Validation of a genomic classifier for prediction of metastasis and prostate cancer-specific mortality in African-American men following radical prostatectomy in an equal access healthcare setting

Background The Decipher 22-gene genomic classifier (GC) may help in post-radical prostatectomy (RP) decision making given its superior prognostic performance over clinicopathologic variables alone. However, most studies evaluating the GC have had a modest representation of African-American men (AAM). We evaluated the GC within a large Veteran Affairs cohort and compared its performance to CAPRA-S for predicting outcomes in AAM and non-AAM after RP. Methods GC scores were generated for 548 prostate cancer (PC) patients, who underwent RP at the Durham Veteran Affairs Medical Center between 1989 and 2016. This was a clinically high-risk cohort and was selected to have either pT3a, positive margins, seminal vesicle invasion, or received post-RP radiotherapy. Multivariable Cox models and survival C-indices were used to compare the performance of GC and CAPRA-S for predicting the risk of metastasis and PC-specific mortality (PCSM). Results Median follow-up was 9 years, during which 37 developed metastasis and 20 died from PC. Overall, 55% (n = 301) of patients were AAM. In multivariable analyses, GC (high vs. intermediate and intermediate vs. low) was a significant predictor of metastasis in all men (allp < 0.001). Consistent with prior studies, relative to CAPRA-S, GC had a higher C-index for 5-year metastasis (0.78 vs. 0.72) and 10-year PCSM (0.85 vs. 0.81). There was a suggestion GC was a stronger predictor in AAM than non-AAM. Specifically, the 5-year metastasis risk C-index was 0.86 in AAM vs. 0.69 in non-AAM and the 10-year PCSM risk C-index was 0.91 in AAM vs. 0.78 in non-AAM. However, the test for interaction of race and the performance of the GC in the Cox model was not significant for either metastasis or PCSM (bothp >= 0.3). Conclusions GC was a very strong predictor of poor outcome and performed well in both AAM and non-AAM. Our data support the use of GC for risk stratification in AAM post-RP. While our data suggest that GC may actually work better in AAM, given the limited number of events, further validation is needed.