Purpose: The relationship between prostate-specific antigen (PSA) level and prostate cancer risk remains subject to fundamental disagreements. We hypothesized that the risk of prostate cancer on biopsy for a given PSA level is affected by identifiable characteristics of the cohort under study.
Experimental Design: We used data from five European and three U.S. cohorts of men undergoing biopsy for prostate cancer; six were population-based studies and two were clinical cohorts. The association between PSA and prostate cancer was calculated separately for each cohort using locally weighted scatterplot smoothing.
Results: The final data set included 25,772 biopsies and 8,503 cancers. There were gross disparities between cohorts with respect to both the prostate cancer risk at a given PSA level and the shape of the risk curve. These disparities were associated with identifiable differences between cohorts: for a given PSA level, a greater number of biopsy cores increased the risk of cancer (odds ratio for >6- versus 6-core biopsy, 1.35; 95% confidence interval, 1.18-1.54; P < 0.0005); recent screening led to a smaller increase in risk per unit change in PSA (P = 0.001 for interaction term) and U.S. cohorts had higher risk than the European cohorts (2.14; 95% confidence interval, 1.99-2.30; P < 0.0005).
Conclusions: Our results suggest that the relationship between PSA and risk of a positive prostate biopsy varies, both in terms of the probability of prostate cancer at a given PSA value and the shape of the risk curve. This poses challenges to the use of PSA-driven algorithms to determine whether biopsy is indicated. Clin Cancer Res; 16(17); 4374–81. ©2010 AACR.
Prostate cancer research has hitherto been based almost exclusively on studies of individual cohorts. The result has been a plethora of conflicting findings as to the most basic questions in the field, such as whether risk rises as prostate-specific antigen (PSA) increases above biopsy thresholds or whether there is a substantive risk of cancer at low PSAs. We analyzed prebiopsy levels of PSA on 25,772 prostate biopsies and 8,503 cancers from 8 different cohorts in 6 different countries. There were gross disparities between cohorts with respect to both the prostate cancer risk at a given PSA level and the shape of the risk curve. This poses challenges to the use of PSA-driven algorithms to determine whether biopsy is indicated: neither PSA cut points for biopsy, such as 4 ng/mL, nor the use of “risk calculators” based on PSA, seem likely to give results that are generalizable across different cohorts.
Prostate-specific antigen (PSA) is the only molecular marker used for screening a common cancer. Nonetheless, more than 30 years after its discovery, the relationship between PSA level and prostate cancer risk has not been fully characterized.
Indeed, there remain disagreements even on such fundamental issues as whether risk rises with PSA and whether a low PSA level can rule out prostate cancer. For example, whereas Stamey et al. have claimed that PSA does not predict prostate cancer in contemporary cohorts (1), the Prostate Cancer Prevention Trial (PCPT) investigators reported a very strong association between PSA and the risk of a positive biopsy (2). Similarly, although the results of the PCPT suggest that prostate cancer is common in men with very low PSA levels, an analysis of data from the Rotterdam section of the European Randomized study of Screening for Prostate Cancer (ERSPC) found that very few men with low baseline PSA were diagnosed with prostate cancer after 12 years of follow-up (3).
Such differences have been expressed quantitatively in prediction models that calculate an individual's probability of having biopsy-detectable prostate cancer on the basis of PSA level and other risk factors, such as age and family history. These models, generally known as “risk calculators”, have been shown to provide very different estimates of risk for a given PSA level (4). There are two possible reasons for these differences. First, investigators use different statistical methods to develop their models, including adjustment for different covariates. Second, each prediction model is based on a single cohort of men undergoing biopsy and these cohorts differ in a variety of ways, such as the participants' history of PSA screening, whether or not PSA was used as a sole criterion for biopsy, and the number of cores taken during biopsy.
We hypothesized that the relationship between PSA and the risk of prostate cancer on biopsy depends on these identifiable characteristics of the cohort under study. To test this hypothesis, we obtained raw data from a heterogeneous set of prostate biopsy cohorts. We then analyzed these data sets using standardized statistical methods.
