European Urology

1. Introduction

In recent years, numerous regression model-based nomograms have been introduced to predict either pathologic stage or cancer control outcome after radical prostatectomy on the basis of pretreatment tumor characteristics [1], [2], [3], [4], [5], [6], [7], [8], and [9]. Because of the their inability to perfectly predict the targeted outcome, the search for additional predictors continues.

Recent studies suggest that the amount of cancer and high-grade cancer in diagnostic systematic biopsies of the prostate may improve the accuracy of predictive and prognostic models [6], [7], [10], [11], [12], [13], and [14]. The predictive ability of these two variables defined by the number and the percentage of positive cores has been suggested by some investigators [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], and [20], but was also refuted by others [15], [21], and [22]. The reason for this discrepancy stems from the fact that no study compared the predictive accuracy (PA) of these two variables in a head-to-head fashion. Similarly, no previous single study assessed the effect of these two variables on several simultaneous outcomes, such as their effect on all pathologic stages, as well as on the rate of biochemical recurrence (BCR). Therefore, it is difficult to reconcile some of the conflicting results of these reports [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], and [20].

To address this issue, we performed a systematic and comprehensive analysis of the ability of either the number of positive cores or the percentage of positive cores to predict the pathologic stage and the rate of BCR. Our study spanned a period of over a decade, during which time the biopsy schemes drastically changed. This is reflected by a wide range of the number of biopsy cores that were obtained (2–36). Detailed information about the distribution of the number of biopsy cores taken are shown in Table 1.

2. Material and methods

Table 2 shows patient characteristics of each cohort enrolled. Patients with pretreatment prostate specific antigen (pPSA) values >50 ng/mL were excluded. All preoperative biopsies were performed under transrectal ultrasound guidance. PSA failures were defined as a value of ≥0.1 ng/ml and rising. None of the patients received neoadjuvant therapy or adjuvant treatment before evidence of recurrence. Patients without biochemical failure were censored at the date of last available pPSA value.

We focused on prediction of six different end points: organ confinement (OC), extracapsular extension (ECE), seminal vesicle invasion (SVI), lymph node invasion (LNI), the rate of BCR predicted with clinical variables, as well as the rate of BCR predicted with pPSA and pathologic variables. The prediction of BCR with clinical variables henceforth will be referred to as preoperative prediction of BCR. Conversely, the prediction of BCR with pPSA and pathologic variables henceforth will be referred to as postoperative prediction of BCR. Each end point was addressed in a separate analysis. Since the number of missing fields varied from one analysis to another, six different cohorts were developed. These ranged in size from 1783 to 3374 patients. Details about the exclusion of patients because of missing data are shown in Table 3. All patients were treated with radical retropubic prostatectomy (RRP) for _cT_anyN₀M₀ prostate cancer at two European and one North American institution between January 1992 and July 2005.

In all six models, the main predictors of interest consisted of the number and the percentage of positive biopsy cores. The percentage of positive cores was defined as the sum of all cores that contained invasive cancer divided by the sum of all cores that were taken. The contribution of these variables to multivariable (MVA) models represented the focal point of all analyses. Statistical tests were performed with S-PLUS Professional, version 1. Two-sided tests with significance at 0.05 were used. Logistic regression models tested the added value of the variables of interest to pPSA, clinical stage, and biopsy Gleason sum, and were complemented with models consisting of pPSA, clinical stage, and biopsy Gleason sum, in prediction of pathologic stages at RRP: OC, ECE, SVI, and LNI.

Cox regression models relying on pPSA, clinical stage, and biopsy Gleason sum were used for preoperative prediction of BCR. Conversely, logistic regression models relying on pPSA, clinical stage, and biopsy Gleason sum were used for postoperative prediction of BCR. PA of the MVA models was assessed by calculation of the area under the receiver operating characteristics curve (AUC). We used 200 resamples to reduce overfit bias and to internally validate the MVA estimate of PA. In Cox regression models, the AUC was substituted with Harrell's concordance index. A value of 100% indicates perfect predictions, while 50% is equivalent to a toss of a coin. The statistical significance of the increment in PA related to the addition of the variable(s) of interest was tested with the Mantel-Haenszel test.

3. Results

The descriptive variables of the six cohorts are shown in Table 2. Of all patients, 2285 of 3353 (68.1%) had OC, 731 of 3367 (21.7%) had ECE, 351 of 3326 (10.6%) had SVI, and 132 of 3374 (3.9) had LNI. Mean pPSA ranged from 8.8 to 9.4 (0.1–50.0) ng/ml. Clinical stage distribution was T1c in 61.5–64.8%, T2 in 34.2–37.2%, and T3 in 1.0–1.3% of the patients, according to the end point being analyzed. Biopsy Gleason sum of 2–5 was found in 3.9–7.2%, 6 in 57.7–58.3%, 7 in 30.3–33.8%, and 8–10 in 3.8–4.3% of the patients.

The biopsy core information was distributed as follows. The mean number of total cores taken ranged from 7.4 to 8.4 (2–36), and the mean number of positive cores ranged from 2.6 to 2.9 (1.9), while the mean percentage of positive cores ranged from 36.0% to 37.3% (2.8–100%). BCR was diagnosed in 365 of 1783 (20.5%) patients included in the cohort who were used for the preoperative prediction of BCR (median follow-up: 7.9 yr), and in 369 of 1820 (20.3%) patients included in the cohort who were used for postoperative prediction of BCR (median follow-up: 7.9 yr). Pretreatment PSA, clinical stage, and biopsy Gleason sum (base model) were independent predictors of pathologic stages (p < 0.001). Moreover, when the base variables were complemented with either the number (base model + number of positive cores) or the percentage of positive cores (base model + percentage of positive cores) or with both (base model + number and percentage of positive cores), they maintained independent predictor status in all pathologic stage predictions (p ≤ 0.001).

In MVA analyses, when the variable that defined the number of positive biopsy cores was added to the base model, independent predictor status was reached in OC status (Table 4), and in SVI and LNI (Table 5). The variable defining the number of positive cores also reached independent predictor status in preoperative models predicting BCR, as well as in postoperative models predicting BCR (Table 6). Conversely, the number of positive cores failed to reach independent predictor status in ECE analyses.

The MVA effect of the variable defined by the percentage of positive cores achieved independent predictor status in models addressing OC and ECE (Table 4), and SVI and LNI (Table 5), as well as in preoperative models predicting BCR (Table 6). Conversely, percentage of positive cores failed to reach independent predictor status in postoperative models predicting BCR.

Assessment of the effect of either the number or the percentage of positive cores on PA gains in the prediction of pathologic stage revealed that the maximum increase in PA was related to the consideration of the percentage of positive cores in the model predicting OC status (+1.4%), ECE (+0.8%), SVI (+1.5%), and LNI (+1.4%) (all p < 0.001). Consideration of the number of positive cores resulted in lower PA gains. Finally, consideration of both variables resulted in marginal PA gains relative to the inclusion of percentage of positive cores alone (Table 7).

Assessment of the effect of either the number or the percentage of positive cores on the PA gain in the prediction of BCR revealed that the maximum increase in PA was related to the consideration of the number of positive cores in both preoperative (+1.1%, p < 0.001) and postoperative (+0.3%, p = 0.006) models predicting BCR. Consideration of the percentage of positive cores yielded lower PA gains. Finally, consideration of both variables resulted in either the same PA or lower PA relative to the inclusion of the number of positive cores (Table 7).

4. Discussion

Accurate prediction of pathologic stage and outcome after definitive treatment for localized prostate cancer is important for patient counseling, follow-up, and treatment planning. Recently, many prognostic tools incorporating various clinical parameters have been created to increase disease staging accuracy and relapse predictions [1], [2], [3], and [4]. Many of them have also included detailed biopsy information (such as number and/or percentage of positive cores), which was complemented with established prognostic factors such as pPSA, clinical stage, and biopsy Gleason grade [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], and [22]. The rationale for inclusion of core-derived information relates to their ability to identify adverse pathologic findings and predict BCR [13], and [14].

Despite independent predictor status, Graefen and colleagues [15] showed that the number and/or the percentage of positive cores might not actually add to the PA. From a practical perspective, a predictor that is significant, but does not improve the discriminant ability of a MVA model, is of no practical use. We tested this dichotomy in all of the addressed outcomes, namely in prediction of pathologic stages, preoperative prediction of BCR, and postoperative prediction of BCR.

To the best of our knowledge, our report represents the first systematic and comprehensive analysis of the importance of the information contained within the variable that codes either the number or the percentage of positive cores. Moreover, our analysis represents the first systematic and comprehensive assessment of PA gains related to the consideration of either of these variables in MVA models.

Our analyses of the independent predictor status of the number and the percentage of positive cores confirmed the importance of these variables in the prediction of pathologic stage, except for ECE predictions. Interestingly, when the number and the percentage of positive cores were combined within the same model, the number of positive cores was no longer an independent predictor of either SVI or LNI. In models predicting BCR with preoperative variables, the number and the percentage of positive cores also achieved independent predictor status when only one variable was included at a time. When number and percentage were considered simultaneously, neither reached independent predictor status.

Taken together, these data indicate that either the number or the percentage of positive cores is virtually invariably independent predictors of all pathologic or biochemical outcomes. Moreover, when these variables are combined, they fail to independently contribute to the model. Lack of independent predictor status of the number and the percentage of positive cores in models that combine both variables implies that these variables quantify highly interrelated information. Consideration of this information within the same model results in detrimental effects on the model's PA. From a practical standpoint, this observation makes sense, because inclusion of one variable in absolute format (number of cores) and its simultaneous inclusion in the relative format (percentage of cores) do not add novel information to the model.

Although testing of independent predictor status is important statistically, it is somewhat secondary from a practical standpoint. Independent predictor status does not help to identify individuals at risk of unfavourable pathologic stage or men at risk of BCR. Conversely, a variable that increases PA helps discriminate between individuals destined to fail and those at low risk of failing. Similarly, an informative variable also helps to identify those at high risk of pathologically adverse prostate cancer.

Our study attempted to measure both effects, namely the presence of independent predictor status as well as the increase in PA, when information derived from biopsy cores was considered. We addressed this issue by comparing bootstrap-corrected PA of models with and without data derived from biopsy cores. Our analysis showed that either the number or the percentage of positive cores significantly increased the PA (p < 0.001; Table 7). This benefit was noted in both pathologic and BCR models. It is of interest that the consideration of the percentage of positive cores was invariably related to higher PA gains in models addressing pathologic stage. Conversely, the consideration of the number of positive cores led to higher PA in BCR models.

In all six examined end points, the consideration of both variables simultaneously never increased PA beyond 1% point. Specifically, in models predicting the pathologic stage, the gain related to simultaneous inclusion of both variables never improved the PA. PA actually decreased in the SVI model when both variables were considered. This implies that only information from the percentage of positive cores should be used in models predicting pathologic stage.

In models predicting BCR, the simultaneous consideration of the number and of the percentage of positive cores resulted in PA gains of 0.2% and 0.3%. Such gain is negligible from a practical perspective. This implies that, in models addressing BCR, only one way of coding should be used, and our data indicate that the number of cores provides the optimal PA gain.

Taken together, these data suggest that models predicting pathologic stage and BCR could be improved when information from biopsy cores is considered. Pathologic stage predictions seem to be better when the percentage of cores is added to standard predictors. Conversely, BCR predictions seem to be better when the number of positive cores is added. Inclusion of both variables does not appear to be recommended.

It is important to notice that some of the reported gains are small and may not be clinically relevant. For example, a gain of 0.4%, recorded with ECE predictions, was complemented with the number of positive cores. This finding implies that 4 of 1000 additional patients would be correctly classified if the number of positive cores were considered in ECE models. Such gain might be of importance in clinical trials, but might also be of no importance in daily clinical practice.

Several studies addressed the effect of either the number or the percentage of positive cores [6], [7], [10], [11], [16], [17], [18], and [19]. However, far fewer relied on the analytic approach that we used [10], and [15]. Egawa and colleagues [10] addressed the prediction of pathologic stage in 96 Japanese patients, and found that the number and percentage of positive cores improve PA of ECE and SVI. Increments were rather impressive and ranged from 9.3% to 12.9%. OC and LNI were not examined. Graefen and colleagues [15] examined the effect of core-derived information on the ability to predict BCR in 1152 patients, and found that the number and the percentage of cores improve PA by 1.0–1.4%, which is consistent with our findings.

Our findings have some limitations. First, our analyses did not include more detailed biopsy core information, such as the length of cancer or the percentage of cancer length relative to cumulative core length. These data may provide even more accurate predictions and may increase the PA beyond what we reported [23]. Second, the total number of cores that were obtained ranged from 2 to 36, and the median was between 6 and 7. This finding indicates that our cohort was exposed to a range of biopsy schemes (Table 1) that extended from sampling of digitally or ultrasonically suspicious areas to truly extended or even saturation biopsies. Our sample reflects an averaged yield of these different approaches and is not fully characteristic of either the sextant or the extended biopsy schemes. It is possible that the information from core material may contribute more to PA if it exclusively originates from individuals subjected to extended biopsies. Third, the included population was treated between 1992 and 2005. Therefore, a part of our cohort is very contemporary. Conversely, some patients may not reflect the effect of stage migration that swept over Europe and North America. Fourth, the definition of BCR (0.1 and rising) is stricter than that of some centers where a cut-off of 0.4 was used. Fifth, there was no central pathology, different surgical techniques were used, and not all patients were followed in the same fashion. Despite these differences, we were able to demonstrate a benefit from inclusion of core-derived information. The heterogeneity of our population adds to the strength of our findings, because variability may undermine the strength of tenuous relationships. This was not the case in our study, which shows that the effect of positive cores is robust and significant.

5. Conclusions

The information derived from biopsy cores is important and can improve the ability to predict pathologic stage and BCR. It appears that the percentage of cores is most helpful in stage predictions. Conversely, the number of cores appears to improve mostly BCR predictions. Consideration of both variables might not be helpful because of the similarity of information that they encode.

Conflicts of interest

There are no conflicts of interest.