Cross-Site Validation

Nested cross-validation with leave-one-site-out validation is presented as a strength because it provides repeated external validation and less biased performance estimates. The protocol uses nested cross-validation to evaluate the complete modelling process. The outer loop estimates performance on held-out external sites using leave-one-site-out cross-validation. Models are trained on four sites and validated on the remaining site, repeated until every site has been held out once. The inner loop handles feature selection, hyperparameter tuning, threshold tuning, and model selection.