An artificial neural network considerably improves the diagnostic power of percent free prostate-specific antigen in prostate cancer diagnosis: Results of a 5-year investigation



Our study was performed to evaluate the diagnostic usefulness of %fPSA alone and combined with an ANN at different PSA concentration ranges, including the low range 2–4 ng/ml, to improve the risk assessment of prostate cancer. A total of 928 men with prostate cancer and BPH without any pretreatment of the prostate in the PSA range 2–20 ng/ml were enrolled in the study between 1996 and 2001. An ANN with input data of PSA, %fPSA, patient's age, prostate volume and DRE status was developed to calculate the individual's risk before performing a prostate biopsy within the different PSA ranges 2–4, 4.1–10 and 10.1–20 ng/ml. ROC analysis and cut-off calculations were used to estimate the diagnostic improvement of %fPSA and ANN in comparison to PSA. At the 90% sensitivity level, %fPSA and ANN performed better than PSA in all ranges, enhancing the specificity by 15–28% and 32–44%, respectively. For the low PSA range 2–4 ng/mL, we recommend a first-time biopsy at an ANN specificity level of 90%. For PSA 4–10 ng/mL, we recommend a first-time biopsy based on the ANN at the 90% sensitivity level. Use of an ANN enhances the %fPSA performance to further reduce the number of unnecessary biopsies within the PSA range 2–10 ng/ml. © 2002 Wiley-Liss, Inc.

Prostate cancer is the most common neoplasia among men in the Western world and PSA is recognized as the best marker for its early detection.1 Due to lack of specificity, various techniques, such as PSA density, velocity and age-specific ranges, have been developed but are reported to be only partially successful.1 Measurements of the molecular forms of PSA improve specificity over tPSA alone.2, 3 Approximately 20–25% of all biopsies could be avoided using %fPSA in the tPSA range 4–10 ng/ml4, 5 as well as for tPSA values lower than 4 ng/ml.6, 7, 8 For further enhancement of %fPSA specificity to differentiate between prostate cancer and benign prostate diseases, logistic regression9, 10 and ANNs have been successfully used.11, 12 Briefly, these models can predict diagnostic outcome using a variety of factors. Based on the data of 28 studies, both methods perform equally, especially for large cohorts.13 However, ANNs can predict an outcome for an individual patient, which cannot currently be done with traditional statistics and they can better handle a greater number of variables with more nonlinear relations.14 In a study of 656 men within the 4–10 ng/ml tPSA range, Finne et al.11 demonstrated an advantage of ANNs compared to logistic regression, avoiding approximately one-third of unnecessary biopsies at 95% sensitivity. For the lower tPSA range 2.6–4 ng/ml, Babaian et al.12 used a combination of 3 different ANNs and saved 63.6% of all unnecessary biopsies. To date, no study has evaluated 1 valid, clinically practicable ANN on %fPSA data for the whole tPSA range 2–10 or 2–20 ng/ml.


ANN, artificial neural network; AUC, area under the receiver operating characteristic curve; BPH, benign prostate hyperplasia; DRE, digital rectal examination; MAb, monoclonal antibody; PSA, prostate-specific antigen; %fPSA, percent free PSA; ROC, receiver operating characteristic; tPSA, total PSA; TRUS, transrectal ultrasound.

We used data from a 5-year period of %fPSA measurements with only 1 assay over the whole period with the following aims: (i) to evaluate the diagnostic power of %fPSA on different tPSA ranges, including the low range 2–4 ng/ml, to differentiate between prostate cancer and benign prostate diseases; (ii) to develop an ANN based on %fPSA combined with patient's age, prostate volume (measured by TRUS) and DRE status to improve the prostate cancer detection rate; and (iii) to evaluate this ANN for clinical use and to calculate the individual patient's risk before performing a prostate biopsy.


Patient selection

A total of 928 men (40–86 years old) with no pretreatment of the prostate, in the tPSA range 2–20 ng/ml, were enrolled in the study between March 1996 and March 2001. The distribution of prostate cancer patients (n = 606) and patients with BPH (n = 322) in the tPSA ranges 2–4, 4.1–10 and 10.1–20 ng/ml is shown in Table I. All men were investigated at the Department of Urology or the affiliated outpatient department at the University Hospital Charité (Berlin, Germany). Indications for referral were PSA elevations, BPH symptoms, abnormal DRE or biopsy-proven prostate cancer, which explains the high number of prostate cancer patients. Therefore, this population does not represent a typical urologic referred population or a screening population.

