Answers are grounded in this paper and cite their sources.
Imbalanced class distributions can limit minority-outcome detection in healthcare prediction. This study investigates a threshold-aware Bayesian optimization protocol integrating preprocessing, feature-type-specific resampling, and fold-wise decision-threshold selection. Each candidate is assessed on validation data using a threshold selected on a separate calibration subset. The contribution is the specification and empirical assessment of this integrated procedure using established learning components. Experiments on five healthcare datasets compare Logistic Regression, Random Forest, XGBoost, and CatBoost under nine configurations, including class weighting, post-selection threshold adjustment, and F 1 - versus AUPRC-guided search. Across 20 dataset–classifier combinations, threshold-aware F 1 optimization without resampling achieved a mean positive-class F 1 of 0.5266, compared with 0.4233 for its fixed-threshold baseline and 0.5101 for post-selection threshold adjustment. AUPRC-guided search with threshold selection achieved 0.5336, so the results do not establish superiority of the F 1 search objective. Resampling and class weighting provided configuration-dependent benefits. Conditional sensitivity analyses identified useful thresholds below the original grid boundary, while probability calibration improved Brier scores without consistently improving F 1 . No pairwise comparison was significant after Holm correction across five dataset blocks per classifier. The findings support examining the operating threshold alongside model selection, with conclusions restricted to the evaluated datasets and single outer split.
No takes yet. Share an insight, caveat, or question.
Mekhaldi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: