Answers are grounded in this paper and cite their sources.
Class imbalance may distort performance estimates in clinical prediction models for myofascial pelvic pain syndrome (MPPS). To compare class weighting, categorical-aware synthetic oversampling, and constrained generative augmentation for contemporaneous classification of MPPS. The retrospective single-center dataset contained 2016 records (568 MPPS and 1448 health-examination controls) and 23 predictors. Exact duplicate predictor signatures were kept in the same partition, producing a development set of 1409 records (397 MPPS) and an untouched test set of 607 records (171 MPPS). Logistic Regression and XGBoost were evaluated with original data, class weighting, SMOTENC, and a constrained tabular GAN. Five-by-three nested cross-validation, calibration, and threshold selection were confined to the development set. Both augmentation methods balanced each fold-training subset to 1:1. Constrained-GAN XGBoost had a test ROC-AUC of 0.852 (95% CI 0.813–0.886), compared with 0.846 (0.808–0.884) for class-weighted XGBoost; the absolute difference was 0.006 (Holm-adjusted p = 0.002). Their PR-AUC estimates were 0.749 (0.694–0.801) and 0.747 (0.686–0.808), respectively. The GAN pipeline had a Brier score of 0.128 (0.111–0.144) versus 0.130 (0.116–0.144), whereas the calibration intercept was closer to zero for class weighting (− 0.013, − 0.171 to 0.172) than for GAN augmentation (− 0.174, − 0.319 to -0.006). Across five algorithmic seeds, GAN-XGBoost had slightly higher mean ROC-AUC (0.856 versus 0.849), while class-weighted XGBoost had slightly higher mean accuracy (0.779 versus 0.776). XGBoost outperformed Logistic Regression. GAN augmentation produced slightly higher discrimination point estimates, but class weighting was comparable across calibration, decision-curve, and threshold-dependent measures. These internal results do not establish overall clinical superiority and require external validation in symptomatic populations.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: