Nomogram for predicting survival in patients with unresectable and/or metastatic urothelial cancer who are treated with cisplatin-based chemotherapy

Authors


Abstract

BACKGROUND

The current study was conducted to develop a pretreatment prognostic model for patients with unresectable and/or metastatic urothelial cancer who were treated with first-line, cisplatin-based chemotherapy.

METHODS

Individual data were pooled from 399 patients who were enrolled on 8 phase 2 and 3 trials evaluating cisplatin-based, first-line chemotherapy in patients with metastatic urothelial carcinoma. Variables selected for inclusion in the model were combined in a Cox proportional hazards model to produce a points-based nomogram with which to predict the median, 1-year, 2-year, and 5-year survival. The nomogram was validated externally using data from a randomized trial of the combination of methotrexate, vinblastine, doxorubicin plus cisplatin versus docetaxel plus cisplatin.

RESULTS

The median survival of the development cohort was 13.8 months (95% confidence interval, 12.1 months-16.0 months); 68.2% of the patients had died at the time of last follow-up. On multivariable analysis, the number of visceral metastatic sites, Eastern Cooperative Oncology Group performance status, and leukocyte count were each found to be associated with overall survival (P < .05), whereas the site of the primary tumor and the presence of lymph node metastases were not. All 5 variables were included in the nomogram. When subjected to internal validation, the nomogram achieved a bootstrap-corrected concordance index of 0.626. When applied to the external validation cohort, the nomogram achieved a concordance index of 0.634. Calibration plots suggested that the nomogram was well calibrated for all predictions.

CONCLUSIONS

Based on routinely measured pretreatment variables, a nomogram was constructed that predicts survival in patients with unresectable and/or metastatic urothelial cancer who are treated with cisplatin-based chemotherapy. This model may be useful in patient counseling and clinical trial design. Cancer 2013;119:3012—3019. © 2013 American Cancer Society.

INTRODUCTION

Contemporary cisplatin-based combination chemotherapy regimens achieve objective responses in approximately 50&percnt to 60% of patients with metastatic urothelial cancer.[1, 2] Although response durations are generally short, and the median overall survival of patients with unresectable and/or metastatic urothelial cancer is only approximately 14 months, significant heterogeneity exists with regard to patient outcomes.[2] Long-term follow-up of clinical trials has demonstrated that approximately 10% to 20% of patients are alive at 5 years.[1, 3] The ability to predict individual patient outcomes has clear implications with regard to both patient counseling and patient stratification on clinical trials.

Prognostic models have previously been developed for chemotherapy-naive patients with unresectable and/or metastatic urothelial cancer. In 1998, Bajorin et al identified Karnofsky performance status and the presence of visceral metastases as independent predictors of survival in a cohort of patients treated with methotrexate, vinblastine, and doxorubicin plus cisplatin (MVAC).[4] Similar prognostic factors have been identified in cohorts of patients treated with other cisplatin-based regimens.[5-7] However, there are some potential limitations to existing prognostic models in this clinical disease state: 1) their performance (discrimination and calibration) has generally not been rigorously assessed and/or they have not been validated externally; 2) they have not been expanded on; and 3) they generally assign equal weight to each prognostic variable, and “lump” patients into risk categories, which may not be optimal for individualized predictions.

Points-based nomograms for individualized risk prediction have been developed across a variety of clinical disease states in patients with genitourinary cancers and other solid tumors. These nomograms differ from traditional prognostic models in several ways.[8] Variable selection is generally performed before model development, and is based on parameters that are believed to be clinically relevant, instead of relying on variables that retain significance on multivariable analysis. The latter approach may be suboptimal for maximizing predictive accuracy because this leads to predictor variable coefficients that are biased in high absolute value and confidence intervals (CIs) that are falsely narrow. In addition, rather than potentially forcing heterogeneity by placing patients into defined risk groups, nomograms allow for differential weights to be assigned to each variable, facilitating individualized risk estimates. Herein, we describe the development and internal and external validation of a point-based nomogram for predicting survival in patients with unresectable and/or metastatic urothelial cancer who were treated with first-line, cisplatin-based combination chemotherapy.

MATERIALS AND METHODS

Inclusion Criteria

The development cohort included 399 patients with unresectable and/or metastatic urothelial carcinoma who were enrolled on 7 phase 2 and phase 3 trials exploring first-line cisplatin-based combination chemotherapy from 1998 through 2011. The validation cohort included patients enrolled on a phase 3 study comparing MVAC with docetaxel plus cisplatin in patients with unresectable and/or metastatic urothelial carcinoma (patients enrolled from 1997-2002). The details of each trial are provided in Table 1 and have been presented or published previously.[9-16] Each study was approved by the Institutional Review Board at the participating institutions and informed consent was obtained before treatment.

Table 1. Phase 2 and Phase 3 Trials Included in Current Analysis
TrialPhaseRegimenCycle Length, DaysEligibilityPrior Perioperative Chemotherapy Permitted?aRenal FunctionPerformance StatusbToxicity AssessmentResponse AssessmentRestaging Assessments
  1. Abbreviations: cT4b, unresectable disease; 5-FU, 5-fluorouracil; dd, dose-dense; ECOG, Eastern Cooperative Oncology Group; M+, metastatic; MVAC, methotrexate, vinblastine, doxorubicin, and cisplatin; NCI CTC, National Cancer Institute Common Toxicity Criteria; RECIST, Response Evaluation in Solid Tumors; WHO, World Health Organization.

  2. a

    Trials permitting prior perioperative chemotherapy required chemotherapy to have been completed >1 y prior to enrollment.

  3. b

    Trials using Karnofsky performance status for eligibility were converted to ECOG performance status as described in Ma C, Bandukwala S, Burman D, et al. Interconversion of three measures of performance status: an empirical analysis. Eur J Cancer. 2010;46:3175-3183.17

Hahn 2011[9]2Gemcitabine, cisplatin, and bevacizumab21cT4b and/or M+YesCreatinine <1.5ECOG 0-1NCI CTC (version 3)RECIST (version 1)Every 6 wk
Galsky [32][10]2Gemcitabine, cisplatin, and sunitinib21cT4b and/or M+YesCreatinine clearance ≥60ECOG 0-2NCI CTC (version 3)RECIST (version 1)Every 6 wk
Ecke 2006[11]2Gemcitabine, cisplatin, and paclitaxel21cT4b and/or M+NoCreatinine clearance ≥60ECOG 0-2WHOWHOEvery 6 wk
Krege 2010[12]2Gemcitabine and cisplatin, ± sorafenib21cT4b and/or M+NoCreatinine clearance ≥60ECOG 0-1NCI CTC (version 3)RECIST (version 1)Every 6 wk
Lin 2006[13]2Cisplatin and 5-FU28cT4b and/or M+NoCreatinine <1.3 or Creatinine clearance >35ECOG 0-2WHOWHOEvery 12 wk
Lin 2007[14]2Cisplatin, 5-FU, and paclitaxel21cT4b and/or M+NoCreatinine <1.3 or Creatinine clearance >40ECOG 0-2NCI CTC (version 3)WHOEvery 6 wk
Bamias 2004[15]IIIMVAC vs docetaxel and cisplatin28cT4b and/or M+YesCreatinine clearance ≥50ECOG 0-2NCI CTCWHOEvery 12 wk
Bamias 2011[16]3ddMVAC vs ddgemcitabine and cisplatin14cT4b and/or M+YesCreatinine clearance ≥50ECOG 0-1NCI CTC (version 3)RECIST (version 1)Every 8 wk

Each protocol required a histologic or cytologic diagnosis of urothelial carcinoma. Patients were required to have unresectable (cT4b) and/or metastatic disease. The pretreatment evaluations were similar among the protocols and included a complete history and physical examination and laboratory testing including a complete blood cell and platelet count, renal function, and hepatic function. Imaging studies were performed at baseline and every 6 to 12 weeks during treatment, depending on the study. Response assessments were performed using either World Health Organization criteria[17] or the Response Evaluation Criteria in Solid Tumors[17] and toxicity assessments were performed using either World Health Organization criteria or the National Cancer Institute Common Toxicity Criteria.

Nomogram Development and Construction

A list of variables that are routinely measured before the initiation of therapy was generated for nomogram development based on published reports and the authors' experience. The variables analyzed included site of the primary tumor (bladder vs other), number of distinct visceral metastatic sites, presence of lymph node metastases, Eastern Cooperative Oncology Group performance status (ECOG PS), and leukocyte count. Lung, liver, and bone metastases were considered visceral metastases, and the number of visceral metastatic sites was based on the number of these organs involved (0-3). As noted in the introduction, for nomogram development, covariates that are anticipated to have an impact on outcomes were defined for inclusion a priori, rather than relying only on variables that retain independent statistical significance on multivariable analysis.

Nomogram construction was performed as previously described.[8] Briefly, this method consists of specifying the predictor variables of interest and the form of the model to be used. In the current analysis, a multivariable Cox proportional hazards model was used to estimate the probability of overall survival at 1 year, 2 years, and 5 years for the site of the primary tumor, number of visceral metastatic sites, presence of lymph node metastases, ECOG PS, and leukocyte count. Next, the proportionality assumption of the Cox model was assessed by including time-varying covariates (an interaction between each predictor and the event time) in the Cox proportional hazards model. Because none of the interaction terms was statistically significant, the proportionality assumption was deemed valid. Finally, the nomogram itself was constructed based on the Cox model parameter estimates. This process begins by assigning 100 points to the prognostic variable in the model with the largest parameter estimate (in absolute value). Then the variable with the next largest parameter (absolute value) is identified; the number of points for this variable is calculated as the percentage obtained when dividing this second variable's parameter estimate by the initial variable's parameter estimate. This process continues for all variables in the model, with the points decreasing for each subsequent variable. The predicted probabilities based on the Cox model can be calculated for each combination of covariate values; similarly, the total points can be calculated for each combination of covariate values. Finally, the nomogram plots the range of predicted probabilities aligned so that they correspond to the observed number of total points for each observed combination of covariate values in the data set.

Nomogram Validation

Nomogram validation included 3 activities: 1) internal bootstrap validation, in which the Harrell concordance index[18] was estimated by bootstrapping with 200 resamples to calculate an unbiased measure of the ability of the nomogram to discriminate among patients; 2) calibration, which was constructed to examine how well the model-based predicted probabilities of survival agreed with the observed probabilities (in this calculation, 200 bootstrap resamples were used to generate the 95% CIs for the plot[8]); and 3) external validation, for which a Harrell c-index was calculated to assess the external validity of the nomogram by comparing the nomogram-predicted probabilities based on the development cohort with observed probabilities in the validation cohort.

The Design, Hmisc, and survival libraries in R Version 2.12.0 (R Project for Statistical Computing, Vienna, Austria) were used to estimate the Cox model, as well as to produce and validate the final nomogram.

RESULTS

Baseline Characteristics

Among the 399 patients in the development set, 15 were excluded based on missing dates for the start of treatment or follow-up, leaving 384 patients for the final analysis. No patients in the validation cohort were excluded (n = 186). The baseline patient characteristics are detailed in Table 2. The median survival for the development cohort was 13.8 months (95% CI, 12.1 months-16.0 months); 68.2% of patients had died at the time of last follow-up. The median survival for the validation cohort was 10.6 months (95% CI, 9.4 months-12.8months); at the time of last follow-up, 94.1% of patients had died.

Table 2. Patient Characteristics
CharacteristicsDevelopment Set (n = 384)Validation Set (n = 186)
  1. Abbreviations: 5-FU, 5-fluorouracil; ECOG, Eastern Cooperative Oncology Group; GC, gemcitabine plus cisplatin; MVAC, methotrexate, vinblastine, doxorubicin, and cisplatin.

Median age (range), y66 (34-90)65 (32-77)
ECOG performance status  
0147 (38.6%)85 (45.7%)
1192 (50.4%)63 (33.9%)
242 (11.0%)38 (20.4%)
Site of primary tumor  
Bladder295 (77.0%)153 (82.3%)
Other88 (23.0%)33 (17.7%)
Prior perioperative chemotherapy20 (5.2%)20 (10.8%)
Median hemoglobin (range), g/dL12.3 (6.8-17.0)12.5 (7.4-16.2)
Leukocyte count above normal range105 (27.7%)58 (32.4%)
Median creatinine (range)1.1 (0.4-2.0)1.1 (0.5-1.9)
No. of visceral metastatic sites  
0196 (51.3%)97 (52.2%)
1136 (35.6%)69 (37.1%)
242 (11.0%)16 (8.6%)
38 (2.1%)4 (2.1%)
Lymph node only metastases161 (42.0%)77 (42.0%)
Treatment regimen  
GC and sunitinib33 (8.6%)
GC and bevacizumab43 (11.2 %)
GC and sorafenib41 (10.7%)
GC48 (12.5%)
GC and paclitaxel15 (3.9%)
Cisplatin and 5-FU35 (9.1%)
Cisplatin, 5-FU, and paclitaxel44 (11.5%)
Dose-dense GC42 (10.9%)
Dose-dense MVAC83 (21.6%)
MVAC93 (50.0%)
Docetaxel plus cisplatin93 (50.0%)

Cox Proportional Hazards Model

The results of the Cox proportional hazards model for the development cohort demonstrated that the number of visceral metastatic sites, ECOG PS, and leukocyte counts above normal limits each were associated with overall survival (P < .05), whereas the site of the primary tumor and the presence of lymph node metastases were not found to demonstrate such an association (Table 3).

Table 3. Cox Proportional Hazards Model
VariablesHRP95% CIOverall P
  1. Abbreviations: 95% CI, 95% confidence interval; ECOG, Eastern Cooperative Oncology Group; HR, hazards ratio.

Site of primary tumor
Bladder vs other0.87.330.65-1.16.33
No. of visceral metastatic sites
1 vs 01.37.02931.03-1.83.0005
2 vs 02.28<.00011.52-3.42 
3 vs 02.49.03671.06-5.87 
Lymph node metastases
Yes vs no1.18.310.85-1.63.31
ECOG performance status
1 vs 01.34.04091.01-1.79.0004
2 vs 02.35<.00011.54-3.60 
Leukocyte count above normal limits
Yes vs no1.78<.00011.35-2.35<.0001

Prognostic Nomogram

The final nomogram including all 5 variables is illustrated in Figure 1. With this nomogram, an individual patient's value for each variable is scored by locating the corresponding position on the variable scale and drawing a vertical line to determine the corresponding points. The total points are then tallied, and a vertical line drawn through the survival scales provides estimates for 1-year, 2-year, 5-year, and median survival. For example, a patient with a bladder primary tumor with lymph node metastases, an ECOG PS of 0, no visceral metastases, and a normal leukocyte count has a 1-year survival probability of approximately 74%, a 5-year survival probability of approximately 37%, and a median survival of approximately 26 months. Conversely, a patient with an upper tract primary tumor, 3 distinct sites of visceral metastases, elevated leukocyte counts, and an ECOG PS of 2 has a 1-year survival probability of < 5% and a median survival of approximately 5 months.

Figure 1.

A nomogram for predicting the survival of patients with metastatic urothelial carcinoma who are treated with first-line, cisplatin-based chemotherapy is shown. Note that the number of sites of visceral metastases refers to the number of the following organs involved: lung, liver, and bone. Instructions for physicians are to locate the patient's leukocyte (“white blood cell”) count value on the leukocyte count axis. Draw a straight line up to the axis labeled “points” to determine the corresponding points. Repeat this process for each of the remaining axes, drawing a straight line each time to the points axis. Sum the points received for each predictive variable and locate this number on the total points axis. Draw a straight line down from the total points to the 1-year, 2-year, 5-year, and median survival axes to determine the patient's predicted survival probabilities. ECOG PS indicates Eastern Cooperative Oncology Group performance status.

Internal Validation

When subjected to internal validation, the nomogram achieved an uncorrected concordance index of 0.639 and a bootstrap-corrected concordance index of 0.626. Internal calibration showed that the predictions reasonably approximated actual survival probabilities (Fig. 2).

Figure 2.

Calibration curve of the final nomogram is shown. The calibration plot is shown, in which nomogram predictions are compared with observed fractions surviving at (A) 1 year and (B) 5 years. The diagonal line represents the performance of an ideal nomogram. The line containing error bars (95% confidence interval) represents the performance of the nomogram applied to the observed fractions surviving.

Impact of Treatment Regimen

Because 2 trials including 5-fluorouracil–based regimens were included in the development set, and such regimens are not as commonly administered in general practice, model discrimination was also assessed after excluding these trials. The remaining data set was comprised of 5 trials exploring the use of gemcitabine plus cisplatin-based or MVAC-based regimens. Notably, the treatment regimen used was not found to be a significant predictor of survival on multivariable analysis (P = .15). Internal validation of the resulting data set from the 5 trials achieved a bootstrap-corrected c-index of 0.631, which is very similar to the discrimination achieved when model development included the 5-fluorouracil–based trials.

External Validation

External validation was accomplished by comparing the nomogram predictions for each patient in the validation data set with the actual outcome. In this analysis, the nomogram had an estimated concordance index of 0.634.

DISCUSSION

Patients with unresectable and/or metastatic urothelial cancer who were treated with cisplatin-based chemotherapy appear to experience heterogeneous clinical outcomes. The ability to reliably predict outcomes before the initiation of treatment may enhance patient-physician communication, shared decision-making, and clinical trial design. In an attempt to address these needs, we have developed a points-based nomogram, with data derived from an international cohort, to predict survival in patients with urothelial cancer treated with first-line cisplatin-based chemotherapy using pretreatment variables measured on a routine basis.

