Prostate-specific antigen velocity (PSAV) risk count improves the specificity of screening for clinically significant prostate cancer

Authors


William J. Catalona, 675 N. Saint Clair Street, Suite 20-150, Chicago, IL 60611, USA. e-mail: wcatalona@nmff.org

Abstract

Study Type – Prognosis (retrospective cohort analysis)

Level of Evidence 2b

What's known on the subject? and What does the study add?

PSA screening reduces prostate cancer mortality but also may lead to unnecessary biopsies and overdiagnosis of insignificant tumours. PSA velocity (PSAV) risk count (number of serial PSAV exceeding 0.4 ng/ml/year) significantly improves the performance characteristics of screening for overall prostate cancer and high-grade disease on biopsy. Risk count may be useful to reduce unnecessary biopsies and prostate cancer overdiagnosis compared to PSA alone.

OBJECTIVE

  • • To determine whether the prostate-specific antigen velocity (PSAV) risk count (i.e. the number of times PSAV exceeds a specific threshold) could increase the specificity of screening for prostate cancer and potentially life-threatening tumours.

PATIENTS AND METHODS

  • • From 1989 to 2001, we calculated two serial PSAV measurements in 18 214 prostate cancer screening-study participants, of whom 1125 (6.2%) were diagnosed with prostate cancer.
  • • The PSAV risk count was determined as the number of PSAV measurements of >0.4 ng/mL/year (0, 1, or 2).
  • • We used receiver operating characteristic (ROC) and reclassification analyses to examine the ability of PSAV risk count to predict screen-detected and high-grade prostate cancer.

RESULTS

  • • The PSAV was >0.4 ng/mL/year twice (risk count 2) in 40% of prostate cancer cases compared with only 4% of those with no cancer (P < 0.001).
  • • After adjusting for age and PSA level, a PSAV risk count of 2 was associated with an 8.2-fold increased risk of prostate cancer (95% confidence interval 7.0–9.6, P < 0.001) and 5.4-fold increased risk of Gleason score 8–10 prostate cancer on biopsy.
  • • Compared with a model with age and PSA level, the addition of the PSAV risk count significantly improved discrimination (area under the ROC curve 0.625 vs 0.725, P= 0.031) and reclassified individuals for the risk of high-grade prostate cancer (net reclassification, P < 0.001).

CONCLUSIONS

  • • Sustained rises in PSA indicate a significantly greater risk of prostate cancer, particularly high-grade disease.
  • • Compared with men with a risk count of ≤1, those with two PSAV measurements of >0.4 ng/mL/year (risk count 2) had an 8-fold increased risk of prostate cancer and 5.4-fold increased risk of Gleason 8–10 disease on biopsy, adjusting for age and PSA level.
  • • Compared to PSA alone, PSAV risk count may be useful in reducing unnecessary biopsies and the diagnosis of low-risk prostate cancer.
Abbreviations
PLCO

the Prostate, Lung, Colorectal, Ovarian trial

ERSPC

the European Randomized Study of Screening for Prostate Cancer

RP

radical prostatectomy

ROC

receiver operating characteristic

NRI

net reclassification improvement

AUC

area under the curve

OR

odds ratio.

INTRODUCTION

PSA-based screening has become increasingly common in the USA and abroad [1]. According to data from the 2001 Behavioral Risk Factor Surveillance System, 54% of men aged 50–69 years had an up-to-date PSA test [2]. A separate cross-sectional survey of male physicians reported that 87% of those aged ≥50 years had undergone PSA testing [3]. In the 2005 National Health Interview Survey, 49% of men aged 50–79 years reported a PSA test within 2 years [4].

Two randomized trials of PSA-based screening with conflicting results were recently reported. The Prostate, Lung, Colorectal, Ovarian (PLCO) trial reported no difference in mortality between the screening and control arms [5]. By contrast, in the European Randomized Study of Screening for Prostate Cancer (ERSPC), PSA screening was associated with a 20% reduction in prostate cancer mortality and 41% reduction in metastatic disease at diagnosis [6]. Nevertheless, ERSPC authors estimated that 1410 men would need to be screened and 48 treated to prevent one prostate cancer death at 9 years. More recent studies have shown a reduction in number needed to treat with longer follow-up [7,8]. Nevertheless, the on-going controversy about the diagnosis of prostate cancer that may not cause harm highlights the need for screening methods with greater specificity for clinically significant prostate cancer.

Both the PLCO and ERSPC used the absolute PSA level to recommend a prostate biopsy [5,6]. PSA velocity (PSAV), or the change in PSA units per year, was not included in the protocol for either trial. Although there is conflicting evidence [9], some studies have shown an association between PSAV with prostate cancer-specific mortality after surgery and radiation therapy [10,11].

In 2007, Carter et al. [12] proposed a novel way to evaluate PSA dynamics, a concept known as the ‘PSAV risk count’. It is calculated by counting the number of times that PSAV exceeds a specific threshold value. Among 717 unscreened men from an ageing study, the relative risk was 1.49 (i.e. ≈50% increased risk) for high-risk prostate cancer when the PSAV was >0.4 ng/mL/year more than once (risk count >1). Moreover, the addition of the PSAV risk count to a model with PSA level, age and date significantly improved model fit.

The PSAV risk count concept has not been examined in a screening population. We sought to determine whether PSAV risk count could improve the specificity of PSA screening for prostate cancer and specifically for high-grade disease.

PATIENTS AND METHODS

From 1989 to 2001, 35 536 men participated in a prostate cancer-screening study. The PSA level was measured and DRE was performed at 6–12 month intervals, as previously described [13]. Prostate biopsy was recommended for PSA levels of >4 ng/mL (before 1995) or >2.5 ng/mL (after 1995), and/or suspicious findings on DRE. The study protocol was approved by the Institutional Review Board, and all participants provided informed consent.

We excluded 17 322 (48.8%) men with less than three PSA measurements before prostate cancer diagnosis or censorship, for whom a PSAV risk count calculation was not possible. Of excluded individuals, 2444 (14.1%) were diagnosed with prostate cancer. The final study population consisted of the remaining 18 214 men, including 1125 (6.2%) diagnosed with prostate cancer.

Two successive PSAV calculations were performed, PSAV2 and PSAV1. PSAV2 included the final PSA (at diagnosis or censorship) and the PSA measurements before that. PSAV1 was defined as the PSAV leading up to PSAV2. In all cases, the PSAV calculation was restricted to PSA values separated by at least 6 months and a maximum of 24 months.

The PSAV risk count was then calculated by counting the number of times that PSAV exceeded the threshold value of 0.4 ng/mL/year. This threshold value was selected based upon the original risk count description by Carter et al. [12], as well as by our own recent studies showing the association between a single PSAV measurement of >0.4 ng/mL/year with prostate cancer detection [14], and with the presence of clinically significant prostate cancer at radical prostatectomy (RP) [15]. Also, the 2011 National Comprehensive Cancer Network Guidelines recommend considering prostate biopsy for men with total PSA levels of ≤2.5 ng/mL and a PSAV ≥0.35 ng/mL/year [16].

In the present study, a risk count of 0 indicated that neither PSAV2 nor PSAV1 exceeded 0.4 ng/mL/year. A risk count of 1 signified that either PSAV2 or PSAV1 was >0.4 ng/mL/year but not both; whereas, a risk count of 2 meant that both PSAV2 and PSAV1 were >0.4 ng/mL/year.

We used SAS for LINUX for all statistical analysis. Demographic characteristics and the PSAV risk count were compared using the t-test, Wilcoxon rank-sum, chi-square and Fisher's exact tests. The Armitage chi-square test was used to test for trend between a risk count of 0, 1 and 2. Logistic regression was used for multivariate analysis to predict prostate cancer using age (continuous), PSA level at diagnosis/censorship (continuous), and PSAV risk count. A separate model also including family history was performed in the 16 230 (89.1%) men with family history data.

