The predictive value of prostate cancer biomarkers depends on age and time to diagnosis: Towards a biologically-based screening strategy

Abstract

Both benign and malignant prostate diseases elevate total prostate-specific antigen (tPSA), and the incidence of benign disease increases markedly with age. There is evidence, however, that free-to-total PSA ratio (%fPSA) and human kallikrein 2 (hK2) more closely reflect the malignant process. We tested the hypothesis that tPSA levels are more strongly predictive of cancer in younger when compared to older men, whereas %fPSA and hK2 are more strongly predictive in men tested closer to diagnosis. The study included 13,676 men age ≥ 44 in Sweden, where PSA screening was uncommon during the study period. fPSA, tPSA and hK2 were measured in archived plasma collected during 1974–1986 in 501 men subsequently diagnosed with prostate cancer up to 1999 and in 1,292 matched controls. The predictive value of tPSA was lower in older men (p = 0.003) but was not strongly affected by time to diagnosis (p = 0.3); the predictive value of hK2 was higher closer to diagnosis (p < 0.0005) but was not modified by age (p = 0.7). A model including tPSA, fPSA and hK2 was superior (p = 0.02) to tPSA alone in older (AUC 0.819 vs. 0.794), but not in younger men (0.758 vs. 0.759). Total PSA can be used as a single marker at early middle age to predict long-term risk of prostate cancer and thus to determine intensity of subsequent screening. In contrast, %fPSA and hK2 add important predictive value in older men and much closer to diagnosis. Strategies for prostate cancer screening should be based on thorough understanding of the interaction of kallikrein-related biomarkers with prostate pathobiology. © 2007 Wiley-Liss, Inc.

Screening for prostate cancer with prostate-specific antigen (PSA) testing is controversial.1, 2 The controversy is partly due to a lack of appropriate randomized evidence,3 but the diagnostic characteristics of PSA testing also remain a key concern.3 For example, in their analysis of the Prostate Cancer Prevention Trial, Thompson and colleagues reported sensitivity and specificity of ∼20 and 95% for the commonly used 4 ng/ml threshold for biopsy. Lowering the threshold to improve sensitivity to 30 or 50% decreased specificity to 85 and 75%, respectively.4

The modest diagnostic accuracy of PSA testing has led investigators to evaluate additional biomarkers, such as ratio of free-to-total PSA (%fPSA)5, 6 and human kallikrein 2 (hK2).7, 8 Population-based studies have generally shown that these biomarkers aid in cancer detection.

We have previously shown that a single PSA test taken at age 44–50 is a strong predictor of prostate cancer diagnosed up to 25 years later.9 In some contrast to the prior literature, we reported that although %fPSA and hK2 univariately predicted subsequent prostate cancer, a multivariable model including these markers alongside PSA was not superior to a model with PSA alone. The men in our cohort were generally younger than those in typical studies in the literature and we considered whether the explanation for the disparate findings was related to age.

An elevated level of total PSA can result from benign enlargement of the prostate10 or from inflammation,11 as well as from malignancy. Benign prostate conditions are relatively rare in the young, but very common in older men: population-based studies typically report a prevalence of 2–3% for symptomatic disease in men aged 44–50, the age range in our study, compared to 20–25% in men aged 65–75.12 We therefore hypothesize that PSA is a stronger predictor of cancer in younger than older men.

Whereas PSA is associated with both benign and malignant processes, there is evidence that levels of %fPSA and hK2 relate more closely to the malignant process. Compared to serum PSA, serum hK2 has shown stronger correlations with a variety of disease indicators, including tumor volume,13 high-grade prostate cancer,14 and extracapsular tumor growth.15, 16 Similarly, free and complexed PSA (cPSA) have been shown to be differentially associated with benign and malignant prostate tissue.13, 14 We therefore hypothesize that %fPSA and hK2 add greater predictive value closer to diagnosis.

In our previous paper,9 we reported associations between cancer and biomarkers measured in blood taken from men aged 50 or less who participated in a cardiovascular prevention project (MPM) in Malmö, Sweden. The project enrolled participants (74% acceptance rate) during 1974–1986. In 1981–1982, a group of men aged ∼60 were also invited to participate. To test our 2 hypotheses concerning changes in the value of the biomarkers with age and time to diagnosis, we analyzed associations between biomarkers and subsequent cancer in 2 groups: an “older cohort,” consisting of men 59–61 and a “younger cohort” consisting of men aged 44–50.

