The findings of the current study suggest that AAM have higher rates of both BCR and adverse pathology after RP compared to CM as would be predicted by both the CAPRA and Kattan risk models. After adjusting for both the pre-operative and post-operative Kattan risk models, we observed a significantly higher rate of BCR among AAM. After adjusting for the CAPRA and CAPRA-S models, there was a trend towards higher rates of BCR in AAM, though this did not reach statistical significance. Further, the rate of adverse pathology was demonstrated to be significantly greater in AAM compared to CM after adjusting for CAPRA and Kattan risk models.
Although there have been multiple prediction models developed for prediction of BCR in men with localized prostate cancer, the D’Amico classification, CAPRA score, and Kattan nomogram hold a majority stake in clinical practice. In a recent national survey of Urologists and Radiation Oncologists, only 8.7% of physicians used a model other than D’Amico classification, CAPRA scores, or Kattan nomograms for counseling and clinical decision making in men with localized prostate cancer.2 The D’Amico risk model bins patients into 3 risk strata according to risk of BCR, whereas the CAPRA and Kattan models output a numeric risk score which we was more informative for the analysis performed in this study. Both the CAPRA score and Kattan nomograms do not include race as a model parameter for prediction of BCR; however, the training databases that were used to determine model parameters were comprised of a small proportion of AAM compared to CM. Even though these models were subsequently validated on patient cohorts with larger relative numbers of AAM, only a few studies have tested the predictive value of race in the determination of risk of BCR when applied to cohorts with greater proportions of AAM.13,14
In similar studies of the Shared Equal Access Regional Cancer Hospital (SEARCH) and Duke Prostate Cancer (DPC) datasets, the authors attempted to determine if there was less discriminatory accuracy of prostate cancer risk models in prediction of BCR in AAM compared to CM.13 After adjusting for the CAPRA and Kattan models, AAM were found to have a higher rate of BCR than CM in both the SEARCH and DPC patient cohorts, which was found to be at or approaching statistical significance for each model. Additionally, they found that the concordance indices were greater for AAM than CM in DPC, but lower in AAM than CM in SEARCH. Although the authors interpreted the conflicting c-indices as demonstrating an absence of systematic bias of model performance towards one race, these results may be better explained by intrinsic differences between the two databases, in which there were lower rates of BCR in the DPC cohort and higher rates of BCR in the SEARCH cohort. In a population with a low rate of BCR, model discriminatory accuracy would be expected to be greater for AAM who have higher rates of BCR, whereas in a population with high rates of BCR model discrimination would be expected to be lower for AAM.
In our analysis of a patient cohort from PURC, we demonstrate a consistently greater rate of BCR in AAM compared to CM for each model. Our results were found to be statistically significant after adjusting for the Kattan models, but not for CAPRA models. In our secondary analysis, we found AAM to have a 28% greater rate of adverse pathology after adjusting for the CAPRA model, and a 23% increased rate after adjusting for the Kattan model, both of which were found to be statistically significant. Notably, the rates of adverse pathology between AAM and CM without adjustment were similar. This finding may be due to the PURC patient composition comprising relatively equal proportions of AAM and CM with adverse pathology despite significant differences in their pre-operative risks, as a high level of statistical significance was met after adjustment for prostate cancer risk models.
In our analysis of adverse pathology, we observed statistically significant differences between CM and AAM after adjusting for both models, whereas our analysis of BCR only reached statistical significance when adjusting for the Kattan models. This difference may be explained by the larger cohort size used to predict adverse pathology compared to the cohort used to predict BCR. The BCR cohort had a larger number of patients excluded due to missing surveillance PSA data, decreasing the statistical power for this analysis. The overall trend towards BCR in this study demonstrates a higher risk in AAM compared to CM, despite statistical significance not being met in the CAPRA and CAPRA-S models.
We found greater differences in concordance indices between AAM and CM for the CAPRA score (AAM 0.715 vs CM 0.751) compared to the Kattan model score (AAM 0.733 vs CM 0.736). These results are in agreement with previous studies that showed a small, although potentially clinically insignificant result in comparing c-indices between separate cohorts for AAM and CM.13,14 We did not employ c-indices as a key point of comparison for this study because it has poor utility in comparing alternative models within the same dataset, as we performed in this analysis. Harrell’s concordance index is useful as an intuitive statistic in demonstrating global model discriminative accuracy; however, the c-index has no penalty for differences in discrimination between model certainty (i.e. 100 vs 51%). As a result, some results seem counter-intuitive; for example, in comparisons between the pre-operative Kattan models, the addition of race as a covariate decreases the c-index, although the likelihood ratio is, in fact, greater.
The limitations of this study include the use of a retrospective cohort and relying on self-reporting race with approximately 12% of this data missing in PURC. Additionally, our smaller patient cohort for the survival analysis in prediction of BCR was likely underpowered compared to our patient cohort used to assess rates of adverse pathology. Further, PURC represents a regional cohort and may not be generalizable across the general population. This study also does not distinguish whether the greater rates of adverse pathology and biochemical recurrence seen in AM compared to predicted models are related to underlying tumor biology and/or access to care. In a recent study analyzing prostatectomy rates in PURC, a significantly greater number of CM underwent RP compared to AM during the COVID pandemic, which demonstrates non-medical determinants of health may play a role.16