Prediction of patient-specific risk and percentile cohort risk of pathological stage outcome using continuous prostate-specific antigen measurement, clinical stage and biopsy Gleason score

Authors

  • Ying Huang,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Sumit Isharwal,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Alexander Haese,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Felix K.H. Chun,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Danil V. Makarov,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Ziding Feng,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Misop Han,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Elizabeth Humphreys,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Jonathan I. Epstein,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Alan W. Partin,

    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author
  • Robert W. Veltri

    Corresponding author
    1. Department of Biostatistics, Columbia University Mailman School of Public Health, New York, NY , *Department of Urologic Surgery, University of Minnesota, Minneapolis, MN, USA, Martini-Clinic Prostate Cancer Center, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Department of Urology, Yale University School of Medicine, New Haven, CT, §Biostatistics Program, Fred Hutchinson Cancer Research Center, Seattle, WA, USA, and ††The James Buchanan Brady Urological Institute, The John Hopkins University School of Medicine, Baltimore, MD, USA
    Search for more papers by this author

Dr Robert W. Veltri, Associate Professor, James Buchanan Brady Urological Institute, The Johns Hopkins University School of Medicine, Baltimore, Maryland 21287, USA. e-mail: rveltri1@jhmi.edu

Abstract

Study Type – Therapy (case series)
Level of Evidence 4

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

This international collaboration started in 2008 based upon the possible application of the ‘predictiveness curves’ (multinomial logistic regression method) developed at the Fred Hutchinson Cancer Research Center (FHCRC) to the internationally recognized Partin Tables for staging prostate cancer. Dr. Ying Huang, a biostatistician at the FHCRC, applied the ‘predictiveness curve’ statistical modeling concept to the Partin Tables and then created a new Partin Nomogram using total PSA (tPSA) as a continuous variable.

The new ‘2010 Partin Nomogram’ stage risk prediction capacity among the total cohort and the individual patients is based on the ‘predictiveness curves’ using the method developed in Huang et al.[16]. For each pathological stage, we calculated ‘the risk’ for each subject in the cohort based on the risk model and made a quantile plot based on the estimated risks. If one considers a point on the ‘predictiveness curve’ with an x-coordinate of value v, then the value of its y-coordinate, which we name R(v), is the 100 × vth percentile of risk in the study cohort. On the other hand, for a particular point on the curve with y-coordinate p, the value of its x-coordinate, which we name R−1(p), corresponds to the percentage of subjects in the study cohort with risk ≤p[i.e. the cumulative distribution function (CDF) of risk at p]. It is likely that this CDF of risk will be useful for clinicians and patients (see Fig. 2 in the article). Dr. Huang has also written an operational R-program to calculate patient's risk and next we intend to develop a user friendly computer program based upon this program to allow the easy use by patients and physicians of the 2010 Partin Nomogram and the corresponding predictiveness curves for patient-specific pathological stage outcome prediction.

OBJECTIVES

• To develop a ‘2010 Partin Nomogram’ with total prostate-specific antigen (tPSA) as a continuous biomarker, in light of the fact that the current 2007 Partin Tables restrict the application of tPSA as a non-continuous biomarker by creating ‘groups’ for risk stratification with tPSA levels (ng/mL) of 0–2.5, 2.6–4.0, 4.1–6.0, 6.1–10.0 and >10.0.

• To use a ‘predictiveness curve’ to calculate the percentile risk of a patient among the cohort.

PATIENTS AND METHODS

• In all, 5730 and 1646 patients were treated with radical prostatectomy (without neoadjuvant therapy) between 2000 and 2005 at the Johns Hopkins Hospital (JHH) and University Clinic Hamburg-Eppendorf (UCHE), respectively.

• Multinomial logistic regression analysis was performed to create a model for predicting the risk of the four non-ordered pathological stages, i.e. organ-confined disease (OC), extraprostatic extension (EPE), and seminal vesicle (SV+) and lymph node (LN+) involvement.