AUC, area under the receiver operative characteristic curve; cPSA, complexed prostate specific antigen; fPSA, free prostate specific antigen; %fPSA, free-to-total prostate specific antigen ratio; hK2, human kallikrein 2; MPM, Malmö preventive medicine project; PSA, prostate specific antigen; tPSA, total prostate specific antigen; WHO, world health organization.

Material and methods

Subjects, matching and analytic methods have been described previously.9 In brief, 21,277 MPM participants ≤ 50 gave an EDTA-anticoagulated blood sample and completed medical history and lifestyle questionnaires; of these, 12,509 were aged 44–50. Men aged close to 60 were invited for attendance during 1981–1982. In this subgroup, there were 1,167 men born in 1921 who gave blood samples, completed medical history and lifestyle questionnaires (participation rate 71%) and who were not included in our previous report on the MPM-cohort. There were no further medical check-ups as part of the study unless participants were identified with hypertension, hyperlipidemia, or diabetes.17, 18 Screening for prostate cancer is not recommended in Sweden, and the incidence of PSA testing was low before December 31, 1999, when ascertainment for our study concluded. The cohort is thus a “natural experiment” for long-term prediction of prostate cancer.

Records of participants were linked to the Swedish Cancer Registry to identify men diagnosed with prostate cancer before December 31, 1999. Five men diagnosed with cancer before the date of their baseline blood sample were excluded. Three controls were chosen for each cancer case by matching for age and date of baseline blood sample (±3 months for both factors for ∼95% of controls). Figure 1 shows the flow of participants through the study.

Figure 1. Participant flow through the study.

Archived blood plasmas were retrieved for identified participants and levels of total PSA, free PSA and hK2 measured in plasma as described previously.9 Levels of cPSA were calculated as total PSA minus free PSA. The study was approved by the Ethics Committee of Lund University.

We conducted conditional logistic regression to determine associations between biomarkers and cancer and to analyze interactions between biomarkers and age and time to diagnosis. An interaction between 2 variables exists if the effect of one variable depends on the level of the other variable. A statistically significant odds ratio (OR) of less than 1 for the interaction between, for example, age and total PSA suggests that a 1 unit increase in PSA is associated with a larger increase in risk for a younger man than for an older man. To enable comparisons between biomarkers, we standardized biomarker values by dividing by their standard deviation.

To compare between older and younger men the value of adding biomarkers in addition to total prostate-specific antigen (tPSA), we performed a permutation test with 10,000 replications for the incremental improvement in area under the curve (AUC).

To calculate how the predicted probability of cancer based on a particular level of a biomarker changed with age, we entered the biomarker in a logistic regression model using restricted cubic splines with knots at the tertiles. We excluded observations less than the 2.5th percentile or greater than the 97.5th percentile from the analysis to reduce the effect of outliers (for tPSA we selected 4 ng/ml as the upper limit because it is commonly used for cancer diagnosis and was close to the 97.5th percentile). We then used a Bayes factor to adjust the incidence of prostate cancer to 10%, the estimated incidence by age 75 in this cohort. To calculate the predicted probability of cancer based on time until diagnosis, we entered the marker, age, time to diagnosis and the marker × time to diagnosis interaction term in a logistic regression model. Probability of cancer was calculated using the mean age and with the addition of a Bayes factor, as described earlier. All analyses were 2-sided, and were conducted using Stata 8.2 (Stata, College Station, TX).

Results

According to the Swedish national cancer registry, 556 of the 13,676 male MPM participants aged ≥44 (4.1%) were diagnosed with prostate cancer up to December 31, 1999. Blood samples were missing or could not be analyzed for 53 of the 467 cancer cases in the younger cohort and 2 of the 89 cases in the older cohort, giving 501 cases in total. Three controls were initially matched for each of the 495 cases; for 6 cases only 2 successful matches were made. We later found that 205 controls were not followed until diagnosis of the matched case, usually due to early death, leaving 1,292 controls. Most cases had either 2 (29%) or 3 (65%) matched controls; 31 cases (6%) had only one control followed until diagnosis and 2 cases had no control.

