Quantitative biopsy pathology for the prediction of pathologically organ-confined prostate carcinoma

A multiinstitutional validation study

Authors


Abstract

BACKGROUND

Quantitative biopsy pathology with prostate specific antigen significantly improves the prediction of pathologic stage in patients with clinically localized prostate carcinoma (PCa). The authors recently reported a computational model for predicting patient specific likelihood of organ confinement of PCa using biopsy pathology and clinical data. The current study validates the initial models and presents an new, improved tool for clinical decision making.

METHODS

The authors assessed 10 biopsy pathologic parameters and 2 clinical parameters using data from two institutions. Of 1287 patients, 798 men had pathologically organ confined (OC) PCa, 282 men had nonorgan-confined disease with capsular penetration (NOC-CP) only, and 207 men showed seminal vesicle or lymph node invasion (NOC-AD) after undergoing pelvic lymphadenectomy and radical prostatectomy. Patient input data were evaluated by ordinal logistic (OLOGIT) and neural network (NN) models; and the likelihood of developing OC, NOC-CP, or NOC-AD disease was calculated for the combined and separate data sets and was compared with the results from original presentation. In addition, a new two-output model was constructed (OC/NOC-CP vs. NOC-AD).

RESULTS

The three-output OLOGIT and NN models predicted OC disease with 95.0% and 98.6% accuracy, respectively, for the combined data set and with 93.0% and 98.6% accuracy, respectively, on subset analysis. The combined accuracy for predicting OC, NOC-CP, and NOC-AD disease in the entire validation set was 66.7% for OLOGIT model and 66.0% for the NN model. The two-output OLOGIT and NN models correctly predicted 94.9% and 100.0% of all OC/NOC-CP disease, respectively.

CONCLUSIONS

Both computation models predicted OC PCa with an accuracy of 93.0–98.6% when they were validated with two different data sets. The OLOGIT and NN-based, two-output model permitted an appropriate treatment decision for 85.2–90.2% of patients. These data support the use of quantitative pathology and clinical data-based decision modeling to manage patients with clinically localized PCa. Cancer 2003;97:969–78. © 2003 American Cancer Society.

DOI 10.1002/cncr.11153

Predicting the pathologic stage of clinically localized prostate carcinoma is the most important yet most challenging task for the urologist and patient alike. Pathologic stage is a strong predictor of cure in patients with clinically localized prostate carcinoma and, thus, has significant impact on treatment options, e.g. surgical, radiation therapy, or palliative therapies. Likewise, in patients who undergo surgical procedures, the safety of nerve-sparing radical prostatectomy relies on the presence of pathologically organ-confined (OC) prostate carcinoma, because capsular penetration (CP) near the site of the neurovascular bundle may leave tumor tissue behind, which may hamper therapeutic success. Due to the lack of predictive accuracy in contemporary patients with prostate carcinoma, the evaluation of single preoperative parameters, such as the total serum prostate specific antigen (tPSA) level, digital rectal examination (DRE), and/or the Gleason score, does not permit the accurate prediction of pathologic stage on an individual basis. Therefore, numerous computational models have been developed that combine preoperative variables into multimodal decision-support tools. One of the most common predictive tools is the Partin (SPORE) nomogram. These recently validated tables depict the likelihood of OC disease, CP disease, seminal vesicle invasion, or lymph node metastases of clinically localized prostate carcinoma based on the PSA level, Gleason score, and clinical stage using linear regression.1, 2 A model for the same outcome characteristics has been developed (although it has yet not been validated) by Narayan et al. using transrectal ultrasound-guided biopsy, PSA level, and biopsy Gleason score.3 Graefen et al. presented a nomogram for side specific prediction of pathologically OC prostate carcinoma based on the PSA level and the presence of Gleason Grade 4–5 disease in the prostate biopsy.4

Veltri et al. have developed computational models based on ordinal logistic regression (OLOGIT) and genetically engineered neural networks (GENNs) for the prediction of OC disease, nonorgan-confined (NOC)-CP disease, and locally advanced or metastatic (NOC-AD) disease.5 This model was trained with 817 patients and demonstrated the importance of using quantitative prostate biopsy assessment to increase the accuracy of predictive tools. The predictive accuracy for OC disease was 91% and 95% using the OLOGIT and GENN models, respectively, upon validation of the model with an additional 116 patients.

This UroScore™ training data set, however, included a quite diverse range of patients with pathologic follow-up data, which were comprised of detailed pathology and bone scan reports, from community-based, private-practice urologists and a smaller number of patients from academic institutions, whereas the validation sample came entirely from a single community practice. A common observation of algorithms that are developed in one institution is that their predictive accuracy decreases when they are applied to data from another institution.6, 7 In this study, we exposed the original UroScore™ models to a larger number of patients in more pathologically diverse validation set collected from two major demographically disparate referral centers of excellence. We exclusively used data from patients who underwent pelvic lymph node dissection and radical retropubic prostatectomy to verify the performance and validity of UroScore™ in the two surgical academic settings.

MATERIALS AND METHODS

Demographics

Complete data from 1287 patients who underwent pelvic lymphadenectomy and radical retropubic prostatectomy for clinically localized prostate carcinoma were available for evaluation. Data were collected from the University Hospital of Hamburg (n = 894 patients) and Johns Hopkins Hospital (n = 393 patients). Preoperatively, all patients underwent a systematic sextant biopsy. Patients who received neoadjuvant antiandrogen therapy and patients who did not undergo lymph node dissection during radical prostatectomy were excluded from this evaluation. Postoperative pathologic evaluation employed the most recent 1997 TNM staging criteria. In all patients, the 10 pathologic biopsy parameters listed in Table 1 were available. The clinical parameters used were patient age and serum tPSA. PSA results from all patients were obtained on two different equimolar total PSA assays and were categorized into increments of 2 ng/mL, as described previously.5 The categorized PSA levels were entered as ordinal levels (level 1, PSA 0–2 ng/mL; level 2, PSA > 2–4 ng/mL; level 3, PSA > 4–6 ng/mL; etc.).

Table 1. Descriptive Statistics of Parameters for the Ordinal Logistic and Neural Network Model Split by Patients from the University Hospital of Hamburg and Johns Hopkins Hospital
VariableUsed by OLOGIT/NNHAMJHH
MeanMedianRangeMeanMedianRange
  1. OLOGIT: ordinal logistic model; NN: neural network model; HAM: University Hospital of Hamburg; JHH: Johns Hopkins Hospital: PSA: prostate specific antigen; OC: organ confined: NOC-CP: nonorgan confined-capsular penetration; NOC-AD: nonorgan confined-locally advanced/metastatic.

Preoperative parameter       
 Age (yrs)+/+62.263.043–7359.660.038–73
 PSA (ng/mL)+/+12.28.130.24–1257.86.30.4–43.5
 No. of cores positive for carcinoma+/+2.62.01–62.22.01–6
 Highest Gleason score+/+6.46.04–96.26.04–9
 Average % tumor involvement per core+/−16120–911070–71
 Presence of Gleason pattern 4/5 (no/yes)+/−55534231179
 Midcore with > 5% tumor (no/yes)+/+254643215175
 Base/midcore with > 5% tumor (no/yes)+/−88809155235
 Total % tumor involvement−/+99.6701–54560401–420
 Average % tumor involved/positive core−/+35301–10025201–100
 Base core with > 5% tumor (no/yes)−/+251643241152
 Apex core with > 5% tumor (no/yes)−/+365529207186
Postoperative characteristic (%)       
 OC514(57.5)284(72.3)
 NOC-CP198(22.1)84(21.4)
 NOC-AD182(20.4)25(6.4)

Computational Models

Statistical analyses employed the previously described statistical modeling techniques available from the Stata software program (version 7.0; Stata Corp., College Station, TX).5

