Table 3.

Discrimination of ten machine-learning algorithms for MPR.

Model	Training AUC	Internal AUC	External AUC	5-fold CV AUC	External brier
Logistic regression	0.873	0.816	0.830	0.862	0.170
LASSO logistic	0.872	0.818	0.830	0.864	0.169
Elastic-net	0.872	0.819	0.830	0.864	0.170
AdaBoost	0.907	0.834	0.814	0.850	0.197
Random forest	0.984	0.856	0.802	0.850	0.185
k-nearest neighbors	0.894	0.801	0.798	0.846	0.185
XGBoost	0.991	0.845	0.792	0.838	0.200
Decision tree	0.906	0.832	0.781	0.840	0.204
LightGBM	0.984	0.837	0.781	0.828	0.206
Support vector machine	0.861	0.744	0.770	0.836	0.188

AUC, area under the ROC curve; CV, cross-validation. Models are ordered by external-validation AUC. Regularized logistic regression achieved the highest external AUC and was selected as the ENI-Score backbone; tree-based ensembles reached near-perfect training AUC but lost substantial AUC on external data, indicating overfitting on the compact five-marker space.