Recursive partitioning for risk stratification in men undergoing repeat prostate biopsies

Authors


  • This article is a US Government work and, as such, is in the public domain in the United States of America.

Abstract

BACKGROUND

The current study was performed to identify risk factors and risk groups for carcinoma detection in men undergoing repeat prostate biopsies.

METHODS

The medical records of all men who had a negative initial prostate biopsy and underwent at least one repeat biopsy between 1992 and 2003 were reviewed to extract age, race, family history of prostate carcinoma, body mass index, referral indication, all prostate-specific antigen (PSA) values, digital rectal examination, PSA density (PSAD), the presence of a hypoechoic lesion, and the presence of high-grade prostatic intraepithelial neoplasia (HGPIN) on initial biopsy. Risk factors for a subsequent diagnosis of prostate carcinoma were identified using the log-rank test and a stepwise, stratified Cox regression model. Based on the risk factors identified by Cox regression analysis, recursive partitioning was further used for risk stratification.

RESULTS

A total of 373 patients underwent 975 biopsy procedures. During a median follow-up of 37.0 months, prostate carcinoma was detected in 107 of 373 patients (28.9%). Independent predictors of a positive biopsy (P < 0.05) were PSA doubling time (PSADT), PSAD, referral indication, the presence of HGPIN, patient age, and family history of prostate carcinoma. Recursive partitioning identified 4 distinct risk groups that were characterized by their PSADT and PSAD and the presence of HGPIN and had estimated 2-year and 5-year carcinoma detection rates of 3 ± 1% and 21 ± 4%, 28 ± 5% and 40 ± 7%, 22 ± 6% and 58 ± 8%, and 66 ± 9% and 100%, respectively.

CONCLUSIONS

The authors identified a series of independent risk factors for prostate carcinoma detection after an initial negative prostate biopsy and characterized clinically meaningful and distinct patient risk groups. Despite a negative initial biopsy, patients with high-risk features remain at risk for the detection of prostate carcinoma. Cancer 2005. Published 2005 by the American Cancer Society.

Prostate carcinoma screening has become commonplace in the U.S., in which approximately 70% of men have undergone prostate-specific antigen (PSA) screening.1 There remains a strong desire for patients to be aware of their cancer status even if no therapy would be advised.1 Prostate carcinoma is detected on core needle biopsy in approximately 25% of men with abnormalities detected on initial screening.2 However, a negative biopsy does not ensure the absence of disease because repeat biopsies are found to be positive in approximately 25% of patients.3–7

Several individual risk factors have been identified as predictors of prostate carcinoma in men undergoing repeat biopsies of the prostate. These include suspicious digital rectal examination (DRE), total PSA, PSA velocity, percentage of free PSA, hypoechoic lesions on transrectal ultrasound (TRUS), PSA density (PSAD) and transition zone PSAD, the presence of high-grade prostatic intraepithelial neoplasia (HGPIN), and atypical small acinar proliferation.6, 8–10 Although each of these factors has predictive value for the detection of prostate carcinoma by biopsy, individually they lack sufficient predictive power for clinical use. Prior studies also may have been limited by small sample size, the inclusion of patients whose initial biopsies were performed at other institutions, and limited follow-up. Multifactorial models have been developed for carcinoma prediction on an initial biopsy using logistic regression and neural network methods.11–13 For patients undergoing repeat biopsy, a time-dependent nomogram has been developed by Lopez-Corona et al. based on Cox regression analysis.9 In that study, four factors were found to independently predict the results of biopsy, including number of negative core needle biopsy specimens, PSA slope, the presence of HGPIN, and atypical small acinar proliferation.

The goals of the current study were to determine risk factors for the detection of prostate carcinoma and to identify risk groups using the recursive partitioning technique. This statistical method belongs to a family of nonparametric regression methods and can be implemented using software for censored data,14 which builds a decision tree and is capable of classifying patients into risk groups.