OLOGIT-based three-output model

The OLOGIT model used the eight parameters (two clinical and six pathologic) listed in Table 2. The original three-output model β coefficients, which were derived by means of backward, stepwise, multivariate logistic regression analysis at a stringency of P < 0.20 from these independent variables, were utilized to deliver a patient specific score. The results were expressed as the probability that patients would have OC prostate carcinoma, NOC-CP prostate carcinoma, and NOC-AD prostate carcinoma (Table 2, left column). We used our previously published cut-off value of ≥ 35% for the three-outcome model as follows: If the probability that a patient would have NOC-AD disease was ≥ 35%, then it was predicted that the patient would have NOC-AD disease. If the probability of NOC-AD was < 35% but the probability of NOC-CP was ≥ 35%, then the patient was assigned to the NOC-CP group without AD. Finally, if both probabilities for NOC-AD and NOC-CP were < 35%, then it was predicted that the patient would have OC disease. For the three-output model, published cut-off points for each patient's score, adjusted to maximize the prediction of OC disease, are used to classify the patient into an OC group, an NOC-CP group, or an NOC-AD group. The cut-off score for separating OC disease from NOC-CP disease was 6.408712, whereas the cut-off score for separating NOC-CP disease from NOC-AD disease was 7.960772.

Table 2. β Coefficients for the Ordinal Logistic Model to Calculate the Probability of Organ-Confined Prostate Carcinoma versus Nonorgan-Confined Prostate Carcinoma with Capsular Penetration versus Nonorgan-Confined, Locally Advanced/Metastatic Prostate Carcinoma and for the Probability of Organ-Confined and Nonorgan-Confined Prostate Carcinoma with Capsular Penetration versus Nonorgan-Confined, Locally Advanced/Metastatic Prostate Carcinoma
CharacteristicOC vs. NOC-CP vs. NOC-ADOC + NOC-CP vs. NOC-AD
  1. OC: organ confined; NOC: nonorgan confined; CP: capsular penetration; AD: locally advanced/metastatic; PSA: prostate specific antigen.

Age0.03017920.0664568
PSA, categorized in 2-ng/mL increments0.10952030.1241981
No. of cores positive for carcinoma0.0743832−0.104016
Highest Gleason score0.4427290.6585678
Average % tumor involvement/core0.02917170.0287586
Presence of Gleason pattern 4/5 (no/yes)−0.3386806−0.3774724
Midcore with ≥ 5% tumor (no/yes)−0.5089567−0.0195107
Base and/or midcore with > 5% tumor (no/yes)0.62261910.5674477

OLOGIT-based two-output model

The two-output model used the same eight clinical and pathologic parameters that were used for the three-output model. These parameters were examined using backward, stepwise, multivariate logistic regression analysis at a stringency of P < 0.20 for their ability to predict OC plus NOC-CP disease versus NOC-AD disease. For the two-output model, the cut-off point for discriminating OC disease versus NOC disease was a score of 11.70738. The derived β coefficients of the two-output analysis are shown in the right column of Table 2.

The exponential formula for calculating the patient specific score for either the two-output or the three-output OLOGIT model is presented in Table 3. In brief, the calculation is as follows: For either the two-output model or the three-output model, the formula multiplies the β coefficient A with the independent variable A, the β coefficient B with the independent variable B…, etc., and calculates the sum, which results in the patient specific score for each individual patient.

Table 3. Formulae for Uroscore™ and Outcome Probability Calculation of the Ordinal Logistic Model
ModelFormulaa
  • OC: organ confined; NOC: nonorgan confined; CP: capsular penetration; AD: locally advanced/metastatic; PSA: prostate specific antigen.

  • a

    x1–x8: independent variables of the OLOGIT model; β1–β8: β coefficients of the independent variables x1–x8; e: natural log function.