• Patient-specific risk was modelled as a function of the B-spline basis of tPSA (with knots at the first, second and third quartiles), clinical stage (T1c, T2a, and T2b/T2c) and biopsy Gleason score (5–6, 3 + 4 = 7, 4 + 3 = 7, 8–10).

RESULTS

• The ‘2010 Partin Nomogram’ calculates patient-specific absolute risk for all four pathological outcomes (OC, EPE, SV+, LN+) given a patient’s preoperative clinical stage, tPSA and biopsy Gleason score.

• While having similar performance in terms of calibration and discriminatory power, this new model provides a more accurate prediction of patients’ pathological stage than the 2007 Partin Tables model.

• The use of ‘predictiveness curves’ has also made it possible to obtain the percentile risk of a patient among the cohort and to gauge the impact of risk thresholds for making decisions regarding radical prostatectomy.

CONCLUSION

• The ‘2010 Partin Nomogram’ using tPSA as a continuous biomarker together with the corresponding ‘predictiveness curve’ will help clinicians and patients to make improved treatment decisions.

Abbreviations
CDF

cumulative distribution function

EPE

extraprostatic extension

JHH

Johns Hopkins Hospital

LN+

lymph node involvement

OC

organ-confined

PCa

Prostate cancer

RP

radical prostatectomy

SV+

seminal vesicle involvement

tPSA

total PSA

UCHE

University Clinic Hamburg-Eppendorf.

INTRODUCTION

Prostate cancer (PCa) is the second most commonly diagnosed cancer and the sixth most common cause of death from cancer among men in the world, with an approximately sixfold difference between high- and low-incidence countries [1]. Most patients with clinically localized PCa are treated with radical prostatectomy (RP) or radiotherapy, which provides excellent cancer control [2]. However, there is no consensus regarding the optimal management of locally advanced PCa [3].

The ability to predict accurately the pathological stage of PCa before surgery allows for improved patient counselling, a more appropriate selection of treatment plan and risk stratification for novel clinical trials for those with more advanced disease. Our group and others have published algorithms and nomograms predicting the pathological stage of patients with localized PCa [4–13]. The Partin Tables were updated in 2007 to reflect stage migration and they continue to provide a clinically useful adjunct to predict the pathological stage of patients with PCa [7]. The 2007 Partin Tables have recently been validated successfully [14,15]. The 2007 Partin Tables [7] used clinical stage, biopsy Gleason score and total PSA (tPSA) information to estimate the pathological extent of disease. However, the usefulness of tPSA was limited by creating ‘groups’ for the risk stratification of patients with tPSA levels (ng/mL) of 0–2.5, 2.6–4.0, 4.1–6.0, 6.1–10.0 and >10.0.

Hence, we developed ‘2010 Partin Nomogram’ for predicting patients’ pathological stages based on tPSA measured at a continuous scale, clinical stage and biopsy Gleason score. In addition, we plan to provide clinicians and patients with the risk quantile plots, called the ‘predictiveness curve’[16], for each individual pathological stage outcome as an additional tool for patient-specific decision-making.

MATERIALS AND METHODS

THE JOHNS HOPKINS HOSPITAL (JHH) PATIENT COHORT

The institutional review board at JHH approved the present study and, when required, written informed consent was obtained from study participants. From 2000 to 2005, a total of 5988 men with PCa who underwent RP and staging pelvic lymphadenectomy at the JHH, by any of 22 attending surgeons, were identified. All enrolled patients had preoperative serum tPSA level assessed on an ambulatory basis before RP either before or at least 4 weeks after prostate biopsy, a biopsy Gleason score determined at JHH, and a clinical stage, assigned by the attending physician [American Joint Committee on Cancer Tumour-Node-Metastasis (AJCC-TNM) staging system, 1992/2002], of T1c or T2a/b/c. Patients were excluded from the cohort for the following reasons: they lacked this information (n= 29); had received preoperative neoadjuvant hormonal therapy (n= 107); had pathological diagnoses other than adenocarcinoma of the prostate (n= 7); had an absence of cancer on pathology (n= 4) or missing pathological information (n= 33); had preoperative treatment with 5-alpha reductase inhibitors (n= 71); had undergone chemotherapy (n= 5) or had taken androgenic/oestrogenic herbal therapies (n =1). Additionally, 30 patients with tumour extending to an inked surgical margin of the prostate who could not be interpreted as organ-confined (OC) or extraprostatic extension (EPE) were also excluded. One patients with biopsy Gleason score 2 + 2 = 4 was also excluded [17], leaving a cohort of 5730 patients (Table 1).

