Occult central lymph node metastasis risk assessment
Compare the selected stacking ensemble with XGBoost using the same nine preoperative predictors. The tool is designed for the preoperative decision point after cN0 status has been established.
The manuscript and model are undergoing peer review. Public use is planned only after formal publication and release of the validated model version.
Choose a model
Review each model’s retrospective discrimination before viewing the shared inputs.
Stacking ensemble
Selected for its highest internal test AUC (0.913). Stacking combines 11 base learners with a regularized logistic meta-learner. Its multiple training and prediction stages make implementation and maintenance more complex than XGBoost.
Internal test
0.91395% CI 0.883–0.939Temporal validation
0.89495% CI 0.862–0.920Geographical validation
0.89695% CI 0.864–0.924| Model | Internal test | Temporal | Geographical | Architecture and complexity |
|---|---|---|---|---|
| Stacking | 0.913 | 0.894 | 0.896 | 11 base learners + logistic meta-learner; more fitting stages and model components |
| XGBoost | 0.911 | 0.901 | 0.894 | One boosted-tree algorithm; fewer model components and a simpler deployment workflow |
Performance values are AUCs. The geographical AUC difference was 0.002 (stacking minus XGBoost; P=0.703). Complexity refers to model architecture and implementation. Prediction remains locked during peer review.
Model inputs — Stacking ensemble
Input definitions are shown for transparency. Data entry and prediction are disabled during peer review.