Characteristics of cases are shown in Table I. As expected, the delay between blood sample and diagnosis was shorter in the older cohort; there were no other important differences between groups. Note that although age at diagnosis is greater in the older cohort, this is an artifact relating from the different starting ages of the 2 cohorts and would disappear once all men were followed till the end of their lives.

Table I. Tumor Characteristics of Cancer Cases

	Younger cohort	Older cohort
Data are frequency (%) or median (interquartile range).
Number of cases	414	87
Age at venipuncture (years)
44–46	100 (24%)	–
47–48	216 (52%)	–
49–50	98 (24%)	–
59	–	10 (12%)
60	–	68 (78%)
61	–	9 (10%)
Delay between blood sample and diagnosis (years)	18 (15, 20)	12 (8, 15)
Patient charts available for review	362 (87%)	70 (80%)
WHO grade19
I	137 (37%)	32 (46%)
II	140 (39%)	21 (30%)
III	83 (23%)	17 (24%)
Missing	2 (1%)	0 (0%)
Clinical stage
T1	91 (25%)	12 (17%)
T2	136 (37%)	37 (53%)
≥T3	133 (37%)	21 (30%)
Missing	2 (1%)	0 (0%)
Positive bone scan	57/334 (17%)	18/64 (28%)

Table II shows biomarker levels separately for the younger and older men. In both age groups, median tPSA, cPSA, fPSA and hK2 was higher, and median %fPSA somewhat lower, in plasma collected at baseline for patients who were later diagnosed with prostate cancer compared to those who were not. Irrespective of cancer status, median tPSA, cPSA, fPSA and hK2 was higher for the older cohort. Median %fPSA was higher for older controls, but somewhat lower for older cases.

Table II. Plasma Levels of tPSA, PSA Subforms and hK2 in Cases and Controls by Age at Venipuncture, Given as Median (Interquartile Range)

Biomarker	Cancer cases	Controls
Younger cohort	n = 414	n = 1,089
Total PSA (ng/ml)	1.06 (0.65, 1.77)	0.55 (0.37, 0.81)
Free PSA (ng/ml)	0.31 (0.20, 0.52)	0.18 (0.12, 0.27)
Complexed PSA (ng/ml)	0.69 (0.39, 1.25)	0.33 (0.21, 0.52)
Free to total PSA ratio (%)	30.0 (25.8, 33.7)	32.7 (28.0, 37.9)
Human kallikrein 2 (ng/ml)	0.038 (0.027, 0.055)	0.032 (0.022, 0.044)
Older cohort	n = 87	n = 203
Total PSA (ng/ml)	2.45 (1.02, 4.56)	0.86 (0.51, 1.37)
Free PSA (ng/ml)	0.73 (0.34, 1.24)	0.32 (0.19, 0.49)
Complexed PSA (ng/ml)	1.59 (0.64, 3.36)	0.51 (0.27, 0.85)
Free to total PSA ratio (%)	29.5 (23.3, 34.1)	36.8 (30.9, 41.8)
Human kallikrein 2 (ng/ml)	0.065 (0.041, 0.092)	0.037 (0.026, 0.053)

Table III gives ORs from a univariate conditional logistic regression model performed separately for the older and younger cohorts. Within each age cohort, all of the biomarkers were found to be highly predictive of prostate cancer (p < 0.0005). The ORs of tPSA, cPSA and fPSA were larger in the younger cohort as compared to those of the older cohort, suggesting that tPSA, cPSA and fPSA tell us more about the risk of cancer for younger men. For example, a one-quarter standard deviation increase in tPSA increased the odds of cancer ∼7-fold for younger men, compared to only about 3-fold for older men. The reverse was true for %fPSA and hK2; their coefficients were larger (in absolute value) in the older cohort, suggesting that the predictive power of %fPSA and hK2 is stronger among older men. Note that although we have standardized coefficients by dividing by standard deviation, the coefficients for different markers are not fully comparable due to, for example, the skewed distribution for PSA, which increases standard deviation.