Table 1.  Demographic, clinical and pathological information of patients in the JHH and UCHE cohorts
VariableJHHUCHEP value*
(N= 5730)(N= 1646)
Mean (±SD) or N (%)Mean (±SD) or N (%)
  • *

    P value comparing JHH and UCHE based on two-sample t-test for continuous covariates and based on chi-squared test for discrete covariates.

Age, years  57.4 (6.4)  61.9 (5.7)<0.001
PSA level, ng/mL  <0.001
 ≤2.5 452 (7.9)  67 (4.1) 
 2.6–4.0 946 (16.5)  211 (12.8) 
 4.1–6.01994 (34.8) 434 (26.4) 
 6.1–8.01093 (19.1) 354 (21.5) 
 8.1–10 578 (10.1) 199 (12.1) 
 >10 667 (11.6) 381 (23.1) 
Clinical stage  <0.001
 T1c4419 (77.1)1207 (73.3) 
 T2a 998 (17.4) 297 (18.0) 
 T2b/c 313 (5.5) 142 (8.7) 
Biopsy Gleason score  <0.001
 <74402 (76.8)1059 (64.3) 
 3 + 4 = 7 816 (14.2) 403 (24.5) 
 4 + 3 = 7 348 (6.1) 127 (7.7) 
 >7 164 (2.9)  57 (3.5) 
Organ confined  0.531
 Yes4204 (73.4)1221 (74.2) 
 No1526 (26.6) 425 (25.8) 
Extraprostatic extension  < 0.001
 No4454 (77.7)1370 (83.2) 
 Yes1276 (22.3) 276 (16.8) 
Seminal vesicle involvement  <0.001
 No5550 (96.9)1535 (93.3) 
 Yes 180 (3.1) 111 (6.7) 
Lymph node involvement  0.002
 No5660 (98.8)1608 (97.7) 
 Yes  70 (1.2) 38 (2.3) 

PATHOLOGICAL EXAMINATION

All pelvic lymph nodes removed at surgery were sectioned and examined for the presence of cancer. The surgical specimen, prostate and seminal vesicles were analysed and the pathological stage was defined as OC if all cancer was confined within the prostate; EPE if cancer was evident outside the prostate and the seminal vesicles and pelvic lymph nodes were free of disease; positive seminal vesicle involvement (SV+) if the tumour invaded the muscular wall of the seminal vesicle without lymph node involvement; and lymph node involvement (LN+) if the pelvic lymph nodes showed PCa.

UNIVERSITY CLINIC HAMBURG-EPPENDORF (UCHE) PATIENT COHORT

The cohort used for validating the risk models comprised 1646 men who underwent RP and staging pelvic lymphadenectomy (without neoadjuvant therapy) between 2000 and 2005 at the UCHE. All patients included in the validation set have previously agreed to the scientific evaluation of their data. The characteristics of this validation cohort are presented in Table 1.

STATISTICAL ANALYSIS

All data were analysed using R, version 2.7.1 (http://cran.r-project.org), statistical analysis software. Multinomial logistic regression analysis was performed to develop a model for predicting risk of the four non-ordered pathological stage categories – OC, EPE, SV+ or LN+– based on the JHH cohort. The risk of falling into each pathological stage category was modelled as a function of natural B-spline basis of preoperative tPSA (with knots at the first, second and third quartiles), clinical stage (AJCC-TNM 1992/2002), categorized as T1c, T2a or T2b/c, and biopsy Gleason score, categorized as 5–6, 3 + 4 = 7, 4 + 3 = 7 or 8–10.

