Development and validation of a machine learning-based model for predicting survival in patients with bladder neuroendocrine carcinoma: a large retrospective study

Background Bladder neuroendocrine carcinoma (BNEC) is a rare yet highly aggressive malignancy with an extremely poor prognosis. The current TNM staging system provides limited prognostic information and validated prognostic prediction tools are lacking. This study aimed to develop and validate a machine learning (ML)-based model for individualised survival prediction in BNEC patients using large-scale population data. Methods Data from pathologically confirmed BNEC patients from 2004 to 2021 were extracted from the US Surveillance, Epidemiology, and End Results (SEER) database. Univariate and multivariate Cox regression analyses were performed to identify independent prognostic factors. On the basis of the independent prognostic factors, eight machine learning models, namely, random forest (RF), gradient boosting machine (GBM), LASSO-Cox, CoxBoost, support vector machine (SVM), XGBoost, SuperPC, and plsRcox, were constructed using the training set (70% of the total sample). Model performance was comprehensively evaluated on the validation set (30% of total samples) using the C-index, time-dependent receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Results A total of 2117 patients with BNEC were included in the final analysis. Multivariate analysis revealed age >74 years (HR = 1.51, p<0.001), male sex (HR = 1.17, p=0.038), advanced T stage (T2–T4), positive nodal status (N1–N2), distant metastasis (M1, HR = 1.99, p<0.001), surgery (TURBT: HR = 0.75, P = 0.017; RC: HR = 0.36, P<0.001), radiotherapy (HR = 0.79, p=0.002), and chemotherapy (HR = 0.43, p<0.001) as independent prognostic factors. Among the eight machine learning models, the GBM algorithm demonstrated the best performance in both the training and validation sets, with AUC values for 1-year, 3-year, and 5-year survival prediction being 0.846/0.737, 0.825/0.730, and 0.816/0.742, respectively. Decision curve analysis (DCA) demonstrated greater clinical net benefit across a broad range of threshold probabilities, while the calibration curves indicated high agreement between the predicted and observed outcomes. Risk stratification based on the GBM-derived risk score revealed significant differences in survival between the high-risk and low-risk groups (p < 0.0001). Conclusion The GBM-based machine learning prognostic model demonstrated comparatively favorable performance in predicting survival outcomes among patients with BNEC. This model may provide a useful tool for individualized prognostic assessment and risk stratification, although further validation in multicentre prospective cohorts is warranted.