Materials and Methods
We utilized data sets from eight biopsy cohorts, five from Europe and three from the United States. Descriptions of the study cohorts, including biopsy algorithm and biopsy scheme, are given in Table 1. The data sets were restricted to subjects age 50 to 75 when biopsied, and for ongoing studies to data made available in 2009, and hence may differ from prior or future publications on the same cohort. Data sets were obtained under waivers from local ethics review committees.
Four of the data sets were from studies of PSA screening for prostate cancer, namely, the Göteborg, Rotterdam, and Tarn sections of the ERSPC (5) and the San Antonio Center of Biomarkers of Risk for Prostate Cancer (SABOR; ref. 6), whereas two were from clinical cohorts (Cleveland Clinic, Durham VA). The Tyrol cohort was not expressly established as a screening trial, but involved extensive attempts to encourage men to attend free-of-charge PSA testing, with the result that approximately 80% of the target population attended screening (7). Similarly, the Prostate Testing for Cancer and Treatment (ProtecT) study (8) is not a screening study, but used population-based invitations to PSA testing to accrue patients to a treatment trial.
The Göteborg and Rotterdam cohorts consisted of data from the first biopsy. The cohorts were divided depending on whether biopsy occurred after a first or later PSA test (i.e., recently screened versus previously unscreened). The Tarn cohort consisted of men biopsied during the first round of ERSPC. Although some participants had prior PSA screening, this was not well characterized and so we did not split into two separate cohorts those with and without previous screening. For the SABOR cohort, if a participant had multiple biopsies, then only the last biopsy was included.
In total, we obtained raw data from 8 centers in 6 countries, comprising 25,772 biopsies in 23,070 men with 8,503 cancer cases. All patients in the data set underwent biopsy, so there is no issue of verification bias. As a comparator for our raw data cohorts, the risk curve for the PCPT was obtained using the formula for the PCPT risk calculator.
Measurements of PSA were made in accordance with procedures in place at the respective institutions. The type of assay used at each institution during each year was documented. PSA measures were then recalibrated to the WHO standard (PSA-WHO 96/670) using an appropriate correction factor.
The predicted probability of biopsy-detected prostate cancer for a given PSA was calculated using locally weighted scatterplot (“lowess”) smoothing (9). All PSA values <100 ng/mL were used in the calculation of the risk curve, although risk curves only for PSA values ≤10 ng/mL are displayed.
Logistic regression was used to test for a difference in the biopsy detection rate based on biopsy scheme (six cores versus more than six), previous PSA screening (ever versus never), and geographic location (U.S. versus European cohorts). The regression analyses excluded men with prior biopsy, and therefore included only one biopsy per patient. The analysis of biopsy scheme was restricted to European cohorts because very few patients in the United States received limited biopsy. Similarly, the analysis of previous PSA screening was restricted to European cohorts because all U.S. cohorts included patients with previous PSA screening; the Tyrol cohort was also excluded from this analysis, because it was unknown who had previous screening. Covariates for these analyses were PSA, digital rectal exam (DRE), and family history; the analysis of prior screening also included biopsy scheme as a covariate and vice versa. Patients with PSA values ≥100 ng/mL (n = 104, 0.4%) were excluded from the logistic regression analyses. All analyses were conducted using Stata 10.1 (Stata Corp.).
The clinical characteristics and biopsy results of each cohort are given in Tables 1 and 2. The cohorts are highly comparable in terms of age and prostate volume. Cancer prevalence varied from 26% (ERSPC cohorts) to 47% (Durham VA), although 7 of the 10 cohorts had prevalence <35%. PSA levels varied systematically with cancer grade: cohorts with a lower median PSA, such as ERSPC and SABOR, had a lower proportion of high-grade cancers than cohorts with high median PSA, such as Durham VA and Cleveland Clinic. There are large, unexplained differences in DRE findings between cohorts. For example, the rate of positive DRE was three times higher in the SABOR cohort compared with Tyrol (29% versus 10%), despite a lower incidence of high-grade tumors (8% versus 21%). This may reflect the subjective nature of DRE.
Our principal results are given in Fig. 1, which shows the probability of a positive prostate biopsy by PSA level for each cohort. Figure 1A shows gross disparities between cohorts with respect to both the prostate cancer risk at a given PSA level and the shape of the curve. As race and DRE affect risk and are commonly included in prostate cancer risk calculators, Fig. 1B shows results when men with positive DRE and those of African origin are excluded. Large differences between cohorts remain. For example, at the commonly used PSA threshold of 4 ng/mL, risk varies from 15% (Göteborg round 1) to 40% (SABOR). Figure 1C repeats this analysis for high-grade cancer, that is, patients with Gleason 6 or below were categorized as disease-free. There are large differences in risk by PSA level, with similar patterns between cohorts as for the analysis of all cancers.
One obvious difference between cohorts is that some include men with very low PSAs whereas others, such as ERSPC, only include men with PSA of ≥3 ng/mL. However, there were no important differences in the risk curves when analysis was conducted restricting the sample to men with PSA ≥3 ng/mL (data not shown).
We then examined whether identifiable aspects of the cohorts might explain the different relationships between PSA level and prostate cancer risk. Figure 2 compares ProtecT with the first round of the two ERSPC cohorts. All three cohorts were European, population-based trials of men without a history of PSA screening; ProtecT, however, used 10-core biopsy in contrast to the 6-core scheme of the ERSPC. It is apparent that the use of more biopsy cores finds more cancers at a given PSA level. In an analysis including only European cohorts, the number of cores was significantly associated with cancer risk after adjustment for PSA level and prior screening [odds ratio (OR) for >6- versus 6-core biopsy, 1.35; 95% confidence interval, 1.18-1.54; P < 0.0005].
In Fig. 3 we compare round 1 of the ERSPC, which included only unscreened men, with subsequent rounds of the ERSPC, which included only men with prior PSA screening. Prior PSA screening clearly leads to a flattening of the relationship between PSA level and the likelihood of a positive biopsy. As PSA doubles from 4 to 8 ng/mL, risk increases by about 50% in previously unscreened ERSPC participants, with no large increases in risk for ERSPC participants with prior screening. We tested for an interaction between PSA level and prior screening, which would indicate a different relationship between PSA and cancer in participants with and without prior screening. The interaction term was statistically significant (P = 0.001) and negative (i.e., OR <1), confirming the hypothesis that PSA has a smaller effect on the risk of prostate cancer for participants with prior screening. It can also be seen in Fig. 3 that the curve for ERSPC Göteborg participants with prior screening is flatter than the respective curve for ERSPC Rotterdam participants. This could be due to the shorter screening interval in Göteborg (2 years) compared with Rotterdam (4 years).
One unexplained characteristic of our findings is that the U.S. cohorts generally have higher risk than the European cohorts for a given PSA level. For example, the risk at a PSA of 4 ng/mL is approximately 20% for ERSPC Rotterdam but approximately 40% for the Cleveland Clinic. The effect of geographical location was statistically significant when controlling for PSA level (OR for U.S. versus European cohorts, 2.14; 95% confidence interval, 1.99-2.30; P < 0.0005).
We combined data from 8 prostate biopsy cohorts to create a single data set of 25,772 biopsies with 8,503 cancers. Our results suggest that the relationship between PSA and prostate cancer risk varies, both in terms of the probability of a detectable prostate cancer at a given PSA and the shape of the risk curve. These differences are not trivial, with a 2.5-fold difference in risk between cohorts at the commonly used PSA threshold of 4 ng/mL.