Model a: OC vs. NOC-CP vs. NOC-AD (three-output model) 
 Predicted score (S)(x1 * β1) + (x2 * β2) + …(x8 * β8)
 Cut point for OC vs. NOC-CP carcinoma (c1)6.408712
 Cut point for NOC-CP vs. NOC-AD carcinoma (c2)7.960772
 Probability of OC carcinoma1/(1 + eS − c1)
 Probability of NOC-CP carcinoma1 − [1/1(1 + eS − c2)] − 1/(1 + eS − c1)
 Probability of NOC-AD carcinoma1 − [1/1/(1 + eS − c2)]
Model b: OC + NOC-CP vs. NOC-AD (two-output model) 
 Predicted score (S)(x1 * β1) + (x2 * β2) + …(x8 * β8)
 Cut point for OC + NOC-CP vs. NOC-AD11.70738
 Probability of OC + NOC-CP1/(1 + eS − c1)
 Probability of NOC-AD carcinoma1 − [1/(1 + eS − c1)]

NN model

In our initial publication, the NeuroGenetic Optimizer™ program (version 2.6; BioComp Systems, Inc., Redmond, WA), which is software that builds predictive models using GENNs, was used to construct and validate the three-outcome pathologic stage prediction model for the 817 patient sample. In this study, we used the next generation of NN software (IUnderstand™, version 1.4) from the same company. The current network utilized an error back-propagation learning algorithm with a genetic preselection component and had a design similar to that of the original GENN model.5 The new NN used the same 9 input variables (see Table 2), 2–3 hidden layers, and between 1 and 124 nodes per layer. The final, optimized architecture was determined by the software as a result of a best-fit network. We employed 200 iterations to produce the 3-output model (OC, NOC-CP. or NOC-AD) or the 2-output model (OC + NOC-CP vs. NOC-AD). To ensure the comparability of the earlier GENN compared with the current NN architecture, we analyzed the 817 patients from our initial publication with this newly designed network and achieved virtually identical performance in terms of predictive accuracy for OC, NOC-CP, and NOC-AD disease (data not shown). In the same manner, the 817 patients from the initial publication served as a training set for the new two-output model (discussed below).

RESULTS

Preoperative Data

Complete preoperative characteristics for all examined clinical and pathologic parameters that were used in either the OLOGIT model or the NN model are shown in Table 1.

Postoperative Pathologic Data

Pathologic work-up of 1287 validation patients revealed OC disease in 798 patients (62%), NOC-CP disease in 282 patients (22%), and NOC-AD disease in 207 patients (16%). The distribution of each pathologic parameter in patients from the Johns Hopkins Hospital and the University Hospital of Hamburg is shown in Table 1 (postoperative characteristics). Of 207 combined patients with NOC-AD disease, 134 patients had pT3bN0 disease, whereas 73 patients had pN1 disease.

Three-Output Model Evaluation Using the OLOGIT Model

When it was exposed to the combined data, the OLOGIT model correctly classified 758 of 798 patients (95%) in the validation set with OC disease. This performance was maintained in the subset analysis of the two institutions, in which the model correctly identified 93.0% (University Hospital of Hamburg) and 98.6% (Johns Hopkins Hospital) of all patients with OC disease. In the combined evaluation, the model misclassified only 3 of 798 patients who had pathologic OC disease (0.37%) with NOC-AD disease, a decision that would preclude local therapy or, at a minimum, would require additional clinical work-up.

Of 282 patients with NOC-CP disease, overall, the OLOGIT model classified 13.8% correctly. Notably, the model predicted NOC-AD disease in 22 of 282 patients. With the same intention of not treating patients with NOC-AD disease locally, as stated above, this would affect 7.8% of patients. NOC-AD disease was identified correctly in 30% of patients overall; however, there was a marked difference in the performance of the two validation data sets of 32.4% (University Hospital of Hamburg) and 12% (Johns Hopkins Hospital).

Three-Output Model Evaluation Using the NN Model

Analysis of the combined data sets using the NN model resulted in the correct prediction of 787 of 998 patients (98.6%) with OC disease. In the subset analysis, the model correctly identified 97.8% of patients at the University Hospital of Hamburg and 96.8% of patients at Johns Hopkins Hospital with OC disease. The model falsely allocated only 11 of 798 patients (1.38%) with pathologic OC disease to the NOC-AD disease category. Compared with the OLOGIT model, the NN model failed to identify NOC-CP disease in the combined data set or when the data from the University Hospital of Hamburg and Johns Hopkins Hospital were analyzed separately.

With respect to predicting NOC-AD disease, the NN model classified 30.4% of patients correctly, with minimal difference between subsets (32.0% vs. 31.8%). Complete data for OLOGIT and NN analysis for the validation subsets are shown in Table 4.

