Clinical Takeaway Reported discrimination alone is insufficient to determine whether an AI model is clinically reliable or transferable. Differences in outcome definitions, patient populations, input features, and validation strategies can substantially affect model performance across institutions, while algorithm suitability may also depend on the target phenotype and local data structure. Importantly, our review identified substantial gaps in evaluation beyond discrimination: quantitative calibration assessment, prevalence-sensitive measures such as AUPRC, and decision curve analysis were inconsistently reported, and prospective evaluation of the clinical impact of model deployment was absent. Class imbalance was addressed in only approximately half of the included publications, and insufficient reporting of resampling procedures may leave information leakage difficult to exclude. Therefore, evaluation of PPRC models should combine discrimination with calibration and clinical utility, use metrics appropriate to outcome prevalence, ensure leakage-safe model development and validation, and consider explainability and feature feasibility when assessing transferability to a target clinical environment.