MATERIALS AND METHODS

Study Population and Biopsy Procedures

The records of all men who underwent a TRUS with prostate biopsy comprised of a minimum of six biopsy core needle biopsy specimens at the Portland Veterans Administration Medical Center between 1992 and 2003 were reviewed. All patients were referred for the suspicion of prostate carcinoma as part of routine care, not as a component of a formal screening program. Data from all the biopsy procedures were uniformly collected for the purpose of clinical care. Variables recorded at the time of first biopsy included age, family history of prostate carcinoma in a first-degree relative, race, indication for patient referral, history of vasectomy, DRE findings, and serum PSA. An urinalysis was performed before each procedure to rule out the presence of an urinary tract infection. Included in this analysis were all patients who had undergone at least two biopsy procedures during which at least six core needle biopsy specimens were obtained at each setting.

Prostate ultrasounds and biopsies were performed using a Bruel and Kjær® 3535 device (Norcross, GA) as previously described.11 The PSAD was calculated by dividing the serum PSA by the calculated prostate volume. Additional lesion biopsies and transition zone biopsies were obtained when indicated. PSA values, medications, body mass index, biopsy results, and follow-up data (including the histologic diagnosis of prostate carcinoma by other means [i.e., transurethral prostate resection or biopsy of metastatic sites]) were abstracted from the medical records. Patients were censored at the time of last PSA follow-up or at the initiation of therapy with finasteride or other hormonal agents. This study was approved by the Portland Veterans Administration Institutional Review Board and the Research and Development Committee.

Statistical Analysis

The primary endpoint was the time from the initial negative biopsy to the histologic detection of prostate carcinoma by any means. All serum PSA measurements from the time of the initial biopsy until last follow-up or therapeutic intervention were recorded. PSA doubling time (PSADT) was calculated using all available PSA measurements before the initial biopsy as well as during the follow-up as Ln(2)/b, in which b is a slope estimate of Ln(PSA) versus year in a simple linear regression analysis, in which Ln represents natural log. We used either commonly used cutoffs or categories based on the sample distribution of a risk variable for categoric analyses.

The Kaplan–Meier method was used to estimate the probability of a diagnosis of prostate carcinoma since the first prostate biopsy procedure. Univariate analysis was performed using the log-rank test. Multivariate analysis was performed using the stepwise, stratified Cox regression model, which allowed a different baseline detection rate according to the number of repeat biopsies each patient underwent. Predictors found to be significant on the multivariate analysis were evaluated using the recursive partitioning method14 for censored data. To characterize the role of PSADT further, we analyzed PSA over time using a random effects model (i.e., random intercept and slope). We estimated a linear regression of Ln(PSA) as a function of time for those with and without a subsequent carcinoma diagnosis. Statistical significance was defined as a P value of < 0.05.

RESULTS

Patient Characteristics

Based a review of the medical records, a total of 1563 consecutive patients who underwent TRUS with biopsy were identified. Of these patients, 373 (23.9%) had a negative initial biopsy and underwent at least 1 additional biopsy procedure at the study institution. This group underwent a total of 975 biopsy procedures, which yielded 6878 prostate tissue core needle biopsy specimens for histologic evaluation. The median age of the patients was 66 years (mean ± standard deviation [SD], 64.9 ± 6.8 yrs) (Table 1). The median PSA was 6.4 ng/mL (mean ± SD, 7.9 ± 6.14 ng/mL). Finasteride therapy was initiated in 22 patients (5.9%) during the follow-up period and the use of other androgen-suppressive agents was not observed.

Table 1. Patient Characteristics of Men Undergoing a Repeat Prostate Biopsy (n = 373)
VariableResult
  1. PSA: prostate-specific antigen; DRE: digital rectal examination; BMI: body mass index; PSAD: prostate-specific antigen density; HGPIN: high-grade prostatic intraepithelial neoplasia; PSADT: prostate-specific antigen doubling time.

Median follow-up in yrs, (range)3.1 (0.1–−12.6)
Median age in yrs, (range)66 (46–83)
Race, no. 
White340 (91.1%)
Black20 (5.4%)
Other13 (3.5%)
Referral reason, no. 
PSA255 (68.4%)
DRE62 (16.6%)
PSA and DRE49 (13.1%)
Other reason or reason not available7 (1.9%)
Family history, no.67 (18.0%)
Vasectomy, no.119 (31.9%)
Median BMI in (kg/m2), (range)27.5 (11.0–45.7)
Median PSA in ng/mL (range)6.4 (0.2–50.4)
DRE, no. 
Normal214 (57.4%)
Asymmetric16 (4.3%)
Suspicious130 (34.9%)
Carcinoma likely6 (1.6%)
Median prostate volume in (range)45.5 (14.1–235)
Median PSAD in ng/mL/cc, (range)0.15 (0.01–1.84)
No. of biopsy procedures, median, mean (range)2.0, 2.6 (2–7)
Median no. of core needle specimens obtained per patient, (range)17 (12–53)
HGPIN, no.57 (15.3%)
Tumor detected, no.107 (28.7%)
PSADT in yrs 
<219 (5.1%)
2–580 (21.4%)
>5274 (73.5%)

Ultrasound and Biopsy Data

The median PSAD was 0.15 ng/mL/cc (mean ± SD, 0.20 ± 0.19 ng/mL/cc) at the time of the initial prostate ultrasound procedure. The median total number of core needle biopsy specimens per patient was 17 (mean ± SD, 18.4 ± 7.0 specimens). All patients underwent at least 2 biopsy procedures (mean ± SD, 2.6 ± 0.9 procedures; range, 2–7 procedures). The median time between the first and last prostate biopsy was 19.2 months (mean ± SD, 27.6 ± 25.7 mos). The median follow-up period from the time of initial biopsy was 37.0 months for all patients and 42.9 months for patients who were found to be disease-free at last follow-up.

Histologic review of specimens demonstrated HGPIN in 15.3% of the initial biopsy specimens. Adenocarcinoma of the prostate was detected on core needle biopsy in 100 of 373 patients (26.8%). A transurethral prostatectomy or transurethral prostate biopsy was performed in 32 of 373 patients (8.6%), and carcinoma was detected in 6 of these patients (18.8%). Prostate carcinoma was diagnosed by core needle biopsy of a supraclavicular lymph node in one patient. Therefore, prostate carcinoma was detected in 107 of 373 patients (28.7%).

The incidence of prostate carcinoma for those patients undergoing 2, 3, 4, or 5–7 biopsy procedures was found to be 16%, 16%, 19%, and 31%, respectively (Fig. 1). Among those patients with a positive biopsy, a Gleason score of ≥ 7 was noted in 41% of those who underwent 2 biopsy procedures, 36% of those who underwent 3 biopsy procedures, 30% of those who underwent 4 biopsy procedures, and 60% of those who underwent 5–7 biopsy procedures.

Figure 1.

Fraction of patients undergoing each biopsy procedure in whom prostate carcinoma was detected by that procedure. The proportion of patients with a tumor of Gleason score of 7 or higher are shown above the proportion of patients with tumors of Gleason score 6 or lower.

Predictors of Prostate Carcinoma Diagnosis

The following covariates were found to be associated with the time to prostate carcinoma diagnosis on the univariate analysis based on a log-rank test: PSAD (P < .0001), abnormal DRE and PSA referral indication (P = .0011), family history (P = 0.047), HGPIN detected on initial biopsy (P = 0.0021), initial PSA (P = 0.0161), and PSADT (P < 0.0001).

Stepwise, stratified Cox regression analysis was performed to identify a set of covariates that were independently predictive of a diagnosis of prostate carcinoma. A PSAD > 0.25 ng/mL/cc (P = 0.0014), abnormal DRE and PSA referral indication (P = 0.0002), HGPIN detected on initial biopsy (P = 0.0011), and a family history (P = 0.0076) were found to be independently associated with a diagnosis of prostate carcinoma, whereas a PSADT > 5 years (P < 0.0001), PSAD < 0.09 ng/mL/cc (P = 0.0039), PSAD of 0.09–0.13 ng/mL/cc (P = 0.0027), and patient age ≥ 70 years (P < 0.0001) were found to be inversely associated with a diagnosis of prostate carcinoma (Table 2).

Table 2. Results of Multivariate Analysis (Stratified Cox Regression Model) of Risk Factors for Prostate Carcinoma Detection in Patients Undergoing a Repeat Biopsy
VariableHR95% CIP value
  1. HR: hazards ratio; 95% CI: 95% confidence interval; PSADT: prostate-specific antigen doubling time; PSAD: prostate-specific antigen density; PSA: prostate-specific antigen; DRE: digital rectal examination; HGPIN: high-grade prostatic intraepithelial neoplasia.

PSADT > 5 yrs0.2410.156–0.372<0.0001
PSAD > 0.25 ng/ml/cc2.0981.330–3.3100.0014
PSAD 0.09–0.13 ng/mL/cc0.3240.156–0.6760.0027
PSAD < 0.09 ng/mL/cc0.3350.160–0.7040.0039
Referral indication (elevated PSA + abnormal DRE)2.6671.600–4.4460.0002
HGPIN2.2451.379–3.6570.0011
Age ≥ 70 yrs0.5260.335–0.825<0.0001
Positive family history1.9221.186–3.0450.0076

Risk groups were then identified using recursive partitioning analysis of factors that were found to be significant by Cox regression analysis. After initial pruning, recursive partitioning identified five risk groups, which then were collapsed back into four groups because of similarities in outcomes between two groups. The 4 risk groups had estimated 2-year and 5-year cancer detection rates of 3 ± 1% and 21 ± 4%, respectively (Group 1, in which the PSAD was < 0.25 ng/mL/cc, the PSADT was ≥ 5 years, and no HGPIN was present [n = 187]); 28 ± 5% and 40 ± 7%, respectively (Group 2, in which the PSADT was ≥ 5 years, the PSAD was < 0.25 ng/mL/cc, and HGPIN was present; or the PSADT was ≥ 5 years and the PSAD was ≥ 0.25 ng/mL/cc [n = 87]); 22 ± 6% and 58 ± 8%, respectively (Group 3, in which the PSADT was < 5 years and the PSAD was < 0.25 ng/mL/cc [n = 71]); and 66 ± 9% and 100%, respectively (Group 4, in which the PSADT was < 5 years and the PSAD was ≥ 0.25 ng/mL/cc [n = 28]) (Fig. 2). The mean number of biopsy procedures was similar in all 4 groups (2.63 ± 0.92 procedures vs. 2.57 ± 0.69 procedures vs. 2.66 ± 0.97 procedures vs. 2.61 ± 1.03 procedures in Groups 1–4, respectively).

Figure 2.

Kaplan–Meier estimates of disease-free survival for the four risk groups characterized in the current analysis. Group 1 patients had a prostate-specific antigen doubling time (PSADT) of ≥5 years, a prostate-specific antigen density (PSAD) of < 0.25 ng/mL/cc, and no high-grade prostatic intraepithelial neoplasia (HGPIN). Group 2 patients had a PSADT of ≥ 5 years, a PSAD of < 0.25 ng/mL/cc, and HGPIN, or had a PSADT of ≥ 5 years and a PSAD of ≥ 0.25 ng/mL/cc. Group 3 patients had a PSADT of < 5 years and a PSAD of < 0.25 ng/mL/cc. Group 4 patients had a PSADT of < 5 years and a PSAD of ≥0.25 ng/mL/cc.

Analysis of PSA over Time

The analysis of PSA profiles revealed that Ln(PSA) increased linearly as a function of time for both those patients with and those without a subsequent carcinoma diagnosis. However, the PSA trajectory was much greater for those patients with carcinoma than for those patients without carcinoma (Table 3). The estimated PSADT was 4.5 years (95% confidence interval [95% CI], 3.8–5.5 years) for those with carcinoma compared with 18.6 years (95% CI, 13.2–31.3 years) for those patients without carcinoma. Figure 3 shows the Ln(PSA) profiles of a randomly selected subset of patients with carcinoma (n = 30) and noncarcinoma patients (n = 30) with the predicted regression line of each group. It is important to note that the increase was gradual over time and that PSA did not appear to rise sharply just before the diagnosis of prostate carcinoma was made. This finding supports the utility of PSADT as a potential predictor for carcinoma detection in this population and underscores the importance of frequent and regular PSA screening and monitoring so that the PSA trajectory can be measured accurately.

Table 3. Results of Random Effects Model for Ln(PSA) Change Over Time
VariableBetaSEP value
  1. Ln(PSA): natural log prostate-specific antigen; SE: standard error; PSADT: prostate-specific antigen doubling time; 95% CI: 95% confidence interval.