Table 4. Observed versus Predicted Outcome of the Ordinal Logistic and Neural Network Three-Output Model for the Hamburg Data, the Johns Hopkins Data, and Combined Data
Actual outcomeNo.OLOGIT modelNeural network
Predicted outcomeClassified correctly (%)Predicted outcomeClassified correctly (%)
OCNOC-CPNOC-ADOCNOC-CPNOC-AD
  1. OLOGIT: ordinal logistic; OC: organ confined; NOC: nonorgan confined; CP: capsular penetration; AD: locally advanced/metastatic.

a Hamburg validation set
 OC51447833393.050301197.8
 NOC-CP198143361918.21730250.0
 NOC-AD18290335932.412405831.8
 Overall89464.262.7
b Hopkins validation set
 OC2842804098.62750996.8
 NOC-CP8478333.681030.0
 NOC-AD25175312.0170832.0
 Overall39372.572.0
c Combined validation set
 OC79875837395.078701198.6
 NOC-CP282221392213.82540280.0
 NOC-AD207107386230.014406330.4
 Overall128766.766.0

Comparison of Both Models with the Original Presentation

The overall performance of the OLOGIT model was 66.7%, compared with 68.1% in the initial study for the three-output model (Fig. 1). Most important, the predictions of OC disease (90.8% vs. 95%) and NOC-AD disease (33.3% vs. 30.0%) were similar in the initial study and the new validation study, respectively. By contrast, the performance decreased in the NOC-CP group from 22.6% to 13.8%. The overall performance of the NN model was 66%, which compared favorably with the performance of 67.2% in the initial study. The model predicted 98.6% of all OC disease correctly, compared with 97.4% in the initial study.

Figure 1.

Comparison of original and present performance of the ordinal logistic regression OLOGIT (a) and neural network (NN) model (b) when applied to the current combined data from patient groups at Johns Hopkins Hospital and the University Hospital of Hamburg with prostate carcinoma. OC: organ confined; NOC: nonorgan confined; CP: capsular penetration; AD: locally advanced/metastatic. *, three outcome model: OC vs. NOC-CP vs. NOC-AD.

Two-Output Clinical Management Model using the OLOGIT and NN Approaches

Training set

Using the described cut-off value of ≥ 35% with the goal of maximizing correct OC prediction, 645 of 666 NOC-CP tumors (96.8%) and 47% of NOC-AD tumors were assigned correctly (Table 5, training set). The NN model performed equally well in predicting 95.2% of all OC tumors but improved the prediction of NOC-AD tumors to 60.3%.

Table 5. Observed versus Predicted Outcome of the Ordinal Logistic and Neural Network Two-Output Model for the Two-Output Training Set, the Hamburg Data, the Johns Hopkins Data, and Combined Data
Actual OutcomeNo.OLOGIT modelNeural network
Predicted outcomeClassified correctly (%)Predicted outcomeClassified correctly (%)
OC + NOC-CPNOC-ADOC + NOC-CPNOC-AD
  1. OLOGIT: ordinal logistic; OC: organ confined; NOC: nonorgan confined; CP: capsular penetration; AD: locally advanced/metastatic.

a Training set
 NC + NOC-CP6666452198.86343295.2
 NOC-AD151807147.0609160.3
b Hamburg validation set
 NC + NOC-CP7127001298.36763694.9
 NOC-AD1821542815.47310960.0
c Johns Hopkins validation set
 NC + NOC-CP3683680100.0364498.1
 NOC-AD252414.02328.0
d Combined validation set
 NC + NOC-CP108010681298.910404096.3
 NOC-AD2071782914.08512258.9