The ‘2010 Partin Nomogram’ risk prediction capacity among patients meeting current RP criteria in the JHH cohort was characterized by the ‘predictiveness curves’ using the method developed in Huang et al.[16]. We calculated the risk for each subject in the cohort based on the risk model and made a quantile plot based on the estimated risk. If one considers a point on the curve with an x-coordinate of value v, then the value of its y-coordinate is the 100 ×vth percentile of risk in the population. On the other hand, for a particular point on the curve, if one supposes the value of its y-coordinate is p, then the value of its x-coordinate corresponds to the percentage of subjects in the population with risk ≤ p[i.e., the cumulative distribution function (CDF) of risk at p]. The CDF of risk is a conceptually meaningful measure to clinicians and surgeons. In the present study, we used bootstrap resampling with 1000 replications to generate 95% pointwise CIs for these predictiveness curves.

RESULTS

The demographic, clinical and pathological information of the JHH and UCHE cohorts are shown in Table 1. Compared with the JHH cohort, subjects in the UCHE cohort appear to be older with a higher PSA level (P < 0.01 based on two-sample t-test) and a higher Gleason score; a lower proportion of them are EPE, and a higher proportion are SV+ and LN+ (P < 0.01, based on chi-squared test).

Calibration plots for each pathological stage are shown in Fig. 1. In the JHH cohort, both risk models are in general well calibrated for each pathological stage. For LN+, the model with categorical PSA has satisfying calibration when the predicted risk <0.2, and tends to under-estimate the risk as the predicted risk gets larger; the model with continuous PSA has good calibration when the predicted risk < 0.3. Based on the validation UCHE cohort, the ‘2010 Partin Nomogram’ model again has similar calibration to the 2007 Partin Tables model [7]. Both models tend to slightly under-estimate the probability of OC and over-estimate the probability of EPE in the validation cohort. For LN+, the continuous PSA model is well calibrated when predicted risk <0.3, whereas the categorical PSA model is well calibrated only for predicted risk <0.1. The larger spread of the predicted risk as a consequence of modelling PSA continuously is desirable since individual patients get a more accurate estimate of their risk, especially for SV+ and LN+ pathology staging outcomes.

Figure 1.

Calibration plots for the risk model with categorical PSA and the risk model with continuous PSA estimated from the JHH dataset. The first row shows calibration of the risk models in the JHH dataset. The second row shows calibration of the risk models in the UCHE dataset.

The receiver operator characteristic curves for each pathological stage are shown in the Appendix. While having a similar performance in terms of calibration and discriminatory power, the 2010 Partin Nomogram provides an improved quantification of risk by modelling tPSA as a continuous variable, in comparison with the 2007 Partin Tables [7] where only categorized tPSA level was considered, as shown in the examples in Table 2. Patients 6 and 7, for example, have the same predicted risk of OC (46.6%) based on the 2007 Partin Tables [7]; but with the 2010 Partin Nomogram their predicted probabilities of OC are 51.2 and 44.6%, respectively. This could lead to a different treatment recommendation about RP if, for example, a decision rule was in place to recommend only those patients with OC probability >50%. In another example, Patients 16–20 have an estimated risk of LN+ of 4.0–4.5 based on the 2007 Partin Tables [7], but their predicted risks are much more spread out (2.9–7.3%) when PSA is modelled continuously.

Table 2.  Exemplary results for actual patient-specific risk for the four outcomes predicted by the 2007 Partin Tables and new 2010 Partin Nomogram
 2007 Partin Tables2010 Partin Nomogram
tPSAOCEPESN+LV+OCEPESN+LV+
  1. All patients in the table have clinical stage T2a.

