The disease registry
CaPSURE™ (Cancer of the Prostate Strategic Urologic Research Endeavor) is a longitudinal, observational database of men with biopsy-proven prostate adenocarcinoma, recruited from 40 primarily community-based urology practices across the United States. Newly-diagnosed prostate cancer patients are recruited consecutively by participating urologists, who report complete clinical data and follow-up information on diagnostic tests and treatments. Informed consent is obtained from each patient under institutional review board supervision. Patients are treated according to their physicians' usual practices, and are followed until time of death or withdrawal from the study.8, 9
The prostate specific antigen (PSA) value used was the highest PSA value recorded in the nine months prior to diagnosis. 2002 clinical TNM stage was the highest reported from 1 month prior to 3 months after the date of diagnosis. Gleason scores were recorded from the diagnostic biopsy site with the highest total and highest primary scores. Percent positive biopsies (PPB) was calculated from detailed reported biopsy data. Disease recurrence after radical prostatectomy (RP) was defined as two consecutive PSA values ≥0.2 ng/ml at any time postoperatively or any additional treatment more than six months after RP. The date of recurrence was defined as the earlier of the second PSA ≥0.2 ng/ml or the date additional treatment was initiated. If disease recurrence did not occur, the patient's follow-up time was censored at the date of the last recorded PSA.
As of July 2003, 10,018 patients were enrolled in CaPSURE. 4128 of these elected RP as primary treatment for their prostate cancer. We included patients diagnosed between 1992 and 2001 with clinically localized disease (clinical stage T1c-3a, N0/x, M0/x) who did not receive neoadjuvant or adjuvant radiation or hormonal therapy (N=2154). We excluded patients with unknown PSA, Gleason score, clinical T-stage, or PPB. We also limited the analysis to patients with at least a sextant biopsy, PSA ≥2 ng/ml at diagnosis, and at least two follow-up PSAs or evidence of additional treatment more than 6 months after RP. 1439 patients meeting these criteria constituted our analytic dataset.
Development of the UCSF Cancer of the Prostate Risk Assessment (UCSF-CAPRA)
Our goal in developing this predictive index was to maximize the ability of the score to predict disease-free survival (DFS) while maintaining the simplicity and clinical applicability of the tool. The variables initially considered for inclusion in the index included PSA, Gleason score, T stage, PPB, age at diagnosis, and ethnicity. We began by including all of these variables in a Cox proportional hazards model with detailed categories (PSA as 2-4, 4.1-6, 6.1-8, 8.1-10, 10.1-20, 20.1-30, >30; Gleason as 1-2/1-2, 1-2/3, 3/1-2, 3/3, 1-3/4-5, 4-5/1-3, 4-5/4-5; T-stage as T1c, T2a, T2b, T2c, T3a; PPB as <15%, 15-25%, 26-33%, 34-50%, 51-66%, 67-79%, ≥80%; age as <50, 50-54, 55-59, 60-64, 65-69, ≥70; and ethnicity as African-American, Caucasian, other).
The results of the initial model were reviewed to determine whether any variables could be eliminated and which category levels could be collapsed. Ethnicity was not a significant predictor of DFS [hazard ratio (HR) for African-American ethnicity 1.16, p=.53, HR for other/unknown ethnicity 0.31, p=.10]. Further refinement of the model was accomplished primarily through an examination of the model's parameter estimates (PEs): category levels within each variable with similar PEs could be combined to reduce the model. This process was repeated iteratively until the final model was reached.
We also built models incorporating various mathematical transformations of PSA, including linear, logarithmic, truncated, sigmoidal, cubic spline, and piecewise linear. There was in fact a slight improvement in the performance of the model using the piecewise linear PSA function rather than categorized PSA levels. However, the relatively small gain in accuracy would not be worth the large loss in simplicity and ease of use; therefore categorized PSA was retained in the final model. Finally, in addition to the various categorizations of PPB, we examined the percent of biopsy cores positive from the more involved side of the prostate only, as well as the absolute number—rather than percentage—of cores positive. These variables, however, did not provide any additional predictive information.
Once the final model was specified, the PEs were used to assign points for each level of the variables in the model. We decided that each 2-point increase in the final index should represent approximately a doubling of risk for the outcome of disease recurrence. Proportional hazards model PEs are calculated on the log scale; thus a PE of 0.7 would result in a doubling of risk, and each 0.35 increase in PE would be worth 1 point. Points were thus assigned to each level of the final variables in the index. The final CAPRA score for each patient was calculated by summing the points for each variable in the model.
Predictive performance of the CAPRA
CAPRA scores were calculated for the men in the analytic population and were included in Cox proportional hazards regression models, with HR calculated for each CAPRA score. Life table and Kaplan-Meier analysis were used to determine the probability of DFS at 3 and 5 years for each CAPRA score level.
We calculated Kattan nomogram scores7
for each man in the dataset, and classified each according to a modification of D'Amico et al's risk groupings.6
A patient with PSA <10 ng/ml, no Gleason pattern 4 or 5 disease, and a clinical stage of T1 or T2a was low risk. Intermediate risk patients were those with PSA 10.1-20 ng/ml, Gleason 7 or <7 with 4 or 5 as the secondary pattern, or clinical stage T2b-c. High risk patients had PSA >20 ng/ml, Gleason >7 or 4 or 5 as the primary pattern, or clinical stage T3a.
Relationships among the CAPRA, Kattan, and D'Amico scores were assessed by Pearson correlation coefficients (r), frequency analysis of D'Amico categories by CAPRA level, and mean Kattan nomogram score per CAPRA level. We also calculated the concordance index (c-index) for each algorithm. The c-index in survival analysis is the proportion of randomly paired patients for whom the patient with the higher probability of recurrence (higher CAPRA score, higher D'Amico risk category, lower Kattan nomogram score) also had the earlier observed disease recurrence. The concordance index ranges from 0-1, with 1 indicating perfect concordance and 0.5 indicating no concordance.
All analyses were performed using Statistical Analysis Software (SAS) version 8.2, except for the c-index, which was calculated using S-Plus version 6.0.