Intercept1.86610.06699<0.0001
No carcinoma−0.15700.079290.0478
Time (mos)0.012780.00117<0.0001
Time* no carcinoma−0.009680.00133<0.0001
GroupPredicted regression lineEstimated PSADT (95% CI)
No carcinomaLn(PSA) = 1.7091 + 0.0031 time18.63 yrs (13.26–31.33)
CarcinomaLn(PSA) = 1.8661 + 0.01278 time4.52 yrs (3.83–5.51)
Figure 3.

(A) Prostate-specific antigen (PSA) profiles over time among 30 randomly selected noncancer patients. The natural log PSA (Ln(PSA)) was plotted against the time from the PSA at the time of the initial negative biopsy (in months). Note that PSA profiles from a subset of patients are shown to enhance the clarity of the figure. The predicted regression line is based on the random effects model using all 373 patients. (B) PSA profiles over time among 30 randomly selected prostate carcinoma patients and the predicted regression line from the random effects model.

DISCUSSION

Several prostate biopsy strategies have been applied in an attempt to improve the initial detection of cancer, and therefore reduce the rate of false-negative biopsies. However, to our knowledge, no biopsy strategy to date has sufficient negative predictive value to eliminate the need for vigilant follow-up, which may include additional biopsy procedures. Differences in risk factors examined and other study limitations may have precluded a consensus based on prior reports.8 Therefore, the selection of patients who should undergo additional biopsies of the prostate after an initial negative examination remains a common and vexing clinical problem.

To address this challenge, we examined a broad range of potential risk factors to identify covariates that are associated with a subsequent diagnosis of prostate carcinoma. We then used recursive partitioning analysis to identify clinically meaningful risk groups.

We found that a short PSADT, the presence of HGPIN, patient age, referral indication of abnormal DRE and PSA, PSAD, and a family history are independently associated with a diagnosis of prostate carcinoma in men whose initial prostate biopsy is negative. Recursive partitioning analysis then defined 4 groups with distinct 2-year and 5-year cancer detection rates. The members of these groups were characterized by their PSAD, PSADT, and the presence or absence of HGPIN. Patients in the highest risk group (those with a PSADT of ≤ 5 years and a PSAD of >0.25 ng/mL/cc) had estimated carcinoma detection rates of 66 ± 9% and 100%, respectively, at 2 years and 5 years. Therefore, these patients should be strongly considered for a repeat prostate biopsy within a few months. Conversely, patients in the lowest-risk group, which included greater than half of the patients studied (those with a PSAD of < 0.25 ng/mL/cc, a PSADT of > 5 years, and no HGPIN) could be considered for less rigorous follow-up because the estimated carcinoma detection rate was found to be only 3 ± 1% at 2 years.

The identification of HGPIN as the factor that distinguishes between low risk and moderate risk is worthy of further discussion. HGPIN is the purported precursor lesion to invasive prostate carcinoma.5 The finding of HGPIN has been shown to correlate with the detection of prostate carcinoma on a repeat prostate biopsy in some studies.9, 15, 16 However, the results of what to our knowledge is one of the largest studies performed to date demonstrated that HGPIN alone did not increase the risk of prostate carcinoma detection, unless suspicious glands also were present in the sampled material.6 Others have similarly failed to detect a correlation between the presence of HGPIN and prostate carcinoma, perhaps because of a lack of power or insufficient follow-up.4, 17, 18 In the current study, patients with HGPIN had a significant increase in the risk of prostate carcinoma (relative risk [RR] = 2.5; 95% CI, 1.5–4.2) compared with HGPIN-negative patients. Furthermore, analysis with recursive partitioning demonstrated that HGPIN may be most useful in deciding which patients need a repeat biopsy in low-risk to moderate-risk groups. Patients with a PSADT of > 5 years and a PSAD of < 0.25 ng/mL/cc with HGPIN detected in the initial biopsy specimen had a 28% ± 5% risk of carcinoma detection at 2 years compared with only 3% ± 1% when HGPIN was absent. The presence of HGPIN lost significance in our analysis when either the PSADT was short or the PSAD was increased. The detection of this type of unique interaction is a primary strength of recursive partitioning. These data support the concept that HGPIN is a marker for occult prostate carcinoma in select groups.