Gleason 5–6         
 Patient 1 4.171.426.9 1.40.474.624.1 1.10.2
 Patient 2 671.426.9 1.40.470.127.8 1.70.4
 Patient 3 6.168.529.3 1.80.369.928 1.70.4
 Patient 4 868.529.3 1.80.36730.5 1.90.6
 Patient 51068.529.3 1.80.36433.3 20.7
Gleason 3 + 4         
 Patient 6 4.146.643.8 7.12.551.241.4 5.91.5
 Patient 7 646.643.8 7.12.544.644.1 8.52.7
 Patient 8 6.142.845.8 9.12.344.444.2 8.62.8
 Patient 9 842.845.8 9.12.340.946.4 9.23.4
 Patient 101042.845.8 9.12.337.648.8 9.64
Gleason 4 + 3         
 Patient 11 4.134.453.7 7.14.839.351.9 62.8
 Patient 12 634.453.7 7.14.83353.5 8.45.1
 Patient 13 6.131.355.4 8.94.332.853.6 8.45.2
 Patient 14 831.355.4 8.94.329.755.2 8.86.3
 Patient 151031.355.4 8.94.326.857 97.1
Gleason 8–10         
 Patient 16 4.139.245.710.74.543.844 9.32.9
 Patient 17 639.245.710.74.536.745.212.95.1
 Patient 18 6.135.54713.5436.545.3135.2
 Patient 19 835.54713.5433.246.813.76.3
 Patient 201035.54713.530.148.6147.3

The predicted risks for each pathological outcome as a function of continuous PSA are shown in Fig. 2 for each clinical stage and biopsy Gleason score category. For patients with a particular clinical stage and biopsy Gleason score, as tPSA level increases, the risk of OC is always decreasing; the absolute risk of SV+ and LN+ is in general monotone increasing; the risk of EPE, on the other hand, tends to increase first and then decrease. In general, the decrease in the risk of EPE starts at lower tPSA levels for patients with higher biopsy Gleason score and clinical stage. In practice, the 2010 Partin Nomogram, as shown in Fig. 3, can be used by clinicians to predict a patient’s clinical stage.

Figure 2.

The 2010 Partin Nomogram: risk of OC, EPE, SV+, LN+ as a function of continuous PSA, clinical stage and biopsy Gleason score.

Figure 3.

The 2010 Partin Nomogram for predicting risk of OC, EPE, SV+, LN+ as a function of continuous PSA (≤20 ng/mL), clinical stage and biopsy Gleason score.

We generated the predictiveness curves for OC, EPE, SV+ and LN+ based on the proposed risk model. These curves, together with their 95% pointwise bootstrap CIs, are shown in Fig. 4. In these curves, for a chosen x-coordinate value v, the corresponding value on the y-coordinate tells us about the 100vth percentile for the predicted risk of each pathological outcome. For example, the 10% risk percentile (95% CI) is 46.6% (43.4–50.0%) for OC, 11.9% (10.4–13.3%) for EPE, 0.59% (0.30–0.83%) for SV+, and 0.11% (0.03–0.19%) for LN+. On the other hand, fixing a value of interest on the y-axis, the corresponding value on the x-axis tells us about the percentage of subjects in the population with risk smaller than the value of interest. For example, suppose there exist predetermined thresholds for OC (low-risk threshold pL= 0.5, high-risk threshold pH= 0.8) such that, if a patient’s risk of OC < pL (>pH), RP is (is not) recommended. Then from Fig. 3, when OC is considered, the CDF (95% CI) for pL is 12.1% (10.1–14.0%), and 1 − CDF for pH is 53.8% (44.9–59.7%). This implies that, given the particular low- and high- risk thresholds, 53.8% of subjects in the population will be recommended RP and 12.1% will not. Subsequently for 34.1% subjects in the population, the treatment recommendation is indecisive based on the current test. This kind of information is useful for clinical decision-makers to gauge the impact of given thresholds for RP. Moreover, for individual patients, knowing the percentage of subjects in the population with a predicted risk less than one’s own is also of value to one’s decision-making process, in addition to the absolute risk itself.