Table I. Patients Investigated in the tPSA Ranges 2–4, 4.1–10, 10.1–20 and 2–20 ng/ml1
tPSA range (ng/ml)All patientsProstate cancer patientsBPH patients
NumbertPSA (ng/ml)fPSA (%)NumbertPSA (ng/ml)fPSA (%)NumbertPSA (ng/ml)fPSA (%)
  • 1

    Median tPSA concentrations and %fPSA values as well as their significances (*p < 0.05, **p < 0.01; Mann-Whitney U-test) between prostate cancer patients and BPH patients are given.


Prostate cancer group

Of the 606 prostate cancer patients (median age 65 years, range 40–86, median prostate volume 30 cm3), 306 underwent either radical retropubic, perineal or laparoscopic prostatectomy and were diagnosed histopathologically by microscopic examination of prostatic specimens. The remaining 300 cancer patients underwent ultrasound-guided sextant or octant prostate biopsy and received either hyperthermia in combination with radiation therapy, radiation therapy and/or antiandrogenic medication or watchful waiting. Cancer stage was assigned according to the revised TNM system from 1997 and histologic grade was classified as 1, 2 or 3, as previously described.15 The pathologic stages and grades of the 306 patients receiving operative therapy were as follows: pT1pN0M0, n = 7; pT2pN0M0, n = 174; pT3pN0M0, n = 119; pT4pN0M0, n = 6; G1, n = 10; G2, n = 184; G3, n = 112. Gleason grades were available for 217 of the 306 patients with radical prostatectomy and the distribution was as follows: Gleason sum 2–6, n = 108; Gleason sum 7, n = 66; Gleason sum 8–10, n = 43. The remaining 300 prostate cancer patients were clinically staged as follows: T1, n = 46; T2, n = 139; T3, n = 114; T4, n = 1. Pathologic features on biopsy were as follows: G1, n = 45; G2, n = 198; G3, n = 55; grading not available, n = 2. Diagnostic pelvic lymph node dissection was performed on 485 of the 606 prostate cancer patients but not in the remaining 121 patients. The nodal status of these patients was as follows: pN0, n = 460; pN1, n = 21; pN2, n = 4.

BPH group

A total of 322 BPH patients (median age 65 years, range 43–84, median prostate volume 40 cm3) were included in this group. Histologic confirmation was obtained for 292 patients (90.6%) by either transurethral resection of the prostate (n = 118), open adenomectomy (n = 18) or transrectal ultrasound-guided sextant or octant biopsy (n = 156). A further differentiation of this group based on the number of biopsy sessions was not performed. The diagnosis of the remaining 30 patients (9.4%) was established clinically by DRE (enlarged, confined, indolent prostate with smooth surface and no indurations) and TRUS (enlarged gland without hypoechoic areas suggestive of cancer). Clinically confirmed BPH patients had statistically significant lower tPSA values (median 3.95 vs. 6.65 ng/ml); but age, prostate volume and DRE status showed no differences from the histologically confirmed group. Since there were no statistical differences between histologically confirmed BPH patients and all BPH patients, the analysis includes the entire BPH group.

PSA assays

All serum samples were drawn before any prostate manipulation (or at least 3–4 weeks after an earlier manipulation) and centrifuged within 2–3 hr after sampling. Samples were analyzed immediately or stored at –20°C for no longer than 48 hr before assay. Total and free PSA were measured with the Immulite PSA and Free PSA kits (Diagnostic Products, Los Angeles, CA). Assays were solid-phase, 2-site, sequential chemiluminescent immunometric tests, which are automatically performed on the Immulite automated analyzer with analytical sensitivities of 0.02 and 0.03 ng/ml for fPSA and tPSA, respectively. The tests use both polyclonal antibodies and MAbs specific for PSA or an anti-PSA MAb specific only for fPSA. The analytical performance and comparisons to other PSA tests were described earlier.15, 16

Prostate volume and DRE

Prostate volume was determined by TRUS (Combison 330; Kretz Technik, Zipf, Austria) using the prolate ellipse formula (height × width × length × 0.52). A DRE finding nonsuspicious for cancer was defined as negative (0) and a finding suspicious for cancer was positive (1).

Statistical methods

Conventional statistical calculations were performed using the statistical software SPSS 10.0 for Windows and SigmaPlot 2001 for Windows (SPSS, Chicago, IL). The nonparametric Kruskal-Wallis test of variance, the Mann-Whitney U-test, the logistic regression analysis as enter- and forward-method and the rank correlation coefficients according to Spearman (rs) were calculated. The diagnostic validity of tPSA, %fPSA and ANN was evaluated by ROC curve analysis.17 The software GraphROC 2.1 for Windows was used for calculation of AUC.18 Significance was defined as p < 0.05.


The ANN was used to enhance the classification resolution in prostate cancer detection within the tPSA range 2–20 ng/ml. As described, the input layer consists of the 5 variables tPSA, %fPSA, patient's age, prostate volume and DRE status. The 3 variables tPSA, %fPSA and age were available for all patients, whereas data on prostate volume and DRE status were available from 884 of the 928 patients (95.3%) or 892 of all patients (96.1%). The 69 patients (7.4%) with 1 or 2 missing values were excluded from ANN analysis. ANN comparisons including ROC curves were performed for the 859 patients (92.6%) having all input data accessible. The hidden layer used 3 neurons and the output layer had 1 neuron representing the output value as the probability of prostate cancer.

ANN models were constructed with the SPSS Extramodul Neural Connection 2.0. A Borland Delphi 5 computer program called ProstataClass was constructed for an input of the 5 variables and an outcome of a value between 0 and 1, which is the cancer probability. The developed back-propagation ANN is a feed-forward network. The activation function of the 3 hidden layers and the output layer was the hyperbolic tangent function, which produces output values between –1 and 1.


PSA and %fPSA

The distribution of the 928 patients, including their median tPSA values within the different ranges 2–4, 4.1–10, 10.1–20 and 2–20 ng/ml, is shown in Table I. tPSA was significantly lower for BPH patients in the 4.1–10 (p = 0.018), the 2–10 and the 2–20 (p < 0.001) ng/ml tPSA ranges but not in either the 2–4 (p = 0.17) or the 10–20 (p = 0.059) ng/ml tPSA ranges. A study of tPSA and %fPSA values from March 1996 to March 2001 with 5 subsets of 1 year each revealed no differences between years but a slight trend of lower tPSA values in BPH patients since 1997. The number of patients increased continuously over the 5-year period (1996–1997, n = 114; 1997–1998, n = 159; 1998–1999, n = 171; 1999–2000, n = 231; 2000–2001, n = 253).

%fPSA was significantly lower (p < 0.001) in prostate cancer patients for all analyzed tPSA ranges including 2–4 ng/ml. All median %fPSA values for the different tPSA ranges are shown in Table I.

In all 606 prostate cancer patients, %fPSA was negatively correlated to tPSA (rs = –0.317, p < 0.01). There was also a negative correlation for all 322 BPH patients (rs = –0.224, p <0.01).

Age, prostate volume and DRE status

Age was either not statistically different or comparable between BPH (median age 64–67) and cancer (median age 64–65) patients in the different tPSA ranges.

Prostate volume was significantly higher in BPH patients for all subgroups. Prostate volume as an influencing factor of %fPSA was analyzed for the tPSA ranges 2–20 and 4.1–10 ng/ml. %fPSA was positively correlated with prostate volume in both prostate cancer and BPH patients (rs = 0.29, p < 0.01; rs = 0.3, p < 0.01) for the entire tPSA range. In the 4.1–10 ng/ml tPSA range, the correlation did not change for prostate cancer patients but in BPH patients the correlation was stronger (rs = 0.44, p < 0.01). Using 40 cm3 to distinguish between small and enlarged prostate glands, we found in both subgroups of patients with prostate volume ≤40 cm3 (n = 613) and >40 cm3 (n = 271) a significant difference in %fPSA values between BPH and cancer patients. Median %fPSA values were 13.2% for BPH and 8.2% for prostate cancer in the group with small glands and 16.7% for BPH and 12.2% for cancer in the group with enlarged glands.

An abnormal DRE was reported in 70.7% (65–76%) of prostate cancer patients and in 14.6% (11–17%) of BPH patients. In all subgroups, the number of patients with suspicious DRE status was significantly smaller among BPH patients than among prostate cancer patients. The difference between cancer and BPH was evident in patients with abnormal DRE (8.8% vs. 11.4%, p = 0.012, n = 457) and with nonsuspicious DRE (9.5% vs. 15.7%, p < 0.001, n = 435).

Stage and grade

Analysis of the 306 cancer patients undergoing radical prostatectomy revealed a significant difference between pT2 and pT3 stage patients for the median %fPSA (8.7% vs. 7.0%, p = 0.005). However, comparing the grade 2 and grade 3 tumors, the p value did not reach significance (median %fPSA 8.7% vs. 7.3%, p = 0.05).

%fPSA cut-offs

For all tPSA ranges (2–4, 4.1–10, 10.1–20, 2–10 and 2–20 ng/ml), specificity levels at 90% and 95% sensitivity and sensitivity levels at 90% and 95% specificity for the absolute %fPSA values were calculated. These %fPSA values are the so-called cut-offs at specific points (e.g., 90% sensitivity or specificity). There is a clearly visible downward trend of %fPSA with higher tPSA values, which is shown in Tables II and III (column 6). This must especially be considered when using the 95% or 90% sensitivity cut-off for the ranges 2–4 and 4.1–10 ng/ml, where the cut-off difference is >5% (25.7% vs. 19.4%, 22.3% vs. 17.1%; Table II).

Table II. Specificity Levels at Given Sensitivities for tPSA, %fPSA and ANN Values in tPSA Ranges 2–4, 4.1–10, 10.1–20, 2–10 and 2–20 ng/ml
tPSA ranges (ng/ml)Sensitivity (%)Specificity (%)1Cut-offs2
  • 1

    Specificity levels and their 95% confidence intervals (in parentheses) at sensitivity cut-offs of 90% and 95% for all investigated tPSA ranges.

  • 2

    Cut-offs of %fPSA and values obtained with the ANN at the given sensitivities of 90% and 95%.

2–49023 (15–34)38 (27–49)63 (51–73)22.30.38
9520 (13–31)20 (13–31)59 (48–70)25.70.31
4.1–109019 (14–25)37 (30–43)57 (51–64)17.10.36
959.2 (5.8–14)27 (22–34)44 (37–50)19.40.32
10.1–209014 (7.6–23)42 (32–53)46 (35–57)15.10.28
956.3 (2.2–14)27 (18–37)36.5 (26–48)17.80.24
2–109021 (17–26)36 (31–42)65 (59–70)17.70.44
9512 (8.9–17)25 (21–31)43 (38–49)21.30.34
2–209025 (21–30)40 (35–45)62 (57–67)17.00.42
9514 (11–18)26 (22–30)43 (38–48)20.70.35
Table III. Sensitivity Levels at Given Specificities for tPSA, %fPSA and ANN Values in tPSA Ranges 2–4, 4.1–10, 10.1–20, 2–10 and 2–20 ng/ml
tPSA ranges (ng/ml)Specificity (%)Sensitivity (%)1Cut-offs2
  • 1

    Sensitivity levels and their 95% confidence intervals (in parentheses) at specificity cut-offs of 90% and 95% for all investigated tPSA ranges.

  • 2

    Cut-offs of %fPSA and values obtained with the ANN at the given specificities of 90% and 95%.

2–4905.7 (1.6–14)34 (23–46)60 (48–72)9.60.84
950 (0–5.5)28 (2.7–17)28 (18–40)5.80.91
4.1–109016 (13–20)30 (26–35)56 (51–60)7.70.98
953.9 (2.3–6.4)15 (12–19)19 (15–23)5.91.04
10.1–209011 (7.2–15)52 (46–58)80 (75–85)7.20.46
951.9 (0.7–4.4)37 (32–43)52 (46–58)6.40.93
2–109017 (13–20)28 (24–33)64 (59–68)7.80.79
953.9 (2.4–6.1)13 (11–17)42 (38–47)5.80.95
2–209022 (19–25)38 (34–41)63 (60–67)7.70.84
956.9 (5.2–8.9)22 (19–25)49 (45–52)5.90.96

ANN and logistic regression

For the 3 tPSA ranges 2–4, 4.1–10 and 10.1–20 ng/ml as well as for the group at 2–10 ng/ml tPSA and all patients at 2–20 ng/ml tPSA, the ANN was evaluated separately. Logistic regression and ANN were compared at all 3 tPSA ranges within the “gray zone,” 2–10 ng/ml. The number of correctly identified patients is given in Table IV. However, by identifying approximately 80% of all patients correctly in the investigated groups, there was no statistical difference in the outcome of both statistical approaches, which confirms recently summarized results.13 Therefore, all further analysis of performance comparisons to tPSA and %fPSA are shown only with ANN.

Table IV. Comparison of Diagnostic Results Obtained with ANN and Logistic Regression in tPSA Ranges 2–4, 4.1–10 and 2–10 ng/ml1
Classification in different tPSA rangesPercent of patients correctly identified
ANNLogistic regression
  • 1

    All values are percentages. Percentages of correctly classified patients (%) using the 0.5 cut-off for ANN and logistic regression and the corresponding sensitivities and specificities at this point are given. In the tPSA ranges 10–20 and 2–20 ng/mL (not separately shown), the ANN identified 78.5% and 80.2% of all patients correctly, respectively. No statistical differences were observed between ANN and logistic regression analysis.

2–4 ng/ml
 Correctly classified patients (%)79.178.6
 Sensitivity (%)8393
 Specificity (%)7664
4.1–10 ng/ml
 Correctly classified patients (%)80.381.6
 Sensitivity (%)8372
 Specificity (%)7687
2–10 ng/ml
 Correctly classified patients (%)81.081.4
 Sensitivity (%)7578
 Specificity (%)8583

ANN results and biopsy recommendations

All ANN calculations and analyses were performed on the 859 patients for whom all data were available. There were no statistical differences between these patients and all 928 patients from the study. Table II summarizes the specificities for tPSA, %fPSA and ANN for the given sensitivities at 90% and 95% for all tPSA ranges. The saved unnecessary biopsies using %fPSA or ANN compared to tPSA can be calculated as differences between specificity levels. Table III equally shows the sensitivities for the given specificities 90% and 95%.

For all 5 tPSA ranges, the ANN model was initially trained on 90% and tested on 10% of the patients. Training and testing sessions were repeated 10 times so that each of the patients was set once in the testing group. This procedure of cross-validation also avoids overtraining of the ANN. The sum of the correctly identified patients in the 10 runs was the overall number of correctly classified patients, which is given in Table IV. As one of the main results of our investigation, use of the ANN demonstrated a significant enhancement over %fPSA in saving unnecessary prostate biopsies, with approximately 20–30% higher specificity levels than for %fPSA. Compared to tPSA alone, the specificity in the different tPSA ranges increased between 28.8% and 43.4%.

Based on the equivalent performance of the 5 established ANNs within the 5 tPSA ranges, we recommend using only 1 ANN for the whole range 2–20 ng/ml. Using this ANN and its cut-offs (Tables II, III), we recommend a first-time biopsy at the low tPSA range 2–4 ng/ml at a specificity level of 90%, to avoid new and unnecessary biopsies. In this low PSA range, ANN outperforms %fPSA and further enhances the sensitivity from 34% (%fPSA) to 60% (Table III).

Regarding the high number of unnecessary biopsies, especially in the tPSA range 4.1–10 ng/ml, we wanted to avoid a general biopsy within this tPSA range and recommend an ANN-based first-time biopsy at the 90% sensitivity level. At this point, the specificity increases from 37% (%fPSA) to 57% (Table II).

For high tPSA values of 10.1–20 ng/ml, we recommend a general biopsy but would not re-biopsy if the ANN value indicates a <5% cancer risk (95% sensitivity). This strategy minimizes the risk of missing cancers and saves repeat biopsies in approximately 30% of patients. Taken together, for all calculated tPSA ranges, the ANN verified a significantly better performance than tPSA and %fPSA alone.

Due to different weighting of the 5 input variables, we calculated the output also for only 3 or 4 input variables. Patient's age had the smallest impact on the outcome of the ANN, reducing the number of correctly classified patients 1–2%. The impact of removing the other factors individually was 5% for tPSA, 8% for prostate volume or %fPSA and 10% for DRE. When the ANN is performed without prostate volume and DRE status, approximately 13% of all correctly classified patients are missed. These data indicate the importance of DRE status and, to a lesser extent, prostate volume as clinical factors for the performance of this ANN. An additional subanalysis with only DRE-negative patients at tPSA 2–10 ng/ml revealed a weaker performance of the ANN with about 72.6% of correctly identified patients. In this case, %fPSA is the strongest factor of the ANN.

The ANN generally gives an output value between 0 (higher BPH possibility) and 1 (higher cancer risk). In some cases, the value can also be below 0 or above 1, which has no further relevance. These absolute ANN values are also compared to the tPSA and %fPSA values in Figure 1. In Figure 1d, showing the whole tPSA range 2–20 ng/ml, the difference between the median %fPSA values for BPH and prostate cancer is approximately 1.7-fold whereas the median ANN values differ 2.5-fold. Based on this ANN, we developed a computer program, ProstataClass, where the physician can easily input the 5 variables and obtain the ANN output. This program is now in clinical use to establish the performance of this ANN.

Figure 1.

Box plots for tPSA, %fPSA and absolute values of ANN for all patients analyzed with the ANN (n = 859). Whiskers indicate 90% and 10% percentiles and boxes represent upper and lower quartiles. Medians are given as numbers and a horizontal line in each box. Values for tPSA ranges: (a) 2–4 ng/ml, (b) 4.1–10 ng/ml, (c) 10.1–20 ng/ml and (d) 2–20 ng/ml.

ROC analysis

ROC analyses for tPSA, %fPSA and ANN were performed for all tPSA ranges separately using the tPSA values, %fPSA values and the ANN calculated values to draw the related ROC curve. When using ROC curves and the AUC to obtain diagnostic performance of a test, we previously recommended the matching procedure to avoid misdistribution of tPSA values and, thus, misinterpretation of significance levels to other molecular forms of PSA.17 We performed this matching procedure for the ROC calculation, but since there were no differences between the curves calculated with matched data and the curves calculated with all patients, we show only the latter curves.

For all tPSA ranges, the curves of %fPSA run significantly above tPSA and the curves of the different ANNs run significantly above %fPSA (Fig. 2a–d). All AUC comparisons reached significance (p < 0.01) with exception of the ANN curve for tPSA 2–4 ng/ml (p = 0.013). However, for the 2.5–4 ng/ml tPSA range, which is commonly used by Catalona et al.6 and others,12 the difference between tPSA and %fPSA (p = 0.025) was comparable to the 2–4 ng/ml tPSA range (ROC curve not shown). For all patients in the 2–10 ng/ml tPSA range (ROC curve not shown), the AUC for tPSA was 0.62, the AUC for %fPSA was 0.73 and the AUC for ANN was 0.86. These ROC data verify the high diagnostic validity of all evaluated ANNs and the advantage of using ANN values instead of only %fPSA cut-offs for biopsy recommendations.

Figure 2.

ROC curves and corresponding AUC ± SE for tPSA, %fPSA and ANN for tPSA ranges (a) 2–4 ng/ml, (b) 4.1–10 ng/ml, (c) 10.1–20 ng/ml and (d) 2–20 ng/ml.


When using PSA alone to predict the probability of prostate cancer in the 4–10 ng/ml tPSA range, approximately 75% of all biopsies will be negative. The discovery of molecular forms of PSA renewed clinical research on enhancing the specificity for this tPSA range more than other calculated parameters, such as PSA density, velocity or age-adjusted cut-offs.1, 19 Within the last few years, it has been clearly demonstrated that use of %fPSA can significantly improve specificity by about 15–25% compared to tPSA, with only a minimal loss in sensitivity of detecting prostate cancer.4, 5, 20 This has been shown for the 4–10 ng/ml tPSA range5 as well as for the lower ranges of 2.6–4 and 2.5–4 ng/ml tPSA.6, 7 In the present study, %fPSA provided a substantial improvement in specificity at 90% sensitivity, between 15% and 28% and in sensitivity at 90% specificity, between 14% and 41%, over the several investigated groups. The enhancement in specificity and sensitivity was demonstrated for all tPSA ranges, including 2–4 ng/ml. This confirms other findings using the Immulite total and free PSA assays.21 Especially in the 2.5–4 or 2.6–4 ng/ml tPSA range, other investigators also found a significant increase in specificity.22, 23 However, some authors using other fPSA assays could not find an improvement of differentiation between prostate cancer and BPH for low tPSA values.24, 25

Our results also confirm our previous prostate biopsy recommendations for the tPSA ranges 2–4 and 4.1–10 ng/ml.26, 27 Using a 9.6% cut-off (formerly 9%) for %fPSA at low tPSA values between 2 and 4 ng/ml, we detected 28% more cancer patients while substantially reducing the number of unnecessary biopsies at this 90% specificity level. The number of performed biopsies per detected cancer was 2.2, but these results are preliminary due to a limited number of patients. For all histologically confirmed patients with a nonsuspicious DRE at tPSA 2–4 ng/ml, the cancer detection rate was 23%. Performing a prostate biopsy only below the %fPSA cut-off of 9.6% leads to a high cancer detection rate of 46%. Interestingly, the figure of 46% of cancer patients in these low tPSA and %fPSA ranges agrees with the previously predicted 46% risk of having prostate cancer with %fPSA values below 10%.6 We suspect that cancers detected at %fPSA below 10% may be more aggressive and clinically significant.28, 29 A high detection rate of 29% of prostate cancers by Catalona et al.7 was obtained when every patient was biopsied in the 2.6–4 ng/ml tPSA range. This may argue for a general biopsy within this tPSA range.

For the 4.1–10 ng/ml tPSA range, we obtained 90% sensitivity with a 17.1% cut-off for %fPSA (Table II, column 6), which should be used for repeat biopsies and which is similar to the 18% cut-off used by Wymenga et al.,21 who also validated the shift to lower %fPSA cut-offs with higher tPSA values. This circumstance has to be considered if a single %fPSA cut-off is in use for a wide tPSA range since cut-off differences were greater than 5% in our investigations for the 2 tPSA ranges 2–4 and 4.1–10 ng/ml. For all biopsy recommendations based only on %fPSA, it should be considered that these are valid only if an ANN is not accessible.

ANNs were introduced into urologic decision making in 1994 by Snow et al.30 Since then, an increasing number of studies have been performed using ANNs or conventional comparable algorithms to enhance the outcome of disease findings, especially for prostate cancer detection.13, 31 In 1998, Carlson et al.9 introduced a logistic regression model that includes %fPSA, tPSA and patient's age; they found an 11% specificity increase over the use of %fPSA alone in the 4–20 ng/ml tPSA range. Virtanen et al.10 estimated another logistic regression model and an ANN using %fPSA, tPSA, DRE status and heredity at tPSA 3–10 ng/ml. The results demonstrated better diagnostic accuracy for prostate cancer detection, with %fPSA and DRE status as the most powerful predictors. This agrees with our results, where DRE status and %fPSA were also the best input variables to detect prostate cancer. A comparison of logistic regression and ANN models by Finne et al.11 for the tPSA range 4–10 ng/ml included prostate volume but not age. At 95% sensitivity, the specificities of %fPSA, logistic regression and ANN were 19%, 24% and 33%. However, our aim was to establish a clinically usable program for the individual calculation of prostate cancer risk. Therefore, comparisons to logistic regression were performed only to show the performance of both methods. Here, we could not substantiate an advantage of ANN or logistic regression for any of the analyzed tPSA ranges (Table IV). The advantage of ANN compared to logistic regression is that the trained ANN can predict the outcome for an individual patient. With different calculated ANNs for the 5 tPSA ranges, we demonstrated a generally significant better performance of ANN compared to %fPSA, enhancing the specificity and sensitivity (Tables II, III) as well as the positive and negative predictive values (data not shown) of approximately 20–30%. For the ANN at 2–20 ng/ml, which is now in clinical use, the increases at the 90% sensitivity and 90% specificity levels were 22% and 25%. However, it must be emphasized that these data are valid only if the Immulite PSA and fPSA are used with this ANN. There are large differences between several PSA assays;16, 32 therefore, the results may change with other assays.

In a screening study, a substantial increase in specificity (150–200%) at the 90% and 95% sensitivity level for patients with tPSA below and above 4 ng/ml using an ANN could be obtained.33 However, the advantage of this ANN with 8 input data including %fPSA, DRE status and different prostate volume parameters was less remarkable at low tPSA values.33

Data on an ANN in the 2.5–4 ng/ml tPSA range from Babaian et al.12 demonstrated a remarkable enhancement of specificity at the 92% sensitivity level from 11% (%fPSA) to 62% (ANN). This ANN was based on a combination of 3 different ANNs using %fPSA, tPSA, age and 2 other serum values as input variables but not DRE or prostate volume. Our data at the low tPSA range 2–4 ng/ml are analogous to these promising results. At 90% specificity, the ANN achieved 60% sensitivity whereas tPSA (5.7%) and %fPSA (34%) had significantly lower results. The advantage of our ANN approach is its applicability for the whole tPSA range. The practical clinical use of this ANN with our recommended biopsy cut-off at 90% specificity for low tPSA values <4 ng/ml to achieve 60% sensitivity, the 90% sensitivity cut-off at tPSA 4.1–10 ng/ml to obtain 57% specificity and the 95% sensitivity cut-off for re-biopsies at tPSA 10.1–20 ng/ml must be proven in further clinical use. A first-time biopsy based on %fPSA at tPSA 4–10 ng/ml has been proposed.34 Our trained ANN will further avoid unnecessary first-time prostate biopsies over the use of %fPSA alone.

Our data regarding prostate volume as an influencing factor of %fPSA confirm data on more than 3,300 patients, where %fPSA increased with increasing volume in prostate cancer and BPH patients.35, 36 In contrast to our previous data on smaller numbers of patients,15, 37 we found in the present survey that %fPSA can also distinguish between cancer and BPH patients if prostate volumes exceeds 40 cm3 but the median %fPSA values are significantly higher for both groups.

For staging of prostate cancer, use of %fPSA is controversial. Our results are comparable to those of previous studies showing a lower %fPSA for pT3 cancer compared to pT2 cancer.38, 39 Others could not validate this relationship.40, 41, 42 Despite a lower median %fPSA value for grade 3 tumors compared to grade 2 tumors (7.3% vs. 8.7%), this difference was not significant (p = 0.05), confirming other reports.21 In further studies, %fPSA was an independent predictor of tumor grade and aggressiveness.28, 43 Taken together, these results on staging and grading suggest that %fPSA cannot predict the individual outcome of a patient but may help to identify high-risk patients.

Other molecular forms of PSA or fPSA and human glandular kallikrein 2 show promising preliminary results in improving the differentiation between prostate cancer and BPH.44, 45, 46, 47, 48 Therefore, it is likely that in the near future these new molecular forms of PSA and other kallikreins (especially human glandular kallikrein 2) may add substantial information in extended ANNs, including these markers and %fPSA.

We developed a diagnostic algorithm based on Immulite tPSA and fPSA and clinical data to enhance the %fPSA performance to further reduce the number of unnecessary biopsies in the tPSA range 2–10 and to reduce the number of rebiopsies at tPSA 10.1–20 ng/ml. This ANN is based on preliminary data of a specifically referred population and remains to be proven prospectively in screening studies. In the future, ANNs should be used and hopefully expanded with new promising input data to decrease the number of false-positive results in detecting prostate cancer.


This work was supported by the following sources and grants: MSD (to CS), Deutsche Forschungsgemeinschaft (to KJ), Funds of the German Chemical Industry (to KJ) and the Sonnenfeld-Stiftung (to SAL). We gratefully acknowledge Dr. P. Sibley for helpful suggestions. The study contains part of the doctoral thesis of BV.