Prostate cancer-free survival curves were generated using the Kaplan–Meier method and were compared by the log-rank test. Receiver operating characteristic (ROC) analysis was used to examine the discrimination of a model including age and PSA level at diagnosis/censorship, and to determine whether the addition of PSAV risk count significantly improved discrimination.

We also performed several subset analyses. Due to the reduction in our PSA level threshold for biopsy (from 4 to 2.5 ng/mL) in 1995, separate logistic regression models were used to examine the utility of the PSAV risk count in the time intervals before and after this change in the study protocol. We used separate logistic regression models to evaluate the PSAV risk count after stratifying the study population into the following total PSA level ranges: <2.6, 2.6–4.0, and >4.0 ng/mL. We performed a separate analysis including the 14 024 men with at least four successive PSA measurements to determine the utility of the PSAV risk count calculated over a longer interval. In this subgroup, we determined the association between risk counts ranging from 0 to 3 with prostate cancer detection. A risk count of 3 indicated that all three serial PSAV measurements exceeded 0.4 ng/mL/year.

Finally, due to the possibility for misclassification in participants who did not undergo a biopsy [17], we performed a subgroup analysis in 1524 men undergoing a first prostate biopsy with at least three PSA measurements before the biopsy, from which to calculate the PSAV risk count. These data were used to examine the association between the PSAV risk count with the presence or absence of prostate cancer on biopsy and with the detection of high-grade prostate cancer.

We used ROC analysis to determine whether the PSAV risk count improved discrimination of high-grade disease. The ROC analysis represents the area under the curve (AUC) of sensitivity vs false-positive rate (1-specificity), and is equivalent to the probability that a predictive model will assign a higher probability of an event to subjects who subsequently have an event.

Because ROC analysis may be insensitive to the addition of new risk factors, unless they have extremely high-risk ratios [18], we also used net reclassification analysis to determine whether the PSAV risk count improved the prediction of high-grade disease using two different criteria (Gleason score ≥7 and Gleason score 8–10). As described by Pencina et al. [19], net reclassification analysis was designed to assess the improvement in prediction using a new risk factor for an event when added to standard risk factors. It is based on the net reclassification improvement (NRI) defined when comparing the estimated probability of having an event for a model with standard risk factors to that of a model which adds the new risk factor. NRI represents the sum of the improvement in sensitivity and improvement in specificity when the new risk factor is added to standard risk factors. By assuming independence between event and non-event subjects, a z test was developed to test whether the null hypothesis that NRI is zero can be rejected.

RESULTS

Table 1 shows the demographic characteristics of the study population. Men diagnosed with prostate cancer were slightly younger at the last PSA measurement (67 vs 68 years, P= 0.003) and were significantly more likely to have a family history of prostate cancer. The PSA level at diagnosis/censorship was significantly higher in men with prostate cancer than those without it, as was PSAV (0.8 vs 0.1 ng/mL/year, P < 0.001).

Table 1.  Demographics of the study population
VariableOverallProstate cancerNo prostate cancerP
Mean (range) age at last PSA measurement, years67 (41–95)67 (41–89)68 (41–95)0.003
Caucasian race, n (%)17 036 (94)938 (83.4)16 098 (94)0.20
Family history of prostate cancer, n (%)2 330 (14)128 (19)2 202 (14)<0.001
Median (range) PSA level, ng/mL1.3 (0.0–164.0)4.1 (0.1–45.3)1.2 (0.0–164.0)<0.001
Median (range) last PSAV, ng/mL/year0.1 (−34.9 to 132.0)0.8 (−32.3 to 38.1)0.1 (−34.9 to 132.0)<0.001

Of the men diagnosed with prostate cancer, 230 (20.4%), 440 (39.1%), and 455 (40.4%) had a PSAV risk count of 0, 1, and 2, respectively. By contrast, most men with no prostate cancer (68%) had a PSAV risk count of 0, 4755 (28%) had a risk count of 1, and only 674 (4%) had a risk count of 2 (P < 0.001, comparing trend in prostate cancer vs no prostate cancer).

