Performance of prostate cancer prevention trial risk calculator in a contemporary cohort screened for prostate cancer and diagnosed by extended prostate biopsy.

PURPOSE

Statistical models such as the Prostate Cancer Prevention Trial risk calculator have been developed to estimate the cancer risk in an individual and help determine indications for biopsy. We assessed risk calculator performance in a large contemporary cohort of patients sampled by extended biopsy schemes.

MATERIALS AND METHODS

The validation cohort comprised 3,482 men who underwent a total of 4,515 prostate biopsies. Calculator performance was evaluated by ROC AUC and calibration plots. A multivariate regression model was fitted to address important predictor variables in the validation data set. Prediction error was calculated as the response variable in another multivariate regression model.

RESULTS

Using an average of 13 cores per biopsy prostate cancer was detected in 1,862 patients. The calculator showed an AUC of 0.57 to predict all cancers and 0.60 for high grade cancer. Multivariate analysis of the predictive ability of various clinical factors revealed that race and the number of biopsy cores did not predict overall or high grade cancer at biopsy. Prior negative biopsy, patient age and free prostate specific antigen were significantly associated with prediction error for overall and high grade cancer. Race and family history had a significant association with prediction error only for high grade disease.

CONCLUSIONS

To our knowledge our external validation of the Prostate Cancer Prevention Trial risk calculator was done in the largest cohort of men screened for prostate cancer to date. Results suggest that the current calculator remains predictive but does not maintain initial accuracy in contemporary patients sampled by more extensive biopsy schemes. Data suggest that the predictive ability of the calculator in current clinical practice may be improved by modeling contemporary data and/or incorporating additional prognostic variables.