Head-to-head comparison of the three most commonly used preoperative models for prediction of biochemical recurrence after radical prostatectomy

Abstract

Background: Several models can predict the rate of biochemical recurrence (BCR) after radical prostatectomy (RP).

Objective: We tested the three most commonly used models-the D'Amico risk stratification scheme, the Cancer of the Prostate Risk Assessment (CAPRA) score, and the Stephenson nomogram-in a European cohort of RP patients.

Design, setting, and participants: We relied on preoperative characteristics and prostate-specific antigen follow-up data of 1976 patients, as required by the three tested models. All patients were treated with an open RP between 1992 and 2006.

Measurements: Analyses included tests of accuracy (Harrell's concordance index) and calibration between predicted and observed BCR rates at 3 yr and 5 yr after RP. Additionally, we relied on decision curve analyses to compare the three models directly in a head-to-head fashion.

Results and limitations: The median follow-up of censored patients was 32 mo. BCR-free rates at 3 yr and 5 yr after RP were 80.2% and 72.6%, respectively. The concordance index for 3-yr BCR predictions was 70.4%, 74.3%, and 75.2% for the D'Amico, CAPRA, and Stephenson models, respectively, versus 67.4%, 72.9%, and 73.5% for 5-yr BCR predictions. Calibration results supported the use of either the CAPRA or Stephenson models. Decision curve analyses indicated a small benefit for the CAPRA score relative to the Stephenson nomogram. Our findings apply to German patients treated with RP at a high-volume tertiary care centre. Consequently, the rank order reported in this paper may not be the same in North American or other European cohorts.

Conclusions: Different methods yield different results, and it may be difficult to reconcile concordance index, calibration, and decision curve analysis findings. Our data suggest that the CAPRA score outperforms the other models when decision curve analysis and calibration were used as benchmarks. Conversely, the Stephenson nomogram outperformed the other models when concordance index was used as a metric.