Table III. Univariate Associations Between Biomarkers and Cancer by Age at Venipuncture, as Determined by Conditional Logistic Regression

Biomarker	Younger cohort		Older cohort
	Odds ratio	95% CI	Odds ratio	95% CI
The odds ratios presented for biomarkers are for a one-quarter increase in standard deviation.
Total PSA	6.86	4.91, 9.6	3.27	2.12, 5.05
Free PSA	5.67	4.12, 7.80	2.68	1.85, 3.89
Complexed PSA	6.40	4.62, 8.87	3.35	2.14, 5.25
Free to total PSA ratio	0.85	0.82, 0.89	0.76	0.69, 0.83
Human kallikrein 2	1.75	1.45, 2.11	3.09	1.97, 4.85

The relationships between age and tPSA, %fPSA and hK2 are further depicted in Figure 2. The curves giving the probability of prostate cancer for a given biomarker levels increased more steeply in older than younger men for hK2 and %fPSA; conversely, the curve for total PSA rose more steeply in younger men.

Figure 2. (a) Predicted probability of a prostate cancer diagnosis before age 75 by increasing levels of total PSA in the blood plasma. (b) Predicted probability of a prostate cancer diagnosis before age 75 by decreasing percent free PSA in the blood plasma. (c) Predicted probability of a prostate cancer diagnosis before age 75 by increasing levels of hK2 in the blood plasma. Dashed: younger cohort; solid: older cohort.

To formally analyze the impact of the association between age at venipuncture and the biomarkers, we entered an interaction term into the conditional logistic regression model. Because age is correlated with time to diagnosis, and we would expect biomarkers to be more strongly predictive closer to diagnosis, we also included the interaction term between the biomarker and time to diagnosis. The OR for the interaction term between total PSA and age was statistically significant (p = 0.003) and less than 1 (OR = 0.92 per year of age), illustrating the decreasing value of total PSA with increasing age. The interaction term between total PSA and time to diagnosis was closer to 1 (OR = 0.96 per year to diagnosis) and nonsignificant (p = 0.3), suggesting that total PSA retains its predictive value even many years before a patient is diagnosed with cancer. These associations were reversed for hK2: age had no impact on the value of hK2 (OR = 1.01 per year of age; p = 0.7); however, hK2 was most predictive close to diagnosis (OR = 0.93 per year to diagnosis; p < 0.0005). The effects for %fPSA were moderate in size: it appeared most useful in older men (OR = 0.99 per year of age; p = 0.07) and those closer to diagnosis (OR = 1.01 per year to diagnosis; p = 0.043).

These relationships are compared graphically in Figure 3. This shows the results for younger men (the data for older men were similar and are not shown). The curves for a high and low level of hK2 and %fPSA converge over time until 25 years before diagnosis, when the predicted risk for men with high or low levels of these biomarkers is virtually identical. Therefore, hK2 and %fPSA provided limited information about prostate cancer unless measured close to clinical diagnosis. In contrast, the curves for tPSA are approximately parallel, showing that time until diagnosis had only a small effect on the predictive value of tPSA: a man with a high PSA is more likely to get prostate cancer than a man with a low PSA and this is true up to 25 years before diagnosis.

Figure 3. Predicted probability of a prostate cancer diagnosis before age 75 by time until diagnosis, for a given value of total PSA, percent free PSA or hK2, controlling for age at venipuncture, for men age 44–50. Thick solid lines: total PSA; dashed lines: hK2; thin grey lines: percent free PSA. For each marker, the upper line is the 75th centile and the lower line the 25th centile. The separation between the upper and lower line is an indicator of the prognostic information: the wider the separation, the more prognostic the marker.

We calculated the AUC for predictive models with tPSA alone and tPSA plus fPSA, %fPSA and hK2 separately by age group, using 10-fold cross validation (Table IV). The incremental difference in AUC between age groups (younger vs. older) was 0.026 (p = 0.02 by the permutation test). This suggests that additional biomarkers aid in discrimination of prostate cancer from non-cancer more for older men than for younger men. Note that although the AUC of tPSA was higher in older men when compared to younger men, this difference was consistent with random variation (p = 0.4) and does not take into account time to diagnosis. AUC of the full model (tPSA + PSA + %fPSA + hK2) was significantly higher in older men when compared to younger men (p = 0.03).