Separate and combined validation sets

The models were able to predict correctly 94.9% (NN) and 100% (OLOGIT) of all OC/NOC-CP clinically localized disease in the separate evaluation of the two data sets, resulting in the correct identification of 96.3% (NN) and 98.9% (OLOGIT) of all OC/NOC-CP disease (Table 5: University Hospital of Hamburg, Johns Hopkins Hospital, and combined validation sets). Compared with a correct identification rate of 47.0% of NOC-AD disease in the training set, the performance of the OLOGIT model in predicting NOC-AD disease upon validation decreased to 14% (combined), 15.4% (University Hospital of Hamburg), and 4% (Johns Hopkins Hospital). However, the NN model performed significantly better in the same setting, with a combined overall correct prediction rate of 58.9% for NOC-AD disease. However, the combined validation set in Table 5 shows that the model worked better on the data from the University Hospital of Hamburg compared with the data from Johns Hopkins Hospital (60% vs. 8% correct prediction, respectively), indicating a significant difference in the input parameter profile of these two subsets in relation to NOC-AD disease.

DISCUSSION

The prediction of pathologic stage remains a fundamental requirement for the appropriate counseling of a patient with newly diagnosed, clinically localized prostate carcinoma. Because there is limited clinical outcome predictive value when using individual parameters (clinical stage, serum PSA, Gleason score or grade of the biopsy, imaging analysis for DNA ploidy), a number of multimodal staging tools have been developed. These computational models incorporate several clinical variables to predict specific end points, most importantly, pathologic stage2–4 or disease free survival.8–10 The validity of these tools, however, may become challenged when the test populations come from demographically unrestricted sources. Under such circumstances, it is not uncommon for the performance of a decision-support tool to deteriorate when it is exposed to new data.6, 7 Therefore, tools that do not withstand a rigorous validation from the same or another institution may not be useful to aid in clinical decision making.11

Due to increased PSA screening and public awareness, there is a predominance of clinical Stage T1c prostate carcinoma in the United States population, and the pathologic grade and stage of these tumors has decreased.5, 12–14 Likewise, PSA levels and standard biopsy pathology have consolidated and become more similar (Gleason scores of 6 and 7) among patients.15 Therefore, it is a more challenging task for the urologist to predict the pathologic stage of tumors that clinically present so uniformly. Additional information derived from a more detailed evaluation of the prostate biopsy is readily available, has provided critical information in the past, and currently may outperform the Gleason score on multivariate analysis.16 For example, Wills et al.17 and Badalament et al.18 identified high Gleason score and number of biopsies positive for carcinoma as the two most important variables for pathologic stage prediction. Other evidence comes from Sebo et al.19 and Badalament et al.,18 who noted that the percent of cores positive for carcinoma and surface area positive for carcinoma were the best predictors for pathologic stage and tumor volume using multivariate logistic regression analysis that also included PSA, age, clinical stage, and Gleason score. Aside from the emerging evidence for more detailed biopsy assessment, other parameters, like DNA-ploidy,18 quantitative nuclear morphometric analysis,14, 20 or other serum and tissue markers, may provide information to overcome the challenge of the optimal staging for patients with prostate carcinoma.

The literature supports the fact that postoperative recurrence rates increase with advancing clinical stage, Gleason score, preoperative PSA level, and pathologic stage.2, 9, 16, 21–23 Recently, Han et al.21 studied a series of 2404 men who underwent radical prostatectomy with a mean follow-up of 6.3 ± 4.2 years (range, 1–17 years), and those authors found a recurrence rate of 17%, which broke down to 9.7% biochemical recurrences, 1.7% local recurrences, and 5.8% distant recurrences. The overall actuarial 5-year, 10-year, and 15-year recurrence free survival rates for these men were 84%, 74%, and 66%, respectively. The importance of knowing the anatomic extent of disease pretreatment relates to the fact that approximately 30% of men who are treated for localized disease will develop recurrent disease, and a subset of these men will develop progressive disease.13, 21–24 Most patients who are diagnosed early with OC tumors are curable with radical prostatectomy21, 23, 25, 26 (90–95%) or with radiation therapy (80–95%).25, 26 Patients who have locally disseminated disease at the time of diagnosis are managed best with more aggressive therapy that may involve combinations of surgery with adjuvant treatments or, in patients with distant metastatic disease, with ablative hormonal therapy.25

In our previous study,5 we demonstrated the substantial gain in diagnostic accuracy when a more detailed evaluation of prostate biopsies, termed quantitative prostate biopsy pathology, is used to develop and challenge a statistical (logistic regression) or NN model to predict prostate carcinoma stage based on preoperative parameters. These models were developed previously and underwent internal validation with a subset of 116 new patients, and their high accuracy (> 90%) was validated for identifying patients with OC prostate carcinoma.