In addition to uncertainty regarding who should undergo a repeat biopsy of the prostate, the number and frequency of biopsy procedures needed to rule out the presence of prostate carcinoma remain uncertain.7, 10 In a prospective analysis of planned serial biopsies, Djavan et al. demonstrated that the rate of detection of prostate carcinoma decreased with the number of biopsy procedures performed, with a reported detection rate of 10% when 2 biopsy procedures were performed, 5% when 3 biopsy procedures were performed, and 4% when 4 biopsy procedures were performed.10 However, the interbiopsy interval for these patients was short and follow-up data were not available. In contrast, the carcinoma detection rate in the current study was reported to be 16% after 2 biopsy procedures were performed, 16% after 3 biopsy procedures were performed, 19% after 4 biopsy procedures were performed, and 31% after 5–7 biopsy procedures were performed. Furthermore, the frequency with which tumors with a Gleason scores of ≥ 7 were detected also failed to decrease with the increasing numbers of biopsy procedures performed. This raises the concern that a subset of patients remains at risk for harboring clinically significant prostate carcinoma despite having undergone multiple, negative biopsy procedures.

The strengths of the current study include its size, length of follow-up, the quality control possible in a single institution setting, and the finding that the patient population was unselected and included all patients referred from the primary care clinics of the region's Veterans Administration health care system. Furthermore, all diagnoses of prostate carcinoma were included in the final analysis, which permitted the capture of an additional 7 of 107 cases (6.5%). Unlike other regression-based methods, the recursive partitioning method can be used to identify high-risk subgroups and uncover interactions among prognostic factors without assuming a specific parametric probability model. Recent studies have shown this method to have performance measures that are similar to other linear regression techniques and artificial neural networks.19

The absence of a percentage of free PSA in the analysis is a limitation of the current study. The percentage of free PSA has been shown to predict the occult presence of prostate carcinoma in those patients who undergo repeat biopsy.4, 6, 20 The percentage of free PSA data were not available for the entire cohort and therefore we did not analyze this covariate. However, we included PSAD, which previously was shown to be slightly superior to the percentage of free PSA in patients undergoing repeat biopsies (area under the curve of 0.75 vs. 0.72).20 Sampling bias is a possible cause for affecting incident prostate carcinoma cases in all biopsy series. To address this, we performed Cox regression analysis stratified by the number of repeat biopsies; however, the results from the stratified analysis were very similar to those obtained from the unstratified analysis including the number of core needle specimens as an adjustment variable, and there was no major change noted with regard to parameters selected by recursive partitioning. We used all available PSA data to estimate the PSADT because there were not sufficient PSA data before the initial biopsy with which to estimate the PSADT accurately for each patient. Given that several consecutive PSA measurements may be required to achieve an accurate value, we believed that this would provide the best representation of each patient's true, underlying PSADT. Although this is a limitation of the current study, the analysis of PSA over time suggests that an increase in PSA is gradual over time and that PSA does rise sharply shortly before a diagnosis of prostate carcinoma is made. Therefore, the PSADT calculated based on all data does appear to be a surrogate of the underlying true PSADT. A lack of accurately measured PSADT at the time of the initial biopsy will limit the utility of the proposed risk stratification in clinical settings. Therefore, it is critical to monitor PSA regularly and frequently among high-risk patient groups. The retrospective nature of the current analysis represents another important limitation because the decision to proceed to additional biopsies was a result of the interaction between the patient and the clinician and likely was not random. At the same time, the applicability of these results to clinical practice is enhanced by the clinical practice origins of the data used for this analysis.

Conclusions

Risk factors for prostate carcinoma detection in men undergoing a repeat prostate biopsy were a short PSADT, the presence of HGPIN, patient age, referral indication of an abnormal DRE and PSA, PSAD, and a family history of prostate carcinoma. Recursive partitioning analysis was able to characterize high-risk, intermediate-risk, and low-risk groups for the subsequent detection of prostate carcinoma. This information can serve as a valuable tool to the clinician and patient who are contemplating the risk of occult prostate carcinoma after an initial negative prostate biopsy.

Ancillary