Figure 4.

The predictiveness curves (solid lines) based on the 2010 Partin Nomogram with regression spline of continuous tPSA, clinical stage and biopsy Gleason score, as well as their 95% percentile bootstrap CIs (dashed lines).

DISCUSSION

The accurate prediction of the pathological stage of PCa is critical as it could have an impact on a patient’s selection of treatment options [3]. The pre-treatment status of men diagnosed with PCa today, in the era of PSA screening, has changed, as evidenced by the major shifts that have occurred in tPSA level, biopsy Gleason score and clinical stage at the time of diagnosis, resulting in the increasing probability that they will have pathologically OC disease [3,9,18]. The Partin Tables have evolved over the last 15 years to adjust to major shifts in the presentation of contemporary patients with PCa treated at JHH [4–7].

In the present study, we applied a new model with B-spline basis for tPSA to evaluate the predicted probability (risk) of a specific pathological stage outcome (OC, EPE, SV+ or LN+) for this cohort. The ‘2010 Partin Nomogram’ provides an improved quantification of risk, compared to the 2007 Partin Tables [7], by modelling tPSA as a continuous variable, while achieving a similar calibration and discriminatory performance as shown in a validation set. In addition, we provided the risk quantile plots accompanying the new risk model, which can be used to obtain the percentile value of the risk of a patient with PCa in the cohort and the percentage of subjects in the population who would be recommended to undergo treatment given a risk threshold of interest. This information should prove useful to clinicians in selecting risk thresholds for RP and to individual patients in making treatment decisions.

Notably, in the present study, the JHH patient cohort comprises only 7% African American patients. Therefore the usefulness of Partin Tables for African Americans could be questionable, especially as studies have shown some disparity in the biology of PCa between African American and Caucasian men [19–21]. However, Heath et al.[22], in a large, multi-ethnic cohort of 3748 men, 32% of whom were African American, showed that the 2001 Partin Tables [6] performed equally well in both racial groups despite differences in baseline clinical characteristics. In addition, the percentage of patients in our JHH cohort with PSA level >20 ng/mL was 0.2%, and hence the utility of the risk model developed in men with PSA level >20 ng/mL might be questionable. The risk of OC with biopsy Gleason score 8–10 is slightly higher than that of Gleason score 4 + 3 = 7 because these patients are carefully selected for having limited Gleason score 8–10 cancer on biopsy at JHH after negative evaluation for metastasis (negative bone scan or negative CT). Additional pre-treatment variables such as quantitative pathology information (e.g. number of positive cores, total area of cancer on the biopsies), other PSA derivatives, PSA kinetics and tumour imaging could be useful to predict pathological staging more accurately [23,24]. We plan to attempt such an approach with a database that has all these variables. The use of validated molecular biomarkers from the genome, proteome or metabolome could also prove useful to augment the predictive accuracy for pathological stage prediction of PCa [25,26].

In conclusion, using tPSA as a continuous biomarker to construct the 2010 Partin Nomogram offers a better method to predict pathological stage outcome than the 2007 Partin Tables. We intend to develop a user-friendly computer program to allow the easy use by physicians of the 2010 Partin Nomogram and the corresponding predictiveness curves for patient-specific pathological stage outcome prediction.

CONFLICT OF INTEREST

None declared. Source of Funding: The Patana Fund, The Prostate Cancer Foundation, The Johns Hopkins University Prostate Cancer SPORE (grant number P50CA58236), and the Early Detection Research Network (EDRN) NCI/NIH (grant number CA086323-06).

Appendix

Receiver operator characteristic curves for OC, EPE, SV+, LN+ for the model with categorical PSA and the model with continuous PSA (the first row is based on JHH data and the second row is based on UCHE data). FPF, False Positive Fraction; TPF, True Positive Fraction.

Ancillary