Table 2 shows the results of multivariable analysis in the overall study population. After adjusting for age and PSA level, the PSAV risk count was associated with an 8.2-fold increased odds ratio (OR) for prostate cancer (95% CI, 7.0–9.6, P < 0.001). The results were similar with the addition of family history to the model (family history OR 1.3, 95% CI 1.1–1.6, P= 0.02).

Table 2.  Multivariable analysis for the prediction of prostate cancer in the overall study population
VariableOR (95% CI)P
Age0.99 (0.98–1.00)0.13
PSA level1.25 (1.22–1.28)<0.001
PSAV risk count 2 vs 0,18.2 (7.0–9.6)<0.001

On ROC analysis to predict overall prostate cancer, a model including PSA level and age had an AUC of 0.89, which improved significantly with the addition of the PSAV risk count (AUC 0.90, P= 0.026). Overall, a PSAV risk count of 2 was associated with 40% sensitivity, 96% specificity, 40% positive predictive value, and 96% negative predictive value.

Kaplan–Meier survival analysis for freedom from prostate cancer diagnosis is shown in Fig. 1. The 7-year screening study-detectable cancer-free survival was 96% for men with a PSAV risk count of 0–1, vs 61% with a risk count of 2 (P < 0.001).

Figure 1.

Kaplan–Meier curve for cumulative freedom from prostate cancer, stratified by a PSAV risk count of 0–1 vs 2.

The results of subset analyses before and after 1995 (when the total PSA level threshold for biopsy was reduced from 4.0 to 2.5 ng/mL) are shown in Table 3. Irrespective of the time interval, both PSA level and PSAV risk count were significantly associated with prostate cancer risk after adjustment for age.

Table 3.  Multivariable models to predict prostate cancer in the subsets (a) before and (b) after the PSA threshold for biopsy was lowered in 1995
VariableOR (95% CI)P value
Before PSA threshold for biopsy was lowered in 1995
 Age1.02 (1.00–1.04)0.05
 PSA level1.18 (1.12–1.25)<0.001
 PSAV risk count 2 vs 0,17.8 (5.6–10.9)<0.001
After the PSA threshold for biopsy was lowered in 1995
 Age0.99 (0.98–1.00)0.09
 PSA level1.22 (1.19–1.26)<0.001
 PSAV risk count 2 vs 0,16.9 (5.7–8.4)<0.001

Multivariate models were also used after stratification by the total PSA level. In the strata with PSA levels of <2.6, 2.6–4.0, and >4.0 ng/mL, the PSAV risk count maintained a significant association with prostate cancer detection, with ORs of 4.2, 2.3 and 3.6 (all P < 0.001), respectively.

Among 14 024 men with at least four PSA measurements, 142 (15%) men with prostate cancer had a risk count of 3, compared with 87 (1%) without prostate cancer (P < 0.001). On multivariable analysis with age and PSA level, a risk count of 3 (compared to a risk count of 0–2) was associated with a 7.4-fold increased OR for prostate cancer (95% CI, 5.5–10.0, P < 0.001).

Finally, in 1524 men undergoing a first prostate biopsy, 390 (25.6%) were diagnosed with prostate cancer. The Gleason score was 6 in 318 (81.5%), 7 in 48 (12.3%), 8 – 10 in 21 (5.4%), and missing in three (0.8%) men.

Before biopsy, 140 (36%) of men diagnosed with prostate cancer had a PSAV risk count of 2, compared with 245 (22%) men with a negative biopsy (P < 0.001). In this subgroup, a risk count of 2 before a first prostate biopsy was associated with 1.9-fold increased odds (95% CI 1.5–2.4, P < 0.001) for detecting prostate cancer on biopsy after adjusting for age and PSA level.

Table 4 shows the robust univariate association between the PSAV risk count and biopsy Gleason grade (P= 0.007). On multivariate analysis, the PSAV risk count (OR 5.4, 95% CI 2.2–13.3, P < 0.001) was more strongly associated with Gleason 8–10 prostate cancer than PSA level (OR 1.07, 95% CI 0.95–1.20, P= 0.29) or age (OR 1.06, 95% CI 0.99–1.13, P= 0.07).