We were able to identify characteristics of cohorts that affect the relationship between PSA level and prostate cancer risk: the use of a greater number of biopsy cores increases risk, and the relationship between PSA and risk is flatter if men have had prior PSA tests. Both of these findings are well in keeping with prior literature and make clear clinical sense. With respect to biopsy scheme, prostate tumors are often small and may be missed if only a limited number of biopsy samples are taken. As regards previous screening, most clinicians would agree that in an unscreened man, a PSA of 8 ng/mL constitutes higher risk than a PSA of 6 ng/mL. However, if the man had been screened two years previously, and had a PSA below current biopsy thresholds, many would consider the difference between 8 and 6 ng/mL to be clinically irrelevant “noise” that is most likely associated with benign prostate disease.
We have also seen that U.S. cohorts generally have higher risk than European cohorts. Given that a genetic difference seems unlikely, on the grounds that large differences persisted after exclusion of patients with African origins (OR, 2.02; 95% confidence interval, 1.86, 2.19; P < 0.0005), we see several possible explanations for this observation. First, U.S. biopsies in our data set almost exclusively involved ≥8 cores, whereas many of the European biopsies were 6 core. That said, the OR for biopsy scheme (1.35) is much lower than that for geographic region when European patients with >6 cores were excluded (OR, 2.11).Second, there may be differences in biopsy technique (10) or pathologic analysis, for example, if U.S. urologists were less likely to miss cancer. There is no direct evidence of such an effect. Third, the European cohorts tend to be representative of the population as a whole, whereas the U.S. cohorts include selected groups. The fourth explanation, and to us the most plausible one, is that in the U.S. cohorts, prostate biopsy was more usually at the discretion of the attending urologist. Such a decision might be influenced by considerations such as a workup for benign disease or longitudinal monitoring over time to see whether PSA decreases (11). Our data set includes only the final PSA at the time of biopsy, not the initial PSA prompting consideration of biopsy. Take the case of two patients with an initial PSA of 4 ng/mL: in the United States it may be that only one would proceed to biopsy, because elevated PSA in the other patient is ascribed to a benign prostatic condition discovered during clinical workup; in the ERSPC, both patients would proceed to biopsy as per the study protocol. This would lead to an overall lower proportion of patients found to have cancer on biopsy in the European cohorts.
We are aware that “prostate cancer risk” can refer not only to the probability that a man currently has prostate cancer, the sense in which it is used in this paper, but also to the likelihood that he will develop it subsequently. There is excellent evidence that PSA has a clear and consistent association with subsequent risk of clinically detected cancer, especially advanced cancers likely to prove life threatening (12–14). These findings are unaffected by the results presented here.
However, our findings do have several implications for prostate cancer detection. First, our work casts doubt on the simplistic use of PSA cutoffs to determine biopsy. A common approach has been to call men with a PSA <4 ng/mL “negative” whereas a PSA of 4 ng/mL would indicate biopsy. Here we show that this threshold is associated with risks of prostate cancer ranging from 15% to 40%. It is unsound to give the same recommendation to patients across such a wide spectrum of risk. The alternative to the use of cut points is to accept that risk varies continuously, such that a PSA of 3.9 ng/mL constitutes a different risk from 0.5 ng/mL (15). Combined with the insight that factors other than PSA can affect risk, such as age, race, family history, and prior negative biopsy, this has led investigators to develop multivariable prediction models that provide a patient with a percentage probability of cancer. These risk calculators are easily accessible via web applications and seem to have become widely used in clinical practice.
Nonetheless, our work poses comparable challenges to risk calculators as to PSA thresholds. The underlying statistical models were typically created by analysis of a single cohort, for example, the control arm of the PCPT for the PCPT risk calculator (15) and the Rotterdam section of the ERSPC for the ERSPC risk calculator (16, 17). Although these models include several predictors, they are typically strongly dependent on PSA: the discriminative accuracy of the PCPT calculator, for example, is 0.702 compared with 0.678 for PSA alone (15).
Given that the relationship between PSA level and risk of cancer at biopsy is highly dependent on identifiable cohort characteristics, we do not believe it sound for investigators to analyze a single cohort and develop a risk prediction model unless they either restrict to specific types of patients, such as men without prior screening, or separately evaluate their tool in a wide range of settings.