Table IV. Area Under the Curve for Basic and Full Models by Age at Venipuncture

Model	Area under the curve
	Younger cohort	Older cohort
tPSA	0.759	0.794
tPSA + fPSA + %fPSA + hK2	0.758	0.819
Difference	−0.001	0.025

Discussion

Given the low prevalence of prostate cancer screening in Sweden, most prostate cancers are detected clinically. This is reflected in predominance of palpable tumors (∼75%) in our sample. We therefore believe that the MPM study cohort constitutes a unique “natural experiment” for testing hypotheses related to prostate cancer screening. In our study, we have tested 2 biologically-based hypotheses.

Firstly, our analysis has shown that, after adjustment for time to diagnosis, the predictive value of PSA is inversely related to age. Secondly, we have shown that, after adjustment for age, the predictive value of %fPSA and hK2 (but not tPSA) are inversely related to time to diagnosis. As a result, adding %fPSA and hK2 to tPSA adds predictive value in older, but not younger men. As described earlier, these markers are known to increase diagnostic accuracy in a screened population, but their relation to long-term prediction of clinically presenting disease occurrence has not been previously described.

Our finding that the prognostic value of PSA is not strongly related to time to diagnosis is apparently inconsistent with 2 previous papers on long-term prediction of prostate cancer.20, 21 However, the inconsistency is more apparent than real. In a study of PSA determined from blood samples stored in 1964–1971, Whittemore et al.20 report a change point in PSA around 13 years before diagnosis of prostate cancer. This appears to be based on a visual inspection of the data rather than on any formal statistical analysis: the authors do not, as we do, test the hypothesis that the predictive value of PSA is related to time to diagnosis. Stenman et al.21 report associations between PSA measured from blood stored 1968–1973 and prostate cancer diagnosed before 1980. In particular, the authors suggest that men with higher PSA are diagnosed sooner than men with lower PSA. This is a quite different question from that addressed here: we ask whether the prognostic value of a high versus a low level of a marker is affected by time to diagnosis. It turns out that this is the case for hK2 but not PSA. In other words, a man with higher hK2 will be diagnosed before a man with lower hK2 if this marker is measured within 10, but not 25 years of diagnosis; conversely, the man with the higher PSA will always be diagnosed first, even if diagnosis is a long time in the future. Two other papers support our findings that PSA measured before the age of 50 can predict subsequent prostate cancer: that of Gann et al.,22 who used data from the Physicians' Health Study and that of Whittemore et al.,23 who used blood given by fathers participating in the Child Health and Development Study. However, neither of these 2 studies directly addressed the relationship between predictiveness of PSA and time to subsequent diagnosis.

The results of the study have obvious clinical implications. Current recommendations for prostate cancer screening, such as those from the American Cancer Society, suggest yearly PSA screening starting at age 50, continuing until the man's life expectancy is less than 10 years. Our results suggest a different strategy: men should have an initial PSA test in at age 44–50. The purpose of this initial test would not primarily be to detect incident cancer—though clearly men with high PSA levels could be referred for biopsy—but to stratify men for the intensity of subsequent screening tests. Tests later in a man's life would be used to detect cancer, and thus determine referral to biopsy, and could incorporate biomarkers other than PSA. Potentially, this detection and monitoring strategy might make it possible to decrease risks and increase the opportunity of managing patients with watchful waiting and active surveillance.

Our approach takes into account how the diagnostic test characteristics of markers change with the age of the patient. Moreover, the strategy is developed based on an understanding of both the changing biology of the prostate over time and the biology of the markers themselves, specifically, how marker levels are affected by benign and malignant changes within the prostate. We would argue that this is in distinction to current recommendations, which are empirical in nature and are often based on findings that have little relationship to clinical outcomes. For example, the often cited PSA value of 4 ng/ml PSA as a threshold for referral for biopsy appears to have been derived from population-based reference ranges.24 Interestingly, there have been attempts to update this threshold on the basis of new data for the distribution of PSA in the population.25 Statistical norms derived from observational data are, we feel, inferior to biologically-based prediction models that link biomarker levels to prostate cancer diagnoses in an unscreened population. We aim to develop such models as the data from the Malmö cohort mature.

Acknowledgements

Dr. Vickers' work on this research was funded by a P50-CA92629 SPORE from the National Cancer Institute. Dr. Lilja was funded by P50-CA92629 SPORE Pilot Project 7 from the National Cancer Institute, Swedish Cancer Society project no. 3555, European Union 6th Framework contract LSHC-CT-2004-503011 (P-Mark), and Fundación Federico SA.