Predicting High-Grade Cancer at Ten-Core Prostate Biopsy Using Four Kallikrein Markers Measured in Blood in the ProtecT Study

Background: Many men with elevated prostate-specific antigen (PSA) levels in serum do not have aggressive prostate cancer and undergo unnecessary biopsy. Retrospective studies using cryopreserved serum suggest that four kallikrein markers can predict biopsy outcome. Methods: Free, intact and total PSA, and kallikrein-related peptidase 2 were measured in cryopreserved blood from 6129 men with elevated PSA (≥3.0ng/mL) participating in the prospective, randomized trial Prostate Testing for Cancer and Treatment. Marker levels from 4765 men providing anticoagulated plasma were incorporated into statistical models to predict any-grade and high-grade (Gleason score ≥7) prostate cancer at 10-core biopsy. The models were corrected for optimism by 10-fold cross validation and independently validated using markers measured in serum from 1364 men. All statistical tests were two-sided. Results: The four kallikreins enhanced prostate cancer detection compared with PSA and age alone. Area under the curve (AUC) for the four kallikreins was 0.719 (95% confidence interval [CI] = 0.704 to 0.734) vs 0.634 (95% CI = 0.617 to 0.651, P < .001) for PSA and age alone for any-grade cancer, and 0.820 (95% CI = 0.802 to 0.838) vs 0.738 (95% CI = 0.716 to 0.761) for high-grade cancer. Using a 6% risk of high-grade cancer as an illustrative cutoff, for 1000 biopsied men with PSA levels of 3.0ng/mL or higher, the model would reduce the need for biopsy in 428 men, detect 119 high-grade cancers, and delay diagnosis of 14 of 133 high-grade cancers. Models exhibited excellent discrimination on independent validation among men with only serum samples available for analysis. Conclusions: A statistical model based on kallikrein markers was validated in a large prospective study and reduces unnecessary biopsies while delaying diagnosis of high-grade cancers in few men.

As there are a large proportion of men for whom we do not have a blood sample, we wished to assess whether this missing data had an effect on our results. We therefore built a logistic regression model to predict cancer on biopsy using age and the kallikrein markers with data from three centers with the fewest missing blood samples, and we applied this model to the data collected from the remaining six centers. The discrimination of the model was nearly identical for the centers with the fewest missing blood samples and the other six centers (difference in AUC is less than 0.01), suggesting the missing samples did not affect our results.

Analysis of the use of the Rotterdam serum-based model in ProtecT plasma and serum samples.
We initially analyzed the performance of the previous "Rotterdam model", developed using the Rotterdam screening arm cohort of the European randomized screening trial (ERSPC) study (1), in the 496 ProtecT study men with paired plasma and serum samples. The characteristics of these 496 men are shown in Supplementary table 1. We wondered whether the performance of the Rotterdam model in ProtecT samples might be influenced by factors that were different in ProtecT compared with ERSPC. For example, the Rotterdam model had been built using serumbased measurements, and the ERSPC had used a sextant prostate biopsy protocol. We were interested to see how the previous ERSPC-Rotterdam model might perform when it was applied to our contemporary ProtecT study cohort, which used an extended ten-core prostate biopsy protocol, a factor which at least in part explains the 35% rate of any-grade prostate cancer detection in ProtecT versus 25% in Rotterdam ERSPC (2). We were also interested to see how the Rotterdam model would perform for plasma-based samples, given that at least some of the kallikrein markers are more labile in serum than plasma. Finally, we wondered whether the rate of prior PSA testing in our ProtecT study men might influence the performance of the model.
We tested the ability of the Rotterdam serum-based model to predict the presence of any-grade prostate cancer at biopsy using ProtecT plasma and serum samples. When the Rotterdam model was applied to the 496 men with paired samples we observed an enhancement in predicting biopsy outcome compared with the use of PSA alone. However, we also saw a considerable degree of mis-calibration of the Rotterdam model when it was applied to ProtecT samples, particularly in the case of plasma samples (Supplementary Figure 1).

Analysis of the use of the model in serum samples.
We investigated the performance of our new serum-based ProtecT model in predicting any-grade and high-grade prostate cancer in biopsied men with a serum sample available for analysis. The demographics of men with a serum sample are shown in table 1. We observed an important difference between the cohorts of men with either plasma or serum samples in terms of highgrade cancer, with 13% of 4,765 biopsied men with a plasma sample having high-grade disease versus 9% for 1,860 biopsied men with a serum sample. The base model (age plus tPSA) had a relatively poor predictive accuracy for diagnosing any-grade prostate cancer at biopsy for men with serum samples, with an AUC of 0.665 (Supplementary table 2

Marginal value of kallikrein markers.
We sought to investigate the added value of each of the kallikrein markers for predicting high grade and any grade cancer on biopsy. To assess the marginal value of the kallikrein markers we re-fit the logistic regression models omitting each of the markers (the full model includes age, total PSA, free PSA, intact PSA, and hK2). Ten-fold cross validation was utilized to correct the area under the curve (AUC) estimates for overfit. As expected, when total PSA was omitted from the model the AUC decreased by the greatest magnitude. For the outcome of high grade cancer, the AUC fell from 0.820 to 0.703 among patients with a blood plasma measurement and from 0.859 to 0.765 among patients with a blood serum measurement for high grade cancer (Supplementary Table 5). When intact PSA and hK2 were omitted from the model the AUC decreased by 0.02 among both blood plasma and blood serum cohorts for high grade cancer, suggesting an improvement in model performance when including intact PSA and hK2 over total and free PSA alone.

Further Validation of Kallikrein Predictive Models.
The kallikrein models discriminate well between men with and without high grade cancers, and consequently many men with low grade disease or without cancer can avoid prostate biopsy.
These results are based on an internal 10-fold cross validation of the predictive models.