In the current study, we validated the performance of these UroScore™ models using an expanded, collaborative, external data set from two demographically disparate locations. These data were derived from one United States cancer center of excellence and one German cancer center of excellence with extensive experience in radical retropubic prostatectomy. In both centers (although more in the United States center compared with the German center), the recent trend in increasing numbers of early, PSA-detected prostate carcinoma with accompanying increases in clinical Stage T1c tumors and pathologic Stage T2 tumors12–14, 27 has been observed for several years as a result of increased use of tPSA and much improved public awareness for early detection purposes. This trend in the United States precedes the setting in Germany by 3–5 years. Consequently, in the German center, the clinical and pathologic tumor features reveal a lower percentage of pT2 tumors (57%) compared with the United States center (72%) and higher tPSA concentrations (median, 8.13 ng/mL vs. 6.3 ng/mL). If the trend seen in Germany (which, to date, mirrors the trend in the United States) continues, then the clinical and pathologic distributions seen in the German data likely will be similar to the United States data. Despite the difference in the pathologic features, both data sets show a dominance of pathologically OC disease and lesser numbers of pT3a and ≥ pT3b tumors. When both data sets were combined, they showed distributions of OC disease, NOC-CP disease, and NOC-AD disease comparable to the distributions shown in the initial presentation. Therefore, the data sets, either alone or combined, seem to be appropriate for validation purposes.

The validation of the original model (prediction of OC, NOC-CP, or NOC-AD disease) using either the OLOGIT approach or the NN approach produced comparable results by identifying 95–98.6% of all pathologically OC tumors correctly (Fig. 1), a result that compares favorably with the initial model, which had an accuracy of 90.8–97.4%. In the subset analysis of the two institutions, this performance persisted, in that 93% (University Hospital of Hamburg data validation set, OLOGIT) and 98.6% (Johns Hopkins Hospital validation set, OLOGIT) of all OC disease was classified correctly.

A notable observation is that the OLOGIT model performed differently on NOC-CP and NOC-AD prediction between the two validation sets by correctly identifying 18.2% versus 3.6% of patients from the University Hospital of Hamburg and 32.4% versus 12%, of patients from Johns Hopkins Hospital. Table 4 reveals that the OLOGIT model allocated 143 of 198 patients (72%) at the University Hospital of Hamburg to the NOC-CP group and 78 of 84 patients (93%) at Johns Hopkins Hospital to the OC group but allocated only 19 of 198 patients at the University Hospital of Hamburg and 3 of 84 patients at Johns Hopkins Hospital to the NOC-AD group. With respect to NOC-AD prediction, the NN model performance was comparable to the OLOGIT model performance in the University Hospital of Hamburg validation set (31.8% correct predictions) and almost tripled its performance from 12% to 36% in the Johns Hopkins Hospital validation set, whereas it failed to identify any NOC-CP disease correctly, again allocating 173 of 198 patients with NOC-CP disease at the University Hospital of Hamburg and 81 of 84 patients with NOC-CP disease at Johns Hopkins Hospital to the OC group.

Our interpretation of this phenomenon is two-fold: First it may be assumed that a large proportion of NOC-CP tumors are more suggestive biologically of OC tumors; therefore, the parameters applied in this model are not as reliable for identifying these subtly different tumors. The biologic outcome of surgically removed tumors that demonstrate CP on pathologic examination supports the assumption of a comparable behavior of at least a proportion of pT3 tumors. In a study of 721 men at Johns Hopkins Hospital, 58% of men with evident CP had no evidence of biochemical recurrence 10 years after surgery.28 Second, for the differences in performance between the two validation sets, the demographics (Table 1) show that more NOC-CP disease was found in the University Hospital of Hamburg data base, which may increase the likelihood that a model will identify patterns more typical to NOC-CP disease.

