Related Experiment Video
Updated: Jul 17, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Predicting overall survival in synchronous metastatic nasopharyngeal carcinoma using a stacking ensemble machine
Canwen Che1,2, Shanyue Lin3, Jie Ma4
1Shantou University Medical College, Shantou, China.
Background:
Patients with synchronous metastatic nasopharyngeal carcinoma (smNPC) show marked survival heterogeneity, which complicates individualized prognostic assessment and treatment planning. We developed and validated a stacking ensemble machine learning model for predicting 3-year overall survival (OS) in this population.
Methods:
This multicenter retrospective study included 413 patients from three institutions, randomly assigned to a training set (n = 289) and an internal held-out validation set (n = 124). Candidate predictors included demographic, histopathologic, serological, and treatment-related variables. The 3-year OS label was assigned only when death occurred within 3 years or when follow-up confirmed survival at 3 years. Six base learners were trained: gradient boosting decision tree (GBDT), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), categorical boosting (CatBoost), and random forest (RF). Hard voting, soft voting, and stacking ensembles were constructed and compared. Discrimination, calibration, decision curve analysis, leave-one-center-out (LOCO) validation, overfitting diagnostics, and Shapley Additive Explanations (SHAP) were assessed.
Results:
The stacking ensemble achieved the highest internal held-out validation AUC of 0.8358 for 3-year OS prediction. The 3-year death event rate was 60.0% overall, 59.9% in the training set, and 60.5% in the internal held-out validation set. After multiplicity adjustment, stacking showed significantly higher AUC than GBDT, AdaBoost, and hard voting, whereas differences versus other comparators were not significant. Overfitting diagnostics showed a training AUC of 0.9150, an optimism gap of 0.0792, and an optimism-corrected AUC of 0.8415. SHAP identified immunotherapy, first-line regimen, number of metastatic lesions, and number of metastatic organs as influential predictors.
Conclusion:
The stacking ensemble model provided an interpretable approach for individualized 3-year OS prediction in smNPC and may support multidisciplinary prognostic assessment, although prospective validation remains needed.
