Evaluating Model Complexity in a Small Single-Center Clinical Tabular Dataset: A Leakage-Controlled and Calibration-Aware Case Study of MINOCA–Coronary Slow-Flow Differentiation

This study evaluated whether model complexity adds value when differentiating myocardial infarction with non-obstructive coronary arteries (MINOCA) from coronary slow flow in a small clinical tabular dataset. The retrospective cohort included 107 coronary slow-flow and 104 MINOCA patients represented by 19 variables obtained before angiography and definitive group assignment. Penalized logistic regression was compared with kernel- and tree-based machine-learning models. Imputation, scaling, hyperparameter selection, and calibration were restricted to training folds within 5-fold nested cross-validation repeated 10 times. Discrimination, probabilistic accuracy, calibration, threshold-dependent metrics, learning curves, estimator dynamics, and paired uncertainty were evaluated. Random forest achieved the highest AUROC (0.825; 95% CI, 0.774–0.873), whereas elastic net achieved an AUROC of 0.810 (0.753–0.863). The paired random-forest minus elastic-net AUROC difference was 0.015 (−0.027 to 0.060), and the Brier-score difference was −0.0068 (−0.0291 to 0.0138); no superiority was demonstrated after Holm correction. Learning curves showed continued validation gains for the tree ensembles as training size increased, with persistent training–validation separation. Feature-group analyses indicated model-specific contributions from hemogram and lipid variables. The findings do not establish equivalence or generalizability to all small clinical datasets. They support leakage-controlled, calibration-aware, and uncertainty-aware benchmarking before any consideration of external validation or clinical use.