From a clinical point of view, patients who present with a clinically localized prostate carcinoma should be offered curative treatment approaches. A considerable proportion of patients with clinically OC disease demonstrate pathologically NOC disease; however, the likelihood of cure for patients with only NOC-CP disease, as discussed above, can reach almost 60%.28

The following is one example of how the OLOGIT three-outcome model may be used in practice by a physician. For a hypothetical patient age 65 years with a tPSA level of 7.0 ng/mL, 3–6 positive biopsy cores (left and right base and right midcore), total tumor involvement of 125%, and a Gleason score of 7 (3 + 4), the patient specific probabilities of OC disease, NOC-CP disease, and NOC-AD disease patient would be 60%, 28%, and 12%, respectively. This means that the patient has an 88% probability of having specimen-confined disease (60% OC + 28% NOC-CP = 88%). If the physician feels that this patient potentially is curable with definitive therapy, then the combined OC/NOC-CP probability of 88% would be used to counsel this patient toward treatment options such as curative surgery or radiotherapy. However, in another example, a patient age 65 years with a tPSA level of 4.5 ng/mL, a Gleason 6 tumor, 2–6 positive biopsy cores (left and right apex), and total tumor involvement of 50% would have patient specific OC, NOC-CP, and NOC-AD probabilities of 76%, 18%, and 6%, respectively. This means that the second patient has a 94% probability of having specimen-confined, small tumor burden disease (76% OC + 18% NOC-CP = 94%). The physician may counsel this second patient, based on either current health status or personal choice, to choose deferred treatment and select expectant management with curative intent and to include aggressive monitoring for changes in PSA velocity, PSA density and an annual biopsy, such as that recommended by Carter et al.29 Alternatively, in the event that the OLOGIT probabilities demonstrated a high potential for NOC disease (i.e. probabilities of 35% for NOC-CP disease and ≥ 25% for NOC-AD disease), then the physician would carefully consider all of the probabilities and would present several treatment options and their associated risks to the patient. The range of these treatment options likely would be based on possible additional testing to attempt to rule out local, regional, or distant metastasis. Based on the additional test results, the patient would be counseled about various treatment options and their inherent risks.

In summary, patients who were predicted by either computational model to have OC or NOC-CP disease would be eligible for local, curative therapeutic options (e.g., radical prostatectomy or brachytherapy with or without external beam radiotherapy); whereas patients who were predicted to have NOC-AD disease (seminal vesicle invasion or lymph node metastases) are given palliative treatment options. The numbers of patients who are allocated correctly and incorrectly to either curative or palliative treatment may provide a clinically applicable measure for the performance of the two models. From this point of view, we established an OLOGIT and NN model to produce a two-output result by merging the OC group and the NOC-CP group and opposing it to the NOC-AD group.

The results of the training and validation sets (Table 5) reveal that either model was able to identify 94.9–100% of all patients with OC/NOC-CP disease in the analysis of the separated or combined data sets. Table 5 shows that a total of 1097 patients (1068 patients with OC/NOC-CP disease and 29 patients with NOC-AD disease) were classified correctly using the OLOGIT model. This results in appropriate treatment decisions for 1097 of 1287 patients (85.2%). The same analysis for the NN shows an even better ratio: 1162 of 1287 patients (90.2%) will be classified correctly.

Conclusions

Using either OLOGIT or NN analysis, OC prostate carcinoma was predictable with an accuracy of 93.0–98.6% when the models were validated with two data sets from two major demographically disparate centers. The OLOGIT and NN-based, two-output model permitted an appropriate treatment decision for 85.9–90.2% of patients. These data continue to support the use of algorithms that combine pretreatment quantitative pathology and clinical data, and they provide a valuable aid in treatment decision making for patients with clinically localized prostate carcinoma.

Acknowledgements

The authors thank their academic collaborators from Ohio State University, the University of Michigan, Baylor College of Medicine, Albany Medical College, and Johns Hopkins Medical Institutions for their cooperation on this project. They also thank the numerous community-based, private-practice urologists who provided the pathologic staging outcomes for a large set of patients included in the training set.

Ancillary