Up to 17% of men with a prostate-specific antigen (PSA) level below the accepted prostate biopsy cutoff of 2.5 ng/mL may have prostate cancer. Because identification of these patients represents a difficult task, we assessed the ability of percent free PSA to discriminate between benign and malignant prostate biopsy outcomes in men with PSA ≤2.5 ng/mL.
Between 1999 and 2006, 543 men with a PSA ≤2.5 ng/mL were referred for initial prostate biopsy. Age, total PSA, percent free PSA, and digital rectal examination findings represented predictors of prostate cancer at biopsy in logistic regression models. The area under the receiver operating characteristics curve (AUC) quantified the discriminative ability of the predictors. The pathological characteristics of the detected cancers were assessed in individuals treated with radical prostatectomy.
Of all, 23% had prostate cancer on biopsy, 16.5% of patients treated with radical prostatectomy had pT3 stage, and 35.6% had a pathological Gleason score of 3 + 4 or higher. The most accurate predictor of prostate cancer on biopsy was percent free PSA (0.68) versus age (0.50), total PSA (0.57), or rectal examination findings (0.58). Of patients with percent free PSA below 14%, 59% had prostate cancer. In multivariate models, percent free PSA (P < .001) and rectal examination findings (P = .001) were the only independent predictors of prostate cancer. The combined AUC of all predictors (0.69) was not significantly (P = .7) higher than that of percentage of free PSA alone (0.68).
Total serum prostate-specific antigen (tPSA) is widely used for prostate cancer screening and prostate cancer early detection programs in Western countries. Recently, urological and oncological associations lowered the recommended tPSA cutoff for prostate biopsy from 4 to 2.5 ng/mL.1, 2 This was based on the observation that a significant number of patients with a tPSA lower than 4 ng/mL may harbor prostate cancer.3, 4 Indeed, the prostate cancer prevention trial (PCPT) findings indicated that prostate cancer might be detected in 17% of patients with unremarkable digital rectal examination findings and a tPSA of 1.1 to 2.0 ng/mL, and in 23.9% of patients with a tPSA of 2.1 to 3.0 ng/mL.4 These findings demonstrated that there is no safe tPSA level below which prostate cancer does not exist. Moreover, tPSA in the ≤2.5 ng/mL range has very limited ability to discriminate between prostate cancer and benign findings at prostate biopsy.5 In consequence, accurate identification of men at risk of prostate cancer in the tPSA ≤2.5 ng/mL range represents a challenge, and novel biomarkers are needed for this subgroup.
On the basis of its excellent performance characteristics in tPSA ranges above 2.5 ng/mL, we postulated that the ratio between free PSA and tPSA (%free PSA) may be able to enhance the predictive ability of tPSA in the tPSA range ≤2.5 ng/mL.6–9 Therefore, we investigated the value of %free PSA in predicting prostate cancer at prostate biopsy in men with a tPSA ≤2.5 ng/mL and compared it to other common prostate cancer predictors. Our analyses addressed potential %free PSA cutoffs that could help to identify men at an elevated risk of prostate cancer, despite tPSA ≤2.5 ng/mL. Finally, we assessed the pathological characteristics of the cancers detected in men with tPSA ≤2.5 ng/mL.
MATERIALS AND METHODS
From January 1999 to June 2006, 7880 consecutive men underwent initial prostate biopsy at 2 European referral tertiary care centers (Hamburg, n = 6028; Milan, n = 1852). All biopsy data were entered prospectively in the institutions' databases. Informed consent was obtained from each patient before prostate biopsy. Patients were referred for biopsy for a variety of reasons. These consisted of suspicious digital rectal examination findings performed by experienced urologists (1.8% with missing data) or of various abnormal PSA patterns. Examples include nonzero tPSA velocity, tPSA values perceived abnormal given the patient's young age, or even a simple tPSA rise above previously stable tPSA values. Unfortunately, because of the referral nature of our population, the search for a systematic pattern of referral in patients with unremarkable digital rectal examination findings was not possible. Patients had tPSA and %free PSA measured before any manipulation of the prostate right before prostate biopsy (5.7% with missing data). In Hamburg, the Abbott Axsym PSA assay (Abbott Laboratories, Abbott Park, Ill) was used for tPSA and %free PSA testing. In Milan, the Elecsys assay (F. Hoffmann-La Roche, Basel, Switzerland) was used. At biopsy, the median number of obtained cores was 10 (range, 6-26 cores).
We restricted our analyses to patients with serum tPSA ≤2.5 ng/mL (n = 543 [6.9%]; Hamburg, n = 493; Milan, n = 50), who then represented the focus of this analysis. Of all men with newly diagnosed prostate cancer in Hamburg (n = 111), 71 (63.9%) underwent a radical prostatectomy, and of all men with newly diagnosed prostate cancer in Milan (n = 14), 8 (57.1%) underwent radical prostatectomy. This allowed the assessment of the pathological characteristics of those cancers. The radical prostatectomy specimens were processed according to the Stanford protocol.
All analyses were performed twice: first on all patients (n = 543) and subsequently on men with unremarkable digital rectal examination findings (n = 399). The study was approved by institutional review boards.
The descriptive statistics were generated for the entire group and for the subgroup of men with unremarkable digital rectal examination findings. Chi-square and independent Student t test tested for potential baseline differences between subgroups, according to delivery of therapy (radical prostatectomy vs no radical prostatectomy). Further analyses consisted of univariate and multivariate logistic regression models predicting prostate cancer at initial biopsy. The discriminative ability of the individual predictor variable and of multivariate models was quantified with the area under the receiver operating characteristics curve (AUC), where a value of 0.50 represents a toss of a coin and a value of 1.0 represents perfect prediction. Two hundred bootstrap resamples were used to reduce the overfit bias of the AUC estimates. In univariate analyses the P value approach, for categorization of continuously coded variables, was used to identify the most informative cutoffs for %free PSA.10 To test for a possible bias of the results in favor of the Hamburg group (n = 493), we tested the identified cutoffs for %free PSA also in the Milan group (n = 50) in isolation.
Calibration plots were used to graphically explore the variables' and models' performance characteristics. The plots show the relation between the predicted probability of prostate cancer at prostate biopsy and the observed rate of prostate cancer. In the entire group, a multivariate model based on %free PSA and digital rectal examination was used to develop a nomogram predicting the individual probability of prostate cancer at initial biopsy. The Mantel-Haenszel test was used to compare the statistical significance of AUC differences between models. All statistical tests were performed using S-PLUS software (S-PLUS Professional, version 1; MathSoft, Seattle, Wash). Two-sided tests and a significance level of .05 were used in all statistics.
The characteristics of the entire group (N = 543) and the group with unremarkable digital rectal examination findings (n = 399) are shown in Table 1. The prevalence of prostate cancer at prostate biopsy was 23.0% (n = 125) and 19.3% (n = 77) for, respectively, the entire group and the group with unremarkable digital rectal examination findings. Of all patients, 36.6% (n = 199) underwent sextant biopsy, and prostate cancer detection rate in these patients was 22.6% (n = 45); 58.2% (n = 316) underwent 10- to 12-core biopsy, and detection rate in these was 23.1% (n = 73). Another 5.2% (n = 28) underwent ≥13-core biopsy, and detection rate in these was 25% (n = 7). There was no statistically significant difference in the detection rates between the biopsy schemes (P = .9). The pathological results of the biopsies are shown in Table 2. Of all patients diagnosed with prostate cancer, 20.8% (n = 26) showed Gleason 4 pattern, and 36.8% (n = 46) had 3 or more positive biopsy cores. Of all patients who had prostate cancer at prostate biopsy, 79 underwent radical prostatectomy (Table 2). Pathological stage was pT3 in 16.5% (n = 13), and pathological Gleason score was 3 + 4 or higher in 35.6% (n = 24). In patients with biopsy-proven prostate cancer, no statistically significant differences were found between those treated with radical prostatectomy versus others, except for age (P < .001).
Table 1. Descriptive Table of the 2 Groups Included in the Study
Entire Group No. (%)
Group With Unremarkable DRE No. (%)
a DRE indicates digital rectal examination; %free PSA, ratio between free prostate-specific antigen (PSA) and total serum PSA.
Total no. of patients
15% to 27%
Prostate cancer at biopsy
PSA ≤0.5 ng/mL
PSA 0.6-1.0 ng/mL
PSA 1.1-2.0 ng/mL
PSA 2.1-2.5 ng/mL
Table 2. Descriptive Table of Patients Diagnosed With Prostate Cancer and of Patients Treated With Radical Prostatectomy for Detected Cancers
Patients Diagnosed With Prostate Cancer No. (%)
Patients Treated With Radical Prostatectomy No. (%)
PSA indicates prostate-specific antigen; %free PSA, the ratio between free PSA and total serum PSA.
Total no. of patients
No. of positive cores
5 or more
% positive cores
0% to 20%
21% to 40%
41% to 60%
61% to 80%
81% to 100%
Table 3 shows the univariate AUC estimates for all considered predictors of prostate cancer at initial biopsy. The AUC estimates are provided for both groups. In univariate analyses, %free PSA was the most informative predictor of prostate cancer at initial biopsy in the entire group (0.68), as well as in the group restricted to men with unremarkable digital rectal examination findings (0.68). Figure 1A depicts the calibration plot of %free PSA within the entire group and Figure 1B the calibration plot in men with unremarkable digital rectal examination findings. Both graphs demonstrate that %free PSA has only minor departures from ideal predictions in the low probability ranges but underestimates the true probability of cancer detection in higher probability ranges.
Table 3. Univariate and Multivariate Logistic Regression Models Addressing Prostate Cancer at Initial Biopsy
Group With Unremarkable DRE
AUC is estimated by 200 bootstrap resamples.
DRE indicates digital rectal examination; AUC, area under the receiver-operater characteristics curve; PSA, prostate-specific antigen; %free PSA, the ratio between free PSA and total serum PSA.
%free PSA (continuous)
%free PSA (categorized)
15% to 27% vs ≤14%
≥28% vs ≤14%
In both groups, the P value–based approach identified the value of 14% as the most informative %free PSA cutoff. The second most informative cutoff was 27%. The use of both cutoffs, relative to continuously coded %free PSA, did not result in significantly different AUC (all P values ≥0.4; entire group, 0.67 vs 0.68; men with unremarkable digital rectal examination findings, 0.68 vs 0.68).
Table 4 shows the prostate cancer prevalence according to the %free PSA cutoffs. The use of the 14% cutoff resulted in 59% and 49% positive biopsy rates, respectively, in all men and in men with unremarkable digital rectal examination findings. For patients with %free PSA between 15% and 27%, the prostate cancer prevalence was 24% and 23%, respectively. Finally for patients with %free PSA >27%, prostate cancer prevalence was 13% and 9%, respectively. Therefore, 2 prostate biopsies would have been needed to detect 1 prostate cancer below the 14% cutoff. Approximately 4 individuals needed to be subjected to prostate biopsy to detect 1 cancer in the %free PSA range 15% to 27%. In the Milan group, the prostate cancer prevalence for the %free PSA cutoffs of ≤14%, 15% to 27%, and ≥28% were, respectively 67%, 25%, and 12%.
Table 4. Prevalence of Prostate Cancer According to the %free PSA Cutoffs Identified in the Dataset and According to the 2 Different Groups
% free PSA ≤14% No. (%)
% free PSA 15%-27% No. (%)
% free PSA ≥28% No. (%)
% free PSA indicates the ratio between free prostate-specific antigen (PSA) and total serum PSA; DRE, digital rectal examination.
Group with unremarkable DRE
Table 3 also shows the multivariate models predicting prostate cancer at prostate biopsy in both groups. The model that included all variables demonstrated an AUC of 0.69 in the entire group and of 0.68 in men with unremarkable digital rectal examination findings. In both models, %free PSA achieved independent predictor status. In models based on the entire group, the variable defining the digital rectal examination findings (suspicious vs unremarkable) also achieved independent predictor status. An AUC of 0.70 was recorded when only %free PSA and digital rectal examination were included. Inclusion of other variables did not result in statistically significant AUC gains (all P values ≥0.7). A nomogram predicting the probability of prostate cancer at initial biopsy in patients with tPSA <2.5 ng/mL was developed based on %free PSA and digital rectal examination (Fig. 2A). Figure 2B shows the calibration plot of this model in the entire group and demonstrates only minor departures from ideal predictions in the low probability ranges but underestimates the true probability of cancer detection in higher probability ranges.
Despite the recent lowering of the tPSA cutoff from 4.0 to 2.5 ng/mL in prostate cancer screening and prostate cancer early detection programs, a significant number of patients with tPSA ≤2.5 ng/mL may harbor prostate cancer if prostate biopsy is performed.1, 2, 4 These individuals represent a particularly challenging subset of patients for physicians who perform PSA screening or prostate cancer early detection. The challenge stems from the knowledge that up to 1 in 5 men with tPSA ≤2.5 ng/mL may have prostate cancer at prostate biopsy, and that tPSA provides only limited ability to stratify the risk of prostate cancer in these men.4, 5
To address this challenge, we explored the ability of %free PSA to improve the prostate cancer risk stratification in patients with tPSA ≤2.5 ng/mL. Moreover, we tested the individual and combined abilities of tPSA, %free PSA, and digital rectal examination findings to predict prostate cancer at initial biopsy. The analyses were performed in 2 patient groups, namely in all patients with tPSA ≤2.5 ng/mL and in patients with tPSA ≤2.5 ng/mL and unremarkable digital rectal examination. The restriction to men with unremarkable digital rectal examination was based on the consideration that a suspicious digital rectal examination alone represents an indication for prostate biopsy.11 Moreover, in clinical practice the most challenging cases consist of men with unremarkable digital rectal examination findings and of tPSA ≤2.5 ng/mL. Therefore, it was of interest to explore the ability of the above predictors in these particular patients. To keep our results applicable to all men with tPSA ≤2.5mg/mL and because suspicious digital rectal examination findings may coexist with a tPSA ≤2.5 ng/mL, we also analyzed the ability of %free PSA and of the other predictors in all men with tPSA ≤2.5 ng/mL regardless of digital rectal examination findings.
In the group that included all men, 1 of 5 patients with a tPSA ≤2.5 ng/mL had prostate cancer at prostate biopsy. This high prostate cancer prevalence confirms the PCPT results, where 17% of patients with a tPSA from 1.1 to 2.0 ng/mL harbored prostate cancer.4 This important observation indicates that the high prevalence of prostate cancer in low PSA ranges can be expected not only in screening or observational populations but also in referral populations. The prostate cancer incidence within our population exceeded that of the PCPT population. The referral nature of our population may explain this discrepancy, as referred men might be expected to have a higher prostate cancer risk than men within an observational population.12 The inclusion of patients with suspicious digital rectal examination within our group may also have increased prostate cancer prevalence. Indeed, restriction of our group to patients with an unremarkable digital rectal examination resulted in a decrease of prostate cancer prevalence from 23% to 19%. Finally, the higher prevalence of prostate cancer at prostate biopsy within our study may be explained by more extensive biopsy schemes, relative to the PCPT study. The PCPT patients underwent only sextant biopsies, which can yield substantially lower prostate cancer detection rates relative to the systematic 10-core scheme that was used in the majority of patients within our study.4, 13 However, it is of note that detection rates were not significantly different in our analysis when comparing the detection rate of sextant biopsy schemes with 10- to 12-core and >12-core biopsy schemes (P = .9).
These results emphasize that, regardless of the screening or referral nature of a population, a non-negligible risk of prostate cancer can be observed in low PSA ranges that might be regarded as normal based on the existing and generally accepted cutoff of 2.5 ng/mL.1, 2
The novelty of our findings resides in the identification of a biomarker, which can accurately stratify the risk of prostate cancer at prostate biopsy in men with PSA values ≤2.5 ng/mL. The %free PSA cutoffs of ≤14%, 15% to 27%, and ≥28% can discriminate between men with probabilities of prostate cancer that differ drastically between 59% (%free PSA ≤14%) and 9% (%free PSA ≥28%, Table 4). A certain bias of these cutoffs in favor of the Hamburg group might be present, as the Milan group represented only 9% of the entire population, and as a different PSA assay was used for %free PSA measurement. However, the detection rates in the Milan group were very similar to the detection rates in the entire group (detection rate for %free PSA cutoffs of ≤14%, 15%-27%, and ≥28%: entire group, 59%, 24%, and 13% vs Milan group: 67%, 25%, and 12%, respectively). Our data suggest that %free PSA can provide valuable information when tPSA fails to do so. Moreover, our data suggest that this ability is not dependent on a certain PSA assay. Despite the use of 2 different PSA assays, the identified %free PSA cutoffs allowed discrimination in both groups between patients with high, intermediate, and low risk of prostate cancer. It remains to be seen if this is also true for other assays. The identified cutoffs indicate that 1 of 2 men with %free PSA ≤14% will have prostate cancer, if a prostate biopsy is performed. Conversely, only 1 of 10 men will have prostate cancer, if %free PSA is 28% or higher. If prostate cancer prevalence predictions are made according to these cutoffs, an AUC of up to 0.68 can be expected. Similar AUCs can be expected if %free PSA is used in an unaltered, continuously coded format. Therefore, the choice of the categorized or continuous format for %free PSA may be left to the discretion and the preference of the clinician. Some may prefer the use of 3 simple categories. Others may prefer to quantify the risk on a continuous scale, as is provided in the nomogram format (Fig. 2A). Either model can be used to select candidates for prostate biopsy, as well as for patient counseling. It is noteworthy that the addition of other clinical predictors to %free PSA did not result in improved AUCs. Moreover, %free PSA maintained its discriminative ability even when the analysis was restricted to men with unremarkable digital rectal examination findings. It needs to be emphasized that the restriction to tPSA levels ≤2.5 ng/mL spuriously decreased the predictive value of tPSA, as the restriction of a continuously coded variable to a narrow range decreases its predictive abilities. In recent studies, unrestricted tPSA remained an informative predictor of prostate cancer even in low tPSA ranges.14, 15 Therefore, tPSA remains the base of prostate cancer early detection. Its accuracy, however, can be improved when %free PSA is also considered.
Our AUC estimations corroborate the findings of Thompson et al, who used the observation arm of the PCPT to develop a model that predicts prostate cancer at biopsy.16 The tool yielded an AUC of 0.70 in the original population and of 0.66 in an external validation and provided only a marginally higher AUC than tPSA alone (0.68).16, 17 Our multivariate models showed comparable AUCs (0.68-0.70). Relative to our model, the inclusion of family history and race as well as the use of unrestricted PSA values that ranged up to 287 ng/mL within the PCPT may have contributed to the good performance of the Thompson et al model. Information on family history was not available in our study and, as the vast majority of men were Caucasians, the variable race was not analyzed in our group. The advantage of the PCPT model resides in its applicability to screening populations. Conversely, the advantage of our model resides in its applicability to referral populations. Thus, both models are complementary. However, our findings as well as those of Thompson et al demonstrate that prediction of prostate cancer at biopsy in low PSA ranges remains a challenging task.
The value of %free PSA in our population corroborates the report of Catalona et al and of other investigators, who also found %free PSA highly useful to identify patients with prostate cancer among men with low tPSA (2.5 to 4 ng/mL).18 The AUC in these reports ranged from 0.59 to 0.72.5, 9 Others could show that the combination of tPSA with %free PSA, free PSA, and human kallikrein 2 significantly improved the predictions of biopsy outcome, especially in older patients.19 Our data also corroborate several previous analyses that examined the performance of %free PSA at initial biopsy, repeat biopsy, and saturation biopsy.12, 20–22 In these 3 analyses, %free PSA represented the strongest predictor. The strength of our study relative to the previous studies relates to the consideration of very low tPSA values (≤2.5 ng/mL), as others were restricted to tPSA values higher than 2.0 ng/mL.5–7, 9, 23, 24 Therefore, our findings can be viewed as an extension of the previous reports, which extends the proof of %free PSA usefulness across virtually the entire range of tPSA values.
Because prostate cancer detection in tPSA ≤2.5 ng/mL may result in detection of clinically insignificant prostate cancer, we explored the pathological characteristics of cancers that were treated with radical prostatectomy at the participating institutions. The analysis demonstrated that, despite the favorable tPSA range, 16.5% of cancers had established extracapsular extension or seminal vesicle invasion. Moreover, 35.6% of all patients who underwent radical prostatectomy had a pathological Gleason score of 3 + 4 or higher (Table 2). There was no statistically significant difference in biopsy pathology between the patients treated with radical prostatectomy and those only diagnosed with prostate cancer, which suggests that the patients treated with radical prostatectomy might be considered representative for the entire group diagnosed with prostate cancer. Unfortunately, cancer volume was not available for all patients; therefore, no information on the possible clinical significance of the detected cancers could be provided. Despite this, the surprisingly elevated proportion of unfavorable pathologic characteristics indicates that low tPSA values are neither indicative of nonexistent prostate cancer risk nor of favorable prostate cancer pathology.
Our study is not devoid of limitations. Because of the referral nature of our population, no systematic pattern of referral in patients with unremarkable digital rectal examination findings could be identified. It is reassuring that the majority of patients included in the current study had unremarkable digital rectal examination findings (74%). Therefore, the majority of biopsies were not performed because of suspicious digital rectal examination findings. Moreover, our analyses that were restricted to patients with unremarkable digital rectal examination findings indicated the same AUC for %free PSA (0.68 vs 0.68) and a similar rate of prostate cancer (19.3% vs 23.0%), as recorded within the entire group. Our dataset did not provide PSA velocity or PSA doubling time as predictors, which might be considered another limitation. These 2 variables may represent useful tools for risk stratification of prostate cancer in low PSA ranges, and a direct comparison between %free PSA and these modified PSA forms would have been interesting.25, 26 However, previous reports on PSA velocity demonstrated that PSA velocity was not an independent predictor of prostate cancer at biopsy in PCPT patients.16 In consequence, PSA velocity was not included in the Thompson et al model.16 Moreover, PSA velocity and PSA doubling time are difficult to calculate and require long-term PSA follow-up.27 In consequence, this information will not be available for men who present for their first PSA measurement. The lack of a true external validation of the models is another limitation. A model development in the Hamburg dataset and a model validation in the Milan dataset might have been an option; however, because of the small number of patients in the Milan dataset, a reliable validation was not warranted. We therefore performed model development and internal validation by bootstrapping in the combined dataset only. Finally, no centralized pathology review was available, and variability between the centers cannot be excluded. However, it is unlikely that a centralized pathology review would have substantially changed the results, as the main endpoint of the current study was the detection of prostate cancer at biopsy. Despite these limitations, our study provides important novel findings related to prostate cancer screening and early detection of men with tPSA ≤2.5 ng/mL.
The risk of prostate cancer cannot be ruled out in patients with tPSA ≤2.5 ng/mL. The only informative predictors of prostate cancer at initial biopsy in these men are %free PSA and digital rectal examination. Therefore, the routine use of %free PSA should be strongly recommended in prostate cancer screening or early detection efforts to better risk stratify the probability of prostate cancer at biopsy.