The obvious corollary is that care should be taken with respect to the clinical use of prediction models. It is far from clear that the PCPT risk calculator, developed on U.S. men subject to intense screening, is applicable to unscreened European men. Indeed, the PCPT risk calculator has been applied retrospectively to the ERSPC and has been found to be miscalibrated, giving a higher risk than was found (3, 17). By the same token, it is questionable whether the risk calculator based on European men (16) can be applied to a U.S. clinical population. It may be that risk calculators based on a panel of markers may overcome the limitations of PSA alone (18). Regardless, the value of any model would need to be shown by validating the risk calculator on a variety of different types of cohort, particular those in clinical settings.
Investigators should be cautious about drawing general conclusions about PSA and cancer risk on the basis of a single cohort. For example, Schwartz and colleagues analyzed data from biopsies “performed at our institution”, a hospital in the United States, and concluded that, in recent years, there was essentially no correlation between PSA and biopsy outcome in DRE-negative men with a PSA >2 ng/mL (19). Such a conclusion is well in keeping with our findings for cohorts with recent screening, but not our findings for cohorts consisting of men without a recent PSA test.
This observation has profound consequences for research methodology. Prostate cancer research has hitherto been based almost exclusively on studies of individual cohorts. The result has been a plethora of conflicting findings as to the most basic questions in the field, such as whether risk rises as PSA increases above biopsy thresholds or whether there is a substantive risk of cancer at low PSAs. We have found that sharing data as a collaboration between multiple centers provides robust insights into PSA, and helps explain the discrepancies between previously published findings. We would encourage other researchers to go beyond the culture of competing publications and work together on shared data sets.
In conclusion, by analyzing data from multiple cohorts, we have shown that the relationship between PSA and the risk of biopsy-detectable prostate cancer systematically depends on the type of cohort studied. This poses challenges to the use of PSA to determine biopsy, whether in the form of a PSA threshold, such as 4 ng/mL, or in statistical risk calculators where risk is driven primarily by PSA level.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
We acknowledge the tremendous contribution of all members of the ProtecT study research group. The views and opinions expressed therein are those of the authors and do not necessarily reflect those of the Department of Health.
Grant Support: Statistical center supported in part by funds from David H. Koch provided through the Prostate Cancer Foundation; the Sidney Kimmel Center for Prostate and Urologic Cancers; P50-CA92629 SPORE grant from the National Cancer Institute to Dr. P.T. Scardino. Grants to support the work of the ERSPC include: European Union Grants SOC 95 35109, SOC 96 201869 05F022, SOC 97 201329, SOC 98 32241, the 6th Framework Program of the EU: PMark:LSHC-CT-2004-503011; The Dutch Cancer Society (KWF 94-869, 98-1657, 2002-277, 2006-3518); The Netherlands Organisation for Health Research and Development (ZonMW-002822820, 22000106, 50-50110-98-311). Prostate Cancer Research Foundation of Rotterdam (SWOP); Beckman-Coulter-Hybritech Inc; Abbott Pharmaceuticals, Sweden; Af Jochnick's Foundation; Catarina and Sven Hagstroms Family Foundation; Gunvor and Ivan Svensson's Foundation; Johanniterorden, King Gustav V Jubilée Clinic Cancer Research Foundation; Sahlgrenska University Hospital; Schering Plough, Sweden; Swedish Cancer Society (Contract numbers 09 0107, 080315 and 083455); Swedish Research Council (Medicine) project no. 20095; Fundación Federico SA. Wallac Oy, Turkku, Finland. Grants to support the Tyrol study: Supported by the International Agency for Research on Cancer, Lyon and the Tyrolean Prostate Cancer Early Detection Group. The SABOR project is supported by the San Antonio Center of Biomarkers of Risk for Prostate Cancer U01 CA86402. The ProtecT study is funded by the Health Technology Assessment Programme of the National Institute for Health Research (projects 96/20/06, 96/20/99).
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.
- Received May 17, 2010.
- Accepted May 29, 2010.