Competitive Risk Model for Specific Mortality Prediction in Patients with Bladder Cancer: A Population-Based Cohort Study with Machine Learning

Abstract

Background. Noncancer death accounts for a high proportion of all patients with bladder cancer, while these patients are often excluded from the survival analysis, which increases the selection bias of the study subjects in the prediction model. Methods. Clinicopathological information of bladder cancer patients was retrieved from the Surveillance, Epidemiology, and End Results (SEER) database, and the patients were categorized at random into the training and validation cohorts. The random forest method was used to calculate the importance of clinical variables in the training cohort. Multivariate and univariate analyses were undertaken to assess the risk indicators, and the prediction nomogram based on the competitive risk model was constructed. The model’s performance was evaluated utilizing the calibration curve, consistency index (C index), and the area under the receiver operator characteristic curve (AUC). Results. In total, we enrolled 39285 bladder cancer patients in the study (27500 patients were allotted to the training cohort, whereas 11785 were allotted to the validation cohort). A competitive risk model was constructed to predict bladder cancer-specific mortality. The overall C index of patients in the training cohort was 0.876, and the AUC values were 0.891, 0.871, and 0.853, correspondingly, for 1-, 3-, and 5-year cancer-specific mortality. On the other hand, the overall C index of patients in the validation cohort was 0.877, and the AUC values were 0.894, 0.870, and 0.847 for 1-, 3-, and 5-year correspondingly, suggesting a remarkable predictive performance of the model. Conclusions. The competitive risk model proved to be of great accuracy and reliability and could help clinical decision-makers improve their management and approaches for managing bladder cancer patients.