease decision support often requires identifying a small set of routine clinical tests that preserves reliable multi-class predictive performance while reducing diagnostic burden. However, many metaheuristic wrapper studies emphasize accuracy and rely on implicit or post hoc size control, which can yield larger panels under comparable computational budgets. This work addresses this gap by proposing GA-CombinedFitness, a penalty-embedded genetic algorithm that optimizes a single combined objective of five-fold cross-validated logistic-regression accuracy and subset size under a primary configuration. The approach was evaluated on a harmonized multi-cohort dataset of 920 patients with eight routinely collected predictors and benchmarked against seven alternative search techniques under a controlled protocol with matched iteration and cross-validation settings. GA-CombinedFitness achieved the highest mean combined fitness ( CF = 0 . 4805 ) while selecting a single-feature subset ( cp , chest-pain type) with mean accuracy of 0.5478, mean balanced accuracy of 0.2981, and mean macro-F1 of 0.2506. Accuracy-driven methods achieved higher accuracy and macro-F1 on average but selected larger subsets, which reduced their combined-fitness scores under the stated objective. Post hoc SHAP, PDP+ICE, and ALE analyses were used as plausibility checks of model behavior under the evaluation protocol and of the predictors emphasized by the combined objective. These results indicate that embedding a parsimony penalty within genetic search can improve the accuracy–cost trade-off and yield compact, interpretable panels for cost-constrained heart disease workflows.
No takes yet. Share an insight, caveat, or question.
Stephan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: