A Panel of Kallikrein Marker Predicts Prostate Cancer in a Large, Population-Based Cohort Followed for 15 Years without Screening

Levels of prostate-specific antigen (PSA) in blood are strongly associated with prostate cancer (1). Interim reports from the European randomized trial (ERPSC), albeit not the PLCO trial in the United States, suggest that screening for prostate cancer based on PSA testing contributes a modest (20%–30%) reduction in cancer mortality after 9 years (2, 3). Recent data from the Göteborg Randomized Trial (i.e., Swedish arm of ERSPC; ref. 4), and a subgroup analysis of PLCO (5), show a substantial reduction in prostate cancer mortality at longer follow-up. However, the PSA test has only modest diagnostic specificity at commonly used cutoff points and is therefore an imperfect test for prostate cancer (6). As most PSA elevations are due to benign disease, many men undergo prostate biopsy unnecessarily, with an estimated 750,000 potentially avoidable biopsies each year in the United States (7). This has led to attempts to increase the specificity of the PSA test, most typically by using other molecular markers. For example, the ratio of unbound (“free”) PSA to total PSA is known to be lower in men with prostate cancer (8–11). This has led to suggestions that men with moderately elevated PSA levels should not undergo biopsy unless their free to total PSA ratio is low (8).

We have developed a statistical model to predict the result of prostate biopsy on the basis of 4 kallikrein marker measurements in blood: total PSA, free PSA, intact PSA, and kallikrein-related peptidase 2 (hK2; ref. 7). We have shown that the use of this model to determine biopsy would lead to biopsy rates being approximately halved (12–14). Although attempts to improve the specificity of PSA will naturally lead to a reduction in sensitivity, with the 4-kallikrein panel, only a small number of men with cancer would be advised against biopsy—66 in 1,000 men with elevated PSA levels—and a majority of these men would have the sort of early-stage, low-grade cancers most often thought to constitute overdiagnosis. Yet, the clinical outcome of these cancers, had they not been identified by biopsy, is unclear.

In this study, we apply our prediction model to archived blood samples obtained during 1991–96 from a large population-based cohort of men aged 45 to 74 at baseline (though most were 60–70) followed for prostate cancer outcomes for up to 15 years. As the rate of PSA screening in the cohort was very low, our analyses provide the natural history of prostate cancer for a given level of risk. This allows us to determine the proportion of men with elevated PSA levels advised against biopsy by the model who would go on to develop clinically significant cancers.

The Malmö Diet and Cancer cohort has been described previously (15). In brief, as part of a large European population-based study (EPIC) to identify dietary risk factors of cancer mortality, 11,063 men who were living in the city of Malmö, Sweden, and born between 1923 and 1945 provided an EDTA anticoagulated blood sample during 1991–1996. The Swedish Cancer Registry was used to identify men diagnosed with prostate cancer up to December 31, 2005. Because of a 1-year delay in registration, this information was obtained from the South Swedish Regional Tumor Registry for 2006.

This study is a cohort design based on case–control data. Of the 11,063 men who participated in this study, 943 were diagnosed with prostate cancer during follow-up. As described previously (16), these cases were each matched with 3 controls who were selected at random from men with date of birth and date of baseline venipuncture within 3 months of the case and who were alive and without a prostate cancer diagnosis at the time at which the index case was diagnosed. To estimate absolute risk, however, requires data from the entire cohort. Instead of retrieving blood samples for all remaining men not selected as controls (n = 7,278), we imputed these participants' blood measurements by methods described in detail in the following text (see Fig. 1). Imputing marker values for unmeasured controls allows for robust estimate of population-based risk while dramatically reducing the total cost of marker measurement.

From the full cohort, we excluded men with a PSA level of less than 3 ng/mL at baseline, as our model was developed for men with elevated PSA levels. Moreover, men with a PSA level of less than 3 ng/mL do not routinely undergo biopsy. For 8 participants, samples were insufficient for analyses; these men were excluded from all analyses.

We have previously shown that the levels of total PSA and free PSA in archival plasma closely correspond to those in contemporaneously measured serum (17). The analyses of free, total, and intact PSA and hK2 were carried out in Dr. Lilja's laboratory at the Wallenberg Research Laboratories, Department of Laboratory Medicine, Lund University, Skåne University Hospital, Malmö, Sweden. Both free PSA and total PSA were measured using the dual-label DELFIA Prostatus total/free PSA-Assay (Perkin-Elmer; ref. 18) calibrated against the WHO 96/670 (PSA-WHO) and WHO 68/668 (free PSA-WHO) standards. The measurements of intact PSA and hK2 were done as previously reported with F(ab')2 fragments of the monoclonal capture antibodies being used to significantly reduce the frequency of nonspecific assay interference (19). The intact PSA assay measures only free, uncomplexed intact PSA (i.e., not the nicked PSA which is cleaved at Lys₁₄₅-Lys₁₄₆). All analyses were conducted blind to outcome.

Our aim was to independently validate our previously published model to predict prostate cancer (12) in a cohort of men largely unscreened for prostate cancer, and as such, whose disease followed its natural history. Our previously published model was constructed using an arm of the ERSPC (Rotterdam) and used a panel of 4 kallikrein markers (total PSA, free PSA, intact PSA, and hK2) and age to predict prostate cancer on biopsy (12). As no digital rectal examination (DRE) was conducted at baseline in the Malmö Diet and Cancer cohort, we are using what was described as the “laboratory model” in our original publication. We made 2 modifications to simplify this model, removing nonlinear terms for intact PSA and hK2 and basing risk only on PSA for men with a PSA level of 25 ng/mL or more at baseline. Both of these modifications were made before applying the model to the Malmö Diet and Cancer cohort. Therefore, this is an independent evaluation of a prespecified model on a new cohort.

Imputation of biomarkers for men without a prostate cancer diagnosis and who were not selected as controls was conducted by randomly sampling with replacement from the values of controls without a prostate cancer diagnosis stratified by age. The imputation was repeated a total of 10 times, and each of the imputed data sets was analyzed using the methods described later; the results from these analyses were averaged to produce central estimates. CIs were obtained using bootstrap methods with 1,000 replications.

We then evaluated the predictive accuracy of the model by the concordance index (C-index), which is analogous to the AUC (area under the curve), and can quantify discrimination of a single variable or a multivariable model for survival time data. Our endpoints included clinically diagnosed prostate cancer (“any cancer”), palpable prostate cancer (clinical stage T2 or higher at the time of diagnosis), or prostate cancer of unquestionable clinical significance (“advanced cancer”: clinical stage T3 or higher or evidence of metastasis at diagnosis). We did not build separate models for these outcomes; instead, the prespecified model was used to generate the predicted probabilities of diagnosis of any cancer, these probabilities were then used to calculate separate C-indexes for the 3 endpoints of any, palpable, and advanced cancers. For these analyses, 23 patients with missing information on clinical stage were considered to have low-stage cancer.

To evaluate the implications of using the model in clinical practice, we used decision curve analysis, which estimates a “net benefit” for prediction models by summing the benefits (true positives) and subtracting the harms (false positives), where the latter is weighted by a factor related to the relative harm of a missed cancer compared with an unnecessary biopsy. A model is of clinical value if it has the higher net benefit across the full range of threshold probabilities at which a patient would choose to undergo biopsy. We used parametric smoothing for the decision curve. Statistical analyses were conducted using Stata 10.0 (StataCorp LP).

In total, 474 patients with a PSA level of 3.0 ng/mL or more at baseline were clinically diagnosed with prostate cancer by December 31, 2006. Median follow-up for those not diagnosed with cancer was 12 years. Table 1 shows the characteristics of the study subjects at baseline.

Table 2 shows the C-index of the full kallikrein panel model built using the Rotterdam data set when applied to Malmö Diet and Cancer participants with a total PSA level of 3.0 ng/mL or more at baseline. For prediction of a prostate cancer, the base model (age plus PSA) had a C-index of 0.654 (95% CI: 0.621–0.683), which was significantly increased to 0.751 (95% CI: 0.726–0.777) for the full model (age plus kallikrein panel). Applying these models to men with palpable prostate cancer (stage T2 or higher at diagnosis) and advanced prostate cancer (clinical stage T3 or higher or metastasis at diagnosis) enhanced discriminatory accuracy in a similar manner. The C-index increased from 0.708 (95% CI: 0.671–0.741) to 0.803 (95% CI: 0.774–0.831) for palpable cancers and from 0.716 (95% CI: 0.664–0.762) to 0.824 (95% CI: 0.785–0.858) for prostate cancers that were advanced at diagnosis. Table 2 also shows that the marker panel had superior properties to a model including total PSA and free PSA and that removing either intact or free PSA reduced accuracy; hK2 did not seem to have an important effect.

Table 3 illustrates the clinical implications of several biopsy schemes: performing biopsy of all men with a PSA level of 3.0 ng/mL or more at baseline, performing biopsy of those with a PSA level of 4.0 ng/mL or more, performing biopsy of those based on an age-specific PSA threshold, or performing biopsy of those with a 20% or greater risk of prostate cancer as predicted from the full kallikrein panel. Compared with the scheme of performing biopsy of all with a PSA level of 3.0 ng/mL or more, the scheme of performing biopsy of those with 20% or more predicted risk from the 4-kallikrein model would have avoided approximately 42% of biopsies (421 biopsies avoided per 1,000 men with a PSA level of ≥3.0 ng/mL). Using this strategy, only 21 men per 1,000 with an elevated PSA test (≥3.0 ng/mL at baseline) who received a recommendation to avoid biopsy would be clinically diagnosed with prostate cancer within 5 years and only 2 of these men would be diagnosed with a prostate cancer that was advanced at diagnosis. In comparison, using a the PSA threshold for biopsy most common in the United States (PSA ≥4.0 ng/mL) would also reduce the number of men undergoing biopsy—36% fewer biopsies—but 23 men with a PSA level of 3.0 ng/mL or more who would have a clinically diagnosed prostate cancer within 5 years would be recommended against immediate biopsy, 4 of whom would have advanced disease at diagnosis. Thus a biopsy strategy based on a higher PSA threshold would avoid fewer biopsies than the full kallikrein panel while also recommending against biopsy for a greater number of men who would subsequently be clinically diagnosed with prostate cancer. Performing biopsy according to age-specific thresholds would reduce the number of biopsies to a similar degree as the kallikrein panel (46% fewer biopsies) but would miss a substantially greater number of cancers (40 per 1,000, including 9 with advanced disease at diagnosis).

Our assumption in defining the outcome as prostate cancers clinically diagnosed within 5 years is that cancers diagnosed many years after the initial screen would likely be detected by a subsequent round of screening. Most men who undergo prostate cancer screening undergo regular PSA tests, especially if their PSA level is high. Figure 2 shows a cumulative incidence curve for diagnosis of palpable and advanced cancers in men with a low risk of cancer according to the marker panel. It is clear that although some men classified as low-risk did develop clinically significant cancers, almost all of these occurred many years after baseline, suggesting that even if biopsy was not recommended at an initial screen, the cancer would be detected at a curable stage at a subsequent screen. Table 3 illustrates these results for cancers clinically diagnosed within 10 years as well as within 5 years.

Figure 3 shows a decision curve analysis for the outcome of a prostate cancer diagnosis within 10 years from baseline. The net benefit of using the full kallikrein panel was higher than a “biopsy all” strategy for all threshold probabilities greater than 7%. In other words, unless a man was willing to undergo a biopsy if he had less than a 7% risk of a subsequent clinical diagnosis of prostate cancer, which seems unlikely, using the full kallikrein panel to determine who should receive biopsy would lead to the best clinical decision making.

Finally, we carried out several sensitivity analyses to address the robustness of our findings. For our first sensitivity analysis, we restricted the cohort of men to those 50 to 70 years old at the time of baseline blood measurement, the age range specified in many screening recommendations. All results were very similar to the main analysis. For example, the C-index of any prostate cancer improved from 0.645 for age and total PSA alone to 0.742 for the full kallikrein model. A second sensitivity analysis included very high-grade cancers (Gleason score ≥8) in the definition of advanced cancer. Again, the results were very similar to our main analyses: the C-index for very high-grade cancer improved from 0.749 to 0.824. Our third sensitivity evaluated how well our panel performed if we lowered the PSA threshold to 2.5 or less, 2.0 or less, or 1.5 ng/mL or less. The results were consistent with our main findings, although the panel was of lower marginal benefit at lower PSA levels. The panel improved the C-index for any cancer for all of the 3 lower PSA thresholds (from 0.677 to 0.756, 0.701 to 0.750, and 0.721 to 0.744, respectively).

We have previously shown that using a panel of kallikrein markers to determine referral to prostate biopsy in men with elevated PSA levels would lead to a large decrease in biopsy rates and would delay the diagnosis of only a small number of cancers. We have also claimed that most of the cancers missed by our panel are sort of early-stage, low-grade cancers typically thought to constitute overdiagnosis. Here we examine the natural history of prostate cancer in men with elevated PSA levels but defined as being at low risk from the panel. In doing so, we provide further evidence that the panel would lower the rate both of unnecessary biopsies and of overdiagnosis. We show that, among 1,000 men with elevated PSA levels (≥3.0 ng/mL), only 9 of those advised against biopsy would go on to diagnosis of palpable prostate cancer within 5 years and only 2 would go on to diagnosis of advanced prostate cancer. Moreover, these numbers would presumably be lower if the men were to undergo additional rounds of screening. This provides direct evidence that only very few men found at low risk by the panel would suffer from a delay in diagnosis that would result in a curable prostate cancer to progressing to an incurable stage by the time of diagnosis. As such, the results suggest that the use of the panel to determine prostate biopsy would improve clinical outcome of biopsy decisions in men with elevated PSA levels.

Our study is unique among prostate cancer marker studies in that we applied a prespecified prediction model to a historical data set to assess the association between marker levels and subsequent clinical diagnosis of cancer in unscreened men. Typical prior studies compare marker levels to biopsy results. Yet, biopsy-detectable cancer is a highly problematic endpoint because a significant proportion of men with such cancers would never have become aware of their cancer during their lifetime in the absence of screening—the problem of overdiagnosis (2). Schröder and colleagues, for example, have analyzed data from the Rotterdam section of the ERSPC for men with a PSA level of less than 3.0 ng/mL (20). On the basis of data from the Prostate Cancer Prevention Trial—in which men underwent biopsy irrespective of the PSA level—they determined that prostate cancer would have been detected in 3,472 of 15,773 men had all of these men been undergone biopsy, yet after 12 years of screening, only 700 prostate cancers were diagnosed. In contrast, our primary endpoint of advanced cancer at diagnosis is clinically relevant by definition.

One important limitation of our study concerns the lack of DRE data. Men did not undergo DRE as part of the Malmö Diet and Cancer study and so predictions are based purely on the kallikrein marker measurements in the blood. As a result, it is possible that a PSA level plus DRE would be of equal predictive value as the kallikrein panel. Yet, we have repeatedly shown in biopsy series that the kallikrein panel retains its predictive advantage even when DRE is taken into account. For example, in men without prior screening, the C-index of PSA and DRE was 0.695 compared with 0.776 for the kallikrein panel plus DRE (12). By the same token, however, the accuracy of the kallikrein panel is higher with DRE than without (0.776 vs. 0.764). This suggests that fewer clinically significant cancers would be missed by the use of the full kallikrein panel combined with the results from DRE, using marker values alone, than were missed in the current study.

An assumption of our method is that men who have elevated levels of total PSA (≥3 ng/mL) at the time of blood draw at baseline and who were subsequently diagnosed with a clinically detected cancer would already have had biopsy-detectable cancer at baseline. It is plausible, however, that some men may have had an elevated PSA level due to benign disease and then developed prostate cancer at a later time. However, this effect would strengthen the case for the use of the model to determine biopsy. This is on the grounds that some of the men at low model-predicted risk who were later diagnosed with prostate cancer would not yet have had biopsy-detectable prostate cancer and so would not have been harmed by a recommendation to avoid biopsy. This article is therefore conservative with respect to the clinical value of the model.

Given that the relationship between markers and prostate cancer may vary by environment and ethnic group, we next plan to test our model in a US population of men with serial PSA levels and DRE. Predictions of our model will be compared to the outcome of prostate biopsy.

In conclusion, we have shown that a prespecified statistical model, based on a panel of four kallikrein markers (total PSA, free PSA, intact PSA, and hK2), is a highly accurate predictor that a man with elevated PSA levels would be clinically diagnosed with an advanced cancer in the absence of screening. Men at low risk according to this model can be reassured that even if they harbor cancer, the cancer is unlikely to become harmful in the near future. Accordingly, any such cancer would likely be detected on repeat blood tests. As such, our results confirm the value of the panel for determining referral to prostate biopsy.

H. Lilja holds patents for free PSA, intact PSA, and hK2 assays; K. Pettersson also holds patents for intact PSA assays. A.J. Vickers had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

We thank Gun-Britt Eriksson and Mona Hassan Al-Battat for expert assistance with immunoassays. This investigation was supported by the National Cancer Institute (grant numbers R33 CA 127768-02, P50-CA92629); Swedish Cancer Society (3455); Swedish Research Council (Medicine-20095 and MTBH: 2006-7600); the Sidney Kimmel Center for Prostate and Urologic Cancers; David H. Koch through the Prostate Cancer Foundation; Fundación Federico SA; the AUA Foundation Research Scholars Program; and the Astellas USA Foundation.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.