Using machine learning for predicting cancer-specific mortality in bladder cancer patients undergoing radical cystectomy: a SEER-based study.

Background: Accurately assessing the prognosis of bladder cancer patients after radical cystectomy has important clinical and research implications. Current models, based on traditional statistical approaches and complex variables, have limited performance. We aimed to develop a machine learning (ML)-based prognostic model to predict 5-year cancer-specific mortality (CSM) in bladder cancer patients undergoing radical cystectomy, and compare its performance with current validated models.

Methods: Patients were selected from the Surveillance, Epidemiology, and End Results database and the First Affiliated Hospital of Sun Yat-sen University for model construction and validation. We used univariate and multivariate Cox regression to select variables with independent prognostic significance for inclusion in the model's construction. Six ML algorithms and Cox proportional hazards regression were used to construct prediction models. Concordance index (C-index) and Brier scores were used to compare the discrimination and calibration of these models. The Shapley additive explanation method was used to explain the best-performing model. Finally, we compared this model with three existing prognostic models in urothelial carcinoma patients using C-index, area under the receiver operating characteristic curve (AUC), Brier scores, calibration curves, and decision curve analysis (DCA).

Results: This study included 8,380 patients, with 6,656 in the training set, 1,664 in the internal validation set, and 60 in the external validation set. Eight features were ultimately identified to build models. The Light Gradient Boosting Machine (LightGBM) model showed the best performance in predicting 5-year CSM in bladder cancer patients undergoing radical cystectomy (internal validation: C-index = 0.723, Brier score = 0.191; external validation: C-index = 0.791, Brier score = 0.134). The lymph node density and tumor stage have the most significant impact on the prediction. In comparison with current validated models, our model also demonstrated the best discrimination and calibration (internal validation: C-index = 0.718, AUC = 0.779, Brier score = 0.191; external validation: C-index = 0.789, AUC = 0.884, Brier score = 0.137). Finally, calibration curves and DCA exhibited better predictive performance as well.

Conclusions: We successfully developed an explainable ML model for predicting 5-year CSM after radical cystectomy in bladder cancer patients, and it demonstrated better performance compared to existing models.