Table 4.  The association between the PSAV risk count and Gleason grade on first prostate biopsy
Risk countGleason 6, n (%)Gleason 7, n (%)Gleason 8–10, n (%)P
2106 (33.3)20 (41.7)13 (61.9)0.007
0,1212 (66.7)28 (58.3)8 (38.1) 

Figure 2 shows ROC analysis for the prediction of Gleason 8–10 prostate cancer on biopsy. A model with PSA level and age had an AUC of 0.625 that increased to 0.725 with the addition of the PSAV risk count (P= 0.031). Finally, Fig. 3 shows the results of net reclassification analysis. Overall, the PSAV risk count (using a scale of 0–2) significantly altered the probability of Gleason score ≥7 (z= 6.5, P < 0.001) and Gleason score ≥8 prostate cancer (z= 3.5, P < 0.001), beyond age and the total PSA level. Similarly, in the subset with three serial PSAV measurements, the addition of the risk count (using a scale of 0–3) to a model with age and total PSA level significantly altered the probability of Gleason score ≥7 (z= 6.9, P < 0.001) and Gleason score ≥8 (z= 3.6, P < 0.001) prostate cancer.

Figure 2.

ROC analysis for the prediction of Gleason 8–10 prostate cancer on needle biopsy in a base model with PSA level and age, and with the addition of the PSAV risk count to the model.

Figure 3.

Net reclassification analysis comparing the predicted probability of Gleason ≥7 and Gleason ≥8 prostate cancer based upon a model with PSA level and age (x-axis), vs the expanded model also including the PSAV risk count (y-axis).

DISCUSSION

PSA screening has somewhat fallen victim to its own success. The widespread use of PSA testing has led to a considerable stage migration, and prostate cancer death rates have decreased more than for any other cancer. However, the previous concern over high rates of advanced disease at diagnosis has been replaced by concern over the identification of indolent tumours through screening.

Despite considerable refinement in RP and radiation therapy over the past few decades [20], all forms of definitive prostate cancer treatment are associated with potential morbidity [21]. An alternative management approach for low-risk prostate cancer is active surveillance. However, most men diagnosed with low-risk prostate cancer in the USA undergo definitive therapy within 12 months of diagnosis [22]. Additionally, active surveillance is controversial due to the limitations of assessing ‘true’ disease extent and grade on prostate biopsy, and heterogeneity in existing surveillance protocols for patient selection and triggers for intervention [23]. These complex issues make the diagnosis of cancers that may not cause harm an important concern.

A possible aid in reducing diagnosis of potentially harmless cancer is to increase the specificity of screening for clinically significant prostate cancer through the use of PSA kinetics. D'Amico et al. [10,11] previously showed that PSAV predicted the risk of prostate cancer-specific mortality after treatment. Also, in the Baltimore Longitudinal Study of Aging, a PSAV of >0.35 ng/mL/year more than 10–15 years before diagnosis predicted life-threatening prostate cancer [24]. More recently, in our surgical series, we found that men with a single preoperative PSAV of <0.4 ng/mL/year were twice as likely to have pathologically ‘insignificant’ prostate cancer in the RP specimen [15].

However, other studies have reported differing results. For example, Wolters et al. [26] examined the utility of PSAV to predict ‘insignificant’ prostate cancer based upon a nomogram probability in men with a PSA level of ≥3 ng/mL who underwent biopsy in the second round of the ERSPC Rotterdam. Unlike the present study, the authors calculated PSAV using two PSA measurements separated by 4 years. Although PSAV was significantly associated with ‘significant’ prostate cancer on univariate analysis, it lost significance in the multivariable model. Possible explanations for the conflicting findings in previous studies include the use of different criteria to define ‘insignificant’ disease and different methods of PSAV calculation.