Nomogram development began with the selection of variables and included both variables previously associated with outcomes in this clinical disease state and other variables believed to be potentially relevant based on our experience and that of others. The final model consisted of number of visceral metastatic sites, ECOG PS, presence of lymph node metastases, site of the primary tumor, and leukocyte count. It is important to note that for all the variables, the direction of the predictive effect was consistent with what one would expect clinically. Performance status and sites of metastases have been identified as prognostic factors in several prior studies.[4, 5, 7] Patients with metastatic urothelial cancers originating in the upper urinary tract have been previously demonstrated to have inferior outcomes and such tumors may have a different biology than cancers originating in the bladder, despite sharing the same histology.[19-22] Although the site of the primary tumor did not reach independent prognostic significance on multivariable analysis, this may have been due to an insufficient sample size and the primary tumor site was still included in the final model as per standard nomogram development methodology. We previously identified leukocytosis as a significant predictor of survival in patients with metastatic urothelial cancer in a preliminary analysis.[23] Notably, tumor-associated leukocytosis, generally neutrophilia, has been associated with poor outcomes in patients with other solid tumors, and has been associated with elevated serum levels of cytokines such as granulocyte–colony-stimulating factor, granulocyte-macrophage–colony-stimulating factor, and interleukin-6.[24-26] However, to our knowledge, this has not been previously demonstrated and validated in patients with advanced urothelial cancer. The etiology of tumor-associated leukocytosis in patients with urothelial cancer, and its contribution to poor outcomes, warrants further evaluation.

The concordance index for the prognostic model was 0.639 for the development cohort and 0.634 for the validation cohort, suggesting moderate discrimination. This nomogram should be considered a z “starting point.” However, this finding should also be viewed within the context of the limited available validated options for individualized risk estimates in patients with metastatic urothelial cancer. Furthermore, the discriminatory power of the current nomogram is within the range of other tools commonly used for routine counseling and decision-making in clinical oncology.[27, 28] Refinement of the nomogram, with the identification of additional clinical, pathologic, laboratory, and molecular predictors may, nonetheless, optimize the value of this tool.

There are characteristics of the development and validation cohorts that both potentially increase and decrease the generalizability of this tool. Data were derived from patients from multiple different countries, enrolled on a series of phase 2 and phase 3 clinical trials, and treated with a variety of different cisplatin-based combination chemotherapy regimens. However, patients also met rigorous clinical trial eligibility criteria and the majority of trials required an ECOG PS of ≤1 or 2. Although patients with an ECOG PS of >2 are less frequently treated with cisplatin-based therapy, even outside of the context of clinical trials, the cohort used to generate the model may not be representative of the general population of patients with metastatic urothelial cancer with respect to comorbidities. It is important to note that the eligibility criteria used by the included trials were similar, but not identical.

There are other potential limitations to the current study. Two 5-fluorouracil–based regimens were included in the development cohort, which are not commonly used in general practice. However, excluding these 2 trials appeared to have no meaningful impact on the model's predictive ability, and the treatment regimen used was not found to be significantly predictive of survival on multivariable analysis. The validation cohort included 2 regimens, one of which resulted in inferior survival in the original randomized trial.[15] Nevertheless, the nomogram performed similarly in the validation cohort compared with the development cohort. The degree to which the differences in treatment regimens may have contributed to the moderate discrimination achieved with the current nomogram is not entirely clear. However, practice pattern data suggest that patients with metastatic urothelial cancer receive a variety of different first-line, cisplatin-based regimens,[29] and the goal of the current study was to include a diverse range of treatments in an effort to increase the generalizability of the tool. Many of the trials excluded patients who previously were treated with perioperative chemotherapy, whereas others only included patients if such therapy was completed > 1 year before study enrollment. Whether patients who developed disease recurrence after having received perioperative chemotherapy should still be considered “first-line” remains controversial because such patients are rarely retreated with cisplatin-based regimens in standard clinical practice and are often included in “second-line” clinical trials. Furthermore, the small percentage of patients previously treated with perioperative chemotherapy included in the current study is in keeping with estimates of perioperative chemotherapy in population-based studies.[30] The validation set in the current study was not selected at random, but was selected based on the finding that it was the largest single available trial for which we had data access that approximated the size of the development cohort. Finally, the nomogram was developed from a cohort of patients who were treated with cisplatin-based chemotherapy. Therefore, the intended use of the tool is for counseling and clinical trial stratification for patients receiving first-line, cisplatin-based chemotherapy and may not apply to the population of cisplatin-ineligible patients or those patients not fit to receive chemotherapy at all.[31]

The current points-based prognostic nomogram, which was validated both internally and externally, may be used to predict outcomes in patients with metastatic urothelial cancer who are treated with cisplatin-based chemotherapy. The identification of novel prognostic variables may further enhance the performance of the nomogram. In addition, studies to define the role of such tools in patient counseling are needed.

FUNDING SUPPORT

No specific funding was disclosed.

CONFLICT OF INTEREST DISCLOSURES

The authors made no disclosures.

Ancillary