Interpretable machine learning prognostication of gastroenteropancreatic neuroendocrine tumors across Chinese and United States cohorts

Abstract

Background Gastroenteropancreatic neuroendocrine tumors (GEP-NETs) display marked clinical heterogeneity, and conventional prognostic indicators such as TNM stage and Ki-67 index provide limited individualized risk discrimination. We aimed to develop and externally validate interpretable machine learning models for survival prediction. Methods This retrospective study enrolled 337 patients with histologically confirmed GEP-NETs, randomly partitioned into a training set (n = 236) and an independent test set (n = 101) with stratification by mortality. Six survival models, including unpenalized Cox regression as a baseline, LASSO Cox, random survival forest (RSF), gradient-boosted survival analysis (GBSA), Extra Trees, and DeepSurv, were trained with 5×5 nested cross-validation. Discrimination, probabilistic accuracy, calibration, and clinical utility were comprehensively evaluated. SHAP analysis identified key predictors and informed nomogram construction. External validation was performed in 51,225 SEER patients. Results On the test set, Extra Trees achieved the highest discrimination (C-index 0.843, 95% CI 0.762-0.914) and balanced calibration (1-year ICI 0.031). Time-dependent AUC values were 0.871, 0.847 and 0.935 at 1, 3 and 5 years. The Extra Trees-based stratification identified high-risk patients with substantially worse survival (HR 8.93, 95% CI 3.35-23.84, P<0.001). The model outperformed TNM stage (C-index 0.772), Ki-67 (0.705) and tumor grade (0.699). Decision curve analysis demonstrated greater net benefit across threshold probabilities of 5%-50%. SHAP analysis identified T stage, tumor grade, M1 status, TNM stage and Ki-67 as the top predictors. External validation in SEER preserved discriminative ability (C-index 0.719) and excellent calibration (slopes 0.901-1.058). Conclusion Interpretable machine learning models outperform conventional approaches in predicting GEP-NET survival. Extra Trees showed the best internal discrimination and calibration, whereas DeepSurv achieved the highest external C-index. The simplified SHAP-derived nomogram provides a practical and well-calibrated but exploratory tool for individualized prognosis that requires prospective validation before clinical use.