In the context of prostate cancer screening, men frequently present to the clinic with a prior PSA history of variable duration, and it is unclear how to best organise these data in a clinically meaningful fashion. To this end, Carter et al. [12] proposed counting the number of times that serial PSAV measurements exceed a specific threshold (0.4 ng/mL/year). In their unscreened population from the Baltimore Longitudinal Study of Aging, the PSAV risk count was significantly associated with the probability of life-threatening prostate cancer.

In the present study, we tested the utility of the PSAV risk count for prostate cancer screening. After adjusting for PSA level and age, men with two serial PSAV measurements of >0.4 ng/mL/year (risk count 2) had an 8.2-fold increased risk of prostate cancer compared with those with a risk count of ≤1. The statistically significant independent association between the PSAV risk count and prostate cancer was maintained in subset analyses stratifying by the study interval or total PSA ranges. Additionally, men with a PSAV risk count of 2 before prostate biopsy had a 5.4-fold increased risk of Gleason 8–10 prostate cancer on multivariable analysis. Both ROC analysis and net reclassification analysis showed a significant improvement in the prediction of high-grade prostate cancer using the PSAV risk count, as compared to PSA level and age alone. These findings suggest that a prostate biopsy trigger based on the PSAV risk count might reduce the number of unnecessary prostate biopsies, the diagnosis of low-grade, indolent cancer, and correspondingly the potential over-treatment of these cancers.

A limitation of the present study is the potential for verification bias [26]. To reduce the potential for misclassification, we performed subgroup analysis including only those individuals who underwent a prostate biopsy, confirming a robust independent association between the PSAV risk count and histological findings.

Another limitation is that despite the large sample size (18 214), an additional 17 322 men were excluded who did not have sufficient PSA level measurements before prostate cancer diagnosis or censorship to enable a PSAV risk count calculation. The requirement for multiple PSA values introduces a bias, as suggested by the higher prostate cancer detection rate in the excluded population. Furthermore, prostate cancer diagnosed by serial screening is more likely to have favourable features [27], as shown by significantly higher median PSA levels and a greater proportion of palpable disease among the excluded men from the present screening study.

The requirement for multiple PSA values will also preclude the use of the PSAV risk count for patients who present to the clinic with a limited PSA history; however, when feasible, the present results confirm that the PSAV risk count has a robust association with prostate cancer risk, and especially with high-grade disease.

Another limitation is that most men were Caucasian, potentially limiting the generalizability of these findings to other ethnic groups. Overall, as in previous studies [25], there are significant drawbacks to the examination of PSA kinetics in screening programmes not designed to evaluate this issue. Future studies are warranted to prospectively evaluate the use of the PSAV risk count as an indication for prostate biopsy, as well as its possible role in multivariable predictive tools.

In conclusion, men with two successive PSAV measurements of >0.4 ng/mL/year (risk count 2) had about an 8-fold increased risk of prostate cancer, controlling for age and serum PSA concentration. Among men undergoing prostate biopsy, there was a 5-fold increased risk of Gleason >8 prostate cancer with a PSAV risk count of 2 when compared with a PSAV risk count of 0 or 1. Furthermore, the addition of the PSAV risk count to a model with PSA level and age, significantly improved the discrimination for Gleason ≥8 prostate cancer on biopsy and led to NRI. These results suggest that the PSAV risk count may be useful to reduce the number of unnecessary prostate biopsies and the diagnosis of low-risk prostate cancer when compared with PSA alone.

ACKNOWLEDGEMENTS

We would like to thank Dr H. Ballentine Carter for critical review of the manuscript.

The research of W.J.C. is supported in part by the Urological Research Foundation, Prostate SPORE grant (P50 CA90386-05S2) and the Robert H. Lurie Comprehensive Cancer Center grant (P30 CA60553). E Jeffery Metter is supported by the Intramural Research Program of the National Institute of Health, National Institute on Aging.

CONFLICT OF INTEREST

William J Catalona is an investigator, consultant and receives honoraria for talks for Beckman Coulter, Inc. (a manufacturer of PSA tests); he is an investigator, consultant and research support for OHMX; he is a consultant for Nanosphere; and he is an investigator and research support for deCODE Genetics.

Ancillary