A nomogram predicting metastatic progression after radical prostatectomy

Authors


Pierre I Karakiewicz md frcsc, Cancer Prognostics and Health Outcomes Unit, University of Montreal Health Center (CHUM), 1058, rue St-Denis, Montreal, Quebec, Canada H2X 3J4. Email: pierre.karakiewicz@umontreal.ca

Abstract

Objectives:  To develop and internally validate a nomogram predicting the individual probability of metastatic progression after radical prostatectomy according to the length of disease-free interval.

Methods:  Cox regression modeled the probability of metastatic progression of prostate cancer in 752 patients treated with radical prostatectomy with a mean follow up of 11.6 years (median 11.4; range 0.1–40.5). The significance of the predictors was confirmed in competing risks analysis, which accounts for other causes of mortality. The Cox regression model-based nomogram was internally validated with 200 bootstrap resamples.

Results:  Eighty-five of 752 patients (11.3%) developed metastatic progression. The 5, 10, 15 and 20-year actuarial rates of metastatic progression-free survival were, respectively, 95.9, 90.5, 84.8 and 80.5%. Pathological stage T3, elevated radical prostatectomy Gleason sum and delivery of adjuvant radiotherapy represented independent predictors of metastatic progression in both Cox and competing risks regression models, and constituted the nomogram predictors along with a fourth variable describing the presence of co-morbidities. After 200 bootstrap resamples the nomogram achieved 80.2, 77.7, 77.6 and 76.0% accuracy in predicting metastatic progression at 5, 10, 15 and 20 years after radical prostatectomy.

Conclusions:  Metastatic progression is a sign of poor prognosis in men with prostate cancer. Our nomogram is able to accurately predict the conditional probability of metastatic progression up to 20 years after radical prostatectomy.

Introduction

Several prognostic models can help identify men at risk of biochemical recurrence1,2 however there are no models that can predict the individual probability of either metastatic progression or of prostate cancer (PCa)-specific mortality at the time of radical prostatectomy. Therefore, we decided to use the nomogram approach to quantify the individual risk of metastatic progression.3–5 We relied on a surgical series of radical prostatectomy patients treated between 1954 and 1994 with available long-term follow up.6 Since the risk of metastatic progression decreases with time, we adjusted the nomogram predictions according to the disease-free interval after radical prostatectomy.2,7

Methods

Patient population

Between December 1954 and May 1994, 787 patients underwent radical prostatectomy at Virginia Mason Medical Center (VMMC) by 13 surgeons. No patient received neo-adjuvant therapy. Clinical and pathological data were logged into a prospective database from 1969 to 1999. In later years the records were maintained electronically with institutional review board approval. Preoperative comorbidity was defined according to the Charlson comorbidity index.8 For the purpose of analyses, 24 patients (3.2%) with an index of 2 or more were grouped with those with an index of 1. Of 787 men, 35 were excluded due to missing age (9) or missing pathology (26). This resulted in 752 evaluable patients. Last data follow up was performed on 1 December 2004.

Radical perineal prostatectomy was performed in 428 (56.9%) patients, while 324 (43.1%) underwent retropubic radical prostatectomy. Pelvic lymph node dissection (PLND) at radical prostatectomy was performed in 342 patients (45.5%), who were selected according to individual surgeon preference and surgical approach. PLND pathologies were not recorded in the database and could not be included in the present analyses.

From 1992 to 1994 tumors were routinely classified according to the Gleason grading system.9 Before 1992 tumor grade was recorded as well differentiated (I), moderately differentiated (II) and poorly differentiated (III). To recode these data we followed the paradigm outlined by Roehl et al.,10 in which well differentiated tumors are classified as Gleason sum 3, moderately differentiated tumors are assigned Gleason sum 6 and poorly differentiated tumors are assigned Gleason sum 9. Positive surgical margins were recorded as the presence of cancer cells against the inked resection margin.

Adjuvant or salvage radiotherapy and/or hormonal therapy were delivered according to the individual surgeon's preference. Adjuvant radiotherapy was delivered within 1–3 months after radical prostatectomy. Serum prostate-specific antigen (PSA) testing was initiated at VMMC in 1988. Since that time patients were followed at least quarterly for 2 years, then at least biannually for 2 years and then at least annually. Metastatic progression was diagnosed based on radiographic and/or scintigraphic studies.6 Cause of death was ascertained according to detailed chart review or was obtained from the VMMC cancer registry. The cancer registry uses links with the Washington State Death Certificate Office.

Statistical analysis

The cohort of 752 men was used to develop and internally validate the nomogram for prediction of metastatic progression. Pathological stage, surgical margin status, radical prostatectomy Gleason sum, presence of comorbidities according to the Charlson comorbidity index and the delivery of adjuvant radiotherapy were used as predictors in univariable and multivariable Cox regression models addressing metastatic progression after radical prostatectomy. Since a proportion of patients in our cohort was treated before the PSA era (51.7%), pre-operative serum PSA values were unavailable in a large proportion of these individuals. For that purpose we decided to omit pre-operative PSA from consideration within the nomogram. Adjuvant radiotherapy represented the only variable not known at the time of radical prostatectomy. Since it was invariably delivered within 3 months after radical prostatectomy, it was not modeled as a time-dependent variable. Since adjuvant radiotherapy represents an early event and since the decision to administer adjuvant radiotherapy is made very early after radical prostatectomy, we decided to include adjuvant radiotherapy within the predictors of metastatic progression. Conversely, administration of hormonal therapy or salvage radiotherapy were not considered in the model.

The predictive accuracy and statistical significance of all predictors were quantified in univariable analyses. Subsequently, all predictors were included in a multivariable model. Stepwise backward variable removal was applied to the model with the intent of identifying the most accurate and the most parsimonious set of predictors.

Given that a proportion at risk of metastatic progression die as a result of other causes before developing metastatic progression, competing risks regression was used to test the significance of the nomogram's predictors, after accounting for other-cause mortality. The effect of other-cause mortality cannot be accounted for in Cox regression models. However, the use of competing risks regression, as described by Fine and Gray,11 can remove this limitation. Unfortunately, there are no commercially available statistical packages that would allow applying competing risks regression within a nomogram setting. Consequently, the nomogram was based on Cox regression models. As the risk of metastatic progression decreases with survival without metastatic disease after radical prostatectomy, we used the conditional survival approach to provide nomogram predictions on the basis of survival without metastatic disease.2

The accuracy of the final multivariable model was quantified using 200 bootstrap resamples that were applied to its regression coefficients. Because we relied on censored data, Harrell's concordance index substituted the area under the receiver operating characteristics curve.12 Specific accuracy estimates were calculated for 5, 10, 15 and 20 years. To test the nomogram performance characteristics we used the val.surv calibration method devised for censored data. All analyses were performed using the S-Plus Professional, version 1 (MathSoft Inc., Seattle, Washington) and statistical significance was set at 0.05. All studies were undertaken with the approval and oversight of the Institutional Review Boards for the Protection of Human Subjects.

Results

The pathological and treatment characteristics of 752 assessable patients are shown in Table 1. The mean age at radical prostatectomy was 63.6 years (median 64.4, range 42.9 to 77.9). At pathological evaluation, 341 patients (45.3%) had pT3 disease, 63 patients (8.4%) had Gleason sum ≥8, and 283 patients (37.6%) had positive surgical margin status. Mean follow up from radical prostatectomy to metastatic progression or last follow up was 11.6 years (median 11.4, range 0.1–40.5). Metastatic progression was recorded in 85 patients (11.3%). The median actuarial time to metastatic progression was not reached (mean: 34.1).

Table 1.  Descriptive characteristics for 752 patients treated with radical prostatectomy between 1954 and 1994
VariablesNumber of patients
Total752 (100.0%)
Study period
 1954–1987389 (51.7%)
 1988–1994363 (48.3%)
Age
 Mean (median)63.6 (64.4%)
 Range42.9–77.9
pT stage (1992)
 T2a177 (23.5%)
 T2b/c234 (31.1%)
 T3341 (45.3%)
Positive surgical margin283 (37.6%)
Gleason sum
 2–5220 (29.2%)
 6299 (39.8%)
 7170 (22.6%)
 8–1063 (8.4%)
Surgical procedure
 Perineal428 (56.9%)
 Retropubic324 (43.1%)
Lymph node dissection performed342 (45.5%)
Comorbidities (Charlson index)
 0685 (91.1%)
 ≥167 (8.9%)
Adjuvant radiotherapy118 (15.7%)
Hormonal therapy before distant recurrence47 (6.3%)
Metastatic progression85 (11.3%)
Follow-up time (years)
 Mean (median)11.6 (11.4)
 Range0.1–40.5
Actuarial time to metastatic progression (years)
 Mean (median)34.1 (not reached)

Figure 1A shows the rate of metastatic progression survival at 5, 10, 15 and 20 years, which was respectively 4.1, 9.5, 15.2 and 19.5%. Figure 1B shows the rate of other cause mortality at 5, 10, 15 and 20 years, which was 5.4, 15.5, 31.6 and 48.9%. Figure 1 also demonstrates the rate of metastatic progression stratified according to the tested variables (Fig. 1C–G), where pathological T3 stage, positive surgical margins status (P < 0.001), radical prostatectomy Gleason sum >5 (P < 0.001), and delivery of adjuvant radiotherapy (P < 0.001) were all associated with a higher risk of metastatic progression.

Figure 1.

Metastatic progression-free survival (a) and prostate cancer-unrelated mortality (b) at 5, 10, 15 and 20 years. Rate of metastatic progression stratified according to pathological T stages (c), surgical margins status (d), radical prostatectomy Gleason sum (e), comorbidity status (f) and delivery of adjuvant radiotherapy (g).

The univariable Cox regression analyses replicated the relationships described with the Kaplan-Meier survival curves, and are shown in Table 2. Table 2 also shows the multivariable Cox regression model predicting metastatic progression after radical prostatectomy. In the full multivariable model, only radical prostatectomy Gleason sum (P < 0.001) and the delivery of adjuvant radiotherapy (P = 0.02) represented independent predictors of metastatic progression. After stepwise backward variable elimination only radical prostatectomy Gleason sum (P = 0.001), pT stage (P = 0.02), adjuvant radiotherapy status (P = 0.008) and the presence of comorbidities (P = 0.07) were included in the final model. All variables included in the final model contributed to its predictive accuracy.

Table 2.  Univariable and multivariable analyses predicting metastatic progression after radical prostatectomy
PredictorsUnivariableMultivariable
RR (95% CI); P-valueFull model
RR(95% CI); P-value
Reduced model
RR(95% CI); P-value
  1. CI, confidence interval; RR, risk ratio.

Gleason sum–; < 0.001–; 0.001–; 0.001
 6 vs 2–54.2 (1.9–9.4); < 0.0013.5 (1.6–7.9); 0.0023.5 (1.6–7.8); 0.002
 7 vs 2–55.1 (2.1–12.0); < 0.0013.4 (1.4–8.1); 0.0063.4 (1.4–8.2); 0.006
 8–10 vs 2–59.7 (3.9–24.0); < 0.0017.0 (2.8–17.5); < 0.0017.0 (2.8–17.5); < 0.001
pT stage–; < 0.001–; 0.2–; 0.02
 T2b/c vs T2a1.0 (0.5–2.2); 1.01.1 (0.5–2.4); 0.81.1 (0.5–2.5); 0.7
 T3 vs T2a3.2 (1.7–6.0); < 0.0011.9 (0.9–4.0); 0.092.2 (1.1–4.4); 0.02
Adjuvant radiotherapy3.3 (2.1–5.2); < 0.0011.8 (1.0–3.0); 0.022.0 (1.2–3.2); 0.008
Comorbidities (≥1 vs 0)0.4 (0.1–1.3); 0.10.3 (0.1–1.0); 0.060.3 (0.1–1.0); 0.07
Positive surgical margin2.6 (1.7–4.1); < 0.0011.4 (0.7–2.4); 0.3 

Table 3 shows the univariable and multivariable competing risks regression models that were developed on the cohort of 752 patients. In the multivariable competing risks regression model, the statistically significant predictors of metastatic progression maintained their independent prognostic status, after accounting for other-cause mortality.

Table 3.  Univariable and multivariable competing risk-regression analyses predicting metastatic progression after radical prostatectomy
PredictorsUnivariableMultivariable
RR (95% CI); P-valueFull model
RR(95% CI); P-value
Reduced model
RR(95% CI); P-value
  1. CI, confidence interval; RR, risk ratio.

Gleason sum–; ––; ––; –
 6 vs 2–54.2 (1.9–9.3); < 0.0013.5 (1.6–7.9); 0.0023.5 (1.6–7.7); 0.002
 7 vs 2–54.9 (2.1-11.6); < 0.0013.4 (1.4–8.0); 0.0063.4 (1.4–8.1); 0.006
 8–10 vs 2–59.1 (3.6–22.9); < 0.0016.2 (2.4–15.7); < 0.0016.2 (2.4–15.7); < 0.001
pT stage–; ––; ––; –
 T2b/c vs T2a0.9 (0.4–2.0); 0.91.0 (0.5–2.2); 0.91.1 (0.5–2.3); 0.9
 T3 vs T2a3.0 (1.7–5.6); < 0.0011.8 (0.9–3.6); 0.12.1 (1.1–4.0); 0.03
Adjuvant radiotherapy3.4 (2.2–5.3); < 0.0011.9 (1.1–3.2); 0.012.0 (1.2–3.2); 0.006
Comorbidities (≥1 vs 0)0.4 (0.1–1.5); 0.090.3 (0.1–1.0); 0.040.3 (0.1–1.0); 0.05
Positive surgical margin2.6 (1.7–4.0); < 0.0011.3 (0.8–2.3); 0.3

The model subjected to backward variable elimination, which is shown in Table 3, served as the basis for the multivariate nomogram (Fig. 2A). After 200 bootstrap resamples of the nomogram regression coefficients, the combined accuracy of the metastatic progression predictions was 80.2, 77.7, 77.6 and 76.0% at respectively 5, 10, 15 and 20 years. Within the nomogram, radical prostatectomy Gleason sum 8–10 represented the most powerful predictor. The absence of comorbidities, pathological T3 stage and the delivery of adjuvant radiotherapy represented, in this order, increasingly less influential predictors.

Figure 2.

Nomogram (a), calibration plot (b) and conditional probability (c–f) for prediction of individual probability of metastatic progression after radical prostatectomy. The 45° line represents the ideal predictions. The X axis indicates the nomogram predicted metastatic progression probability and the Y axis indicates the observed metastatic progression proportion of patients. It is of note that the curve depicting the relationship between predicted and observed metastatic progression proportions closely approximates the ideal prediction (b).

The nomogram predictions are shown for 5, 10, 15 and 20-year time points (Fig. 2C–F). Each time point prediction is accompanied by a graph that allows adjustment of the metastatic progression-free probability according to the disease-free time interval after radical prostatectomy, which was defined as survival without metastases. The prediction of metastatic progression-free status requires two steps. First, the nomogram axes are used to quantify the total number of risk points. Subsequently the total points are applied on one of the time-specific scales (Fig. 2C–F) with the intent of adjusting for the disease-free interval. For example, Figure 2D shows the 10-year predictions and allows adjustment for the disease-free interval. For example, a patient with 245 risk points has a 50% probability of metastatic-progression free survival at 10 years after radical prostatectomy, if predictions are made immediately after radical prostatectomy. Conversely, after a 60-month disease-free interval, his probability of metastatic progression-free survival at 10 years after radical prostatectomy increases to 70%. Finally, if the disease-free interval persists up to 96 months after radical prostatectomy the probability of metastatic progression-free survival at 10 years increases to 93% (Fig. 2D).

Figure 2B shows the calibration plot of the newly developed nomogram predicting the individual probability of metastatic progression after radical prostatectomy. It is of note that the curve depicting the relationship between predicted and observed proportions closely approximates the ideal predictions, which indicates excellent calibration.

Discussion

Currently no models can predict the individual probability of metastatic progression immediately after radical prostatectomy. Conversely, models addressing metastatic progression do exist for patients with confirmed biochemical recurrence.13–15 To address this void, we decided to develop and internally validate a novel nomogram predicting the individual probability of metastatic progression up to 20 years after radical prostatectomy. Our model was 80, 78, 78 and 76% accurate in predicting metastatic progression at respectively 5, 10, 15 and 20 years after radical prostatectomy. These accuracy estimates compare very favorably with the Slovin et al. model (69% accurate).15 Unfortunately, the Pound et al. model14 was not validated. Our accuracy is somewhat lower than that reported by Dotan et al. (93%).13 This finding can be explained by the design of the Dotan et al. study, which was restricted to patients with an established biochemical recurrence. Therefore, the event of interest was much closer to the starting point than in our study. Since it is easier to predict near future than distant future, understandably the Dotan et al. model was more accurate.13 Moreover, the Dotan et al. study was restricted to patients with biochemical recurrence, in whom the risk of metastatic progression is substantially higher than in all patients treated with radical prostatectomy.13 It is generally easier to predict a more frequent event than a rare event, which further explains the higher accuracy of the Dotan et al.13 model. Based on these considerations, the 76–80% accuracy of our nomogram, which can predict up to 20 years after radical prostatectomy, can be considered as highly respectable.

The effect of radical prostatectomy Gleason sum is consistent with Pound et al., where the pathological Gleason sum within the nomogram occupies a central role. In both the Dotan et al.13 and the Slovin et al.15 models T stage also represent an important predictor. Inclusion of comorbidity is novel relative to the previous models. Patients with no comorbidity are at higher risk of metastatic progression, which can be explained by competing risks. The presence of comorbidities prevents the patients from living enough to develop metastases. Conversely, adjuvant radiotherapy does not protect from metastatic progression, unlike the suggestion made by the recent randomized trial, where adjuvant radiotherapy lowered the risk of both biochemical recurrence and local recurrence.16 The discrepancy of the effect of adjuvant radiotherapy may relate to the nature of our endpoint (metastatic progression) relative to the endpoint addressed by Bolla et al.16 (local control and biochemical recurrence) and to design type (randomized vs non-randomized).

Since a proportion of patients may die of other causes prior to developing metastatic progression, we confirmed the statistical significance of our nomogram predictors in competing risks regression models.11 In competing risks analyses, all variables included in the nomogram maintained their independent predictor status, after accounting for other cause mortality.

Moreover, we complemented the nomogram with conditional probability plots that adjust the predictions according to the disease-free interval,2 defined as survival without metastases. As can be seen in Figure 2C–F, a longer disease-free interval decreases the probability of metastatic progression. It is noteworthy that none of the existing tools for prediction of metastatic progression offer this form of adjustment, which was introduced by Stephenson et al.2

Our results failed to demonstrate a statistically significant difference in the univariable metastatic progression rates between pathological Gleason score 6 and 7. Several explanations may be proposed. The least speculative of those consists of the pathological stage, positive surgical margins status and adjuvant external beam radiotherapy (XRT) differences that existed between those categorizes. The consideration of these differences in the multivariable Cox regression model resulted in the confirmation of independent predictor status of all four pathological Gleason sum categories. Similarly, it might be argued that our series should not contain any pathological Gleason 2–5 cases. It might be argued that those ‘low Gleason scores’ represent examples of misclassification bias. Nonetheless, the univariable as well as multivariable analyses demonstrated that patients with pathological Gleason sums of 2–5 had statistically lower rates of distant progression than their counterparts with pathological Gleason sums of 6 or 7 or 8–10. These differences validate the grading of pathological Gleason sums in the current series with respect to their effect on distant progression rates.

Several limitations apply to our study. First, the sample size is small relative to nomograms for prediction of other endpoints. However, the sample size of 752 patients exceeds that of Pound et al.14 (n = 304), Dotan et al.13 (n = 239) and Slovin et al.15 (n = 148). Second, our model is based on a single institution series. Ideally, our data could originate from several institutions to enhance their generalizability. However, the other nomograms also include patients from a single institution.13,15 Third, our nomogram was not externally validated. Ideally, its accuracy would be confirmed in an independent external cohort. However, other nomograms are also limited by lack of external validation.13,15 Moreover, our previous work demonstrates that bootstrap resamples closely approximate independent external validation.17 Of internal validation methods, bootstrapping represents the preferred approach.18,19 Fourth, we did not have information about nodal metastases. Inclusion of these data could have increased the accuracy of our nomogram.20–24 Fifth, the Gleason sums used in our series have not been centrally reviewed to account for grade migration25. Again, this limitation also affects the competing models.13,15 Fifth, our predictions do not include serum PSA since 51.7% of our patients were treated in the pre-PSA era. It is plausible that the addiction of serum PSA among predictors might have added to the accuracy of the nomogram. Unfortunately, the inclusion of serum PSA among the predictors would have resulted in the exclusion of patients with the longest follow-up which are those with the highest likelihood of developing clinically detectable metastases.

Sixth, adjuvant and salvage therapies were not delivered according to a strict protocol.6 Adjuvant XRT was in general delivered to patients with positive surgical margins, pT3 PCa or to those with pathological Gleason 8–10 disease. Unfortunately, no strict protocol existed in that regard. In consequence, not all patients with those tumor characteristics did receive adjuvant XRT.

Of all 752 patients, 75 (10.0%) received salvage XRT. The indications consisted of palpable recurrence in patients treated in the pre-PSA era and of detectable and rising serum PSA in absence of distant metastases in patients treated in the PSA era. Finally, hormonal therapy was delivered in 114 patients (15.2%). Of these, 26.5% were treated at the time of biochemical recurrence vs 14.2% at local and 59.3% at distant recurrence. It is noteworthy, that PCa specific mortality was lowest in patients treated with early hormonal therapy (at biochemical or local recurrence) vs at the time of distant progression.26

Despite these limitations, our nomogram represents the only available model capable of predicting the individual risk of metastatic progression before biochemical recurrence.

In conclusion, metastatic progression after radical prostatectomy for PCa is a sign of poor prognosis. The nomogram can accurately predict the individual, conditional probability of metastatic progression up to 20 years after radical prostatectomy.

Acknowledgments

Pierre I Karakiewicz is partially supported by the University of Montreal Health Center Urology Associates, Fonds de la Recherche en Santé du Québec, the University of Montreal Department of Surgery and the University of Montreal Health Center (CHUM) Foundation.

Ancillary