The Cancer of the Bladder Risk Assessment (COBRA) score for estimating cancer-specific survival after radical cystectomy: external validation in a large bi-institutional cohort

Objective: To perform an external validation of the Cancer of the Bladder Risk Assessment (COBRA) score for estimating cancer-specific survival (CSS) after radical cystectomy (RC) in a large bi-institutional cohort of patients. Patients and Methods: Patients treated with RC and lymph node dissection (LND) between May 1996 and July 2017 were retrieved from the RC databases of Leuven and Turin. Collected variables were age at RC, tumour stage, lymph node (LN) density, neoadjuvant chemotherapy, the extent of LND, and nodal stage. The primary outcome was CSS visualised using Kaplan–Meier plots. Cox proportional hazard models were used to assess the impact of variables on CSS. We performed a pairwise comparison between the COBRA score levels using a log-rank test corrected by Bonferroni, and developed a simplified COBRA score with three risk categories. To compare models, we assessed concordance indices (C-indices), receiver operating characteristic curves with area under the curve (AUC), calibration plots, and decision curve analysis (DCA). Finally, we compared both COBRA and simplified COBRA models with the established American Joint Committee on Cancer (AJCC) model. Results: A total of 812 patients were included. All COBRA score variables had a significant impact on CSS in a Cox proportional hazard model. However, pairwise comparison of the COBRA subscores could not differentiate significantly between all COBRA score levels. Based on these findings, we developed a simplified COBRA score by introducing three categories within the following COBRA score ranges: low- (0–1) vs intermediate- (2–4) vs high-risk (5–7). A pairwise comparison could discriminate significantly between all COBRA risk categories. When finally comparing COBRA and simplified COBRA models with the AJCC model, AJCC performed better than both. C-indices, AUCs, calibration plots and DCA for AJCC were all better compared with the original and simplified COBRA models. Conclusion: We performed an external validation of the COBRA score in a large bi-institutional cohort. We observed that several risk groups had overlapping CSS, demonstrating suboptimal performance of the COBRA score. Therefore, we constructed a simplified model with three COBRA score risk categories. This model resulted in demarcated risk groups with non-overlapping CSS and good predictive accuracy. However, both COBRA score models were outperformed by the AJCC staging system. Therefore, we conclude that the AJCC staging system should remain the current standard for stratifying patients after RC for CSS.