TABLE 4.

Comparative performance of three validation schemes for the High-risk PARS logistic classifier. Accuracy, precision, recall, and F1 are reported at the fixed clinical threshold of 0.30. Primary analysis uses patient-level 80–20 split stratified by High-risk status. Metrics for precision, recall, and F1 refer to the positive class (High-risk PARS = 1).

Model	Split strategy	AUC	Accuracy	Precision (1)	Recall (1)	F1 (1)
80–20, strat high-risk only	Train/test 80–20 at patient level; stratified by patient indicator of having $\ge$ 1 high-risk PARS visit	0.931	0.908	0.795	0.808	0.802
80–20, strat age + sex + High-risk	Train/test 80–20; combined stratification by age group ( $<$ 65/ $\ge$ 65), sex (F/M), and high-risk PARS (yes/no)	0.938	0.878	0.754	0.781	0.767
5-Fold GroupKFold	5-Fold cross-validation grouped by patient (GroupKFold); all visits of a patient are confined to a single fold	0.939	0.883	0.749	0.818	0.781