Fax: (418) 691-5562
Carcinoma of the upper urinary tract†
Predictors of survival and competing causes of mortality
Article first published online: 11 MAY 2009
Copyright © 2009 American Cancer Society
Volume 115, Issue 13, pages 2853–2862, 1 July 2009
How to Cite
Inman, B. A., Tran, V.-T., Fradet, Y. and Lacombe, L. (2009), Carcinoma of the upper urinary tract. Cancer, 115: 2853–2862. doi: 10.1002/cncr.24339
Presented at the Canadian Urologic Association Annual Meeting, Ottawa, Ontario, Canada, June 26–29, 2005.
- Issue published online: 19 JUN 2009
- Article first published online: 11 MAY 2009
- Manuscript Accepted: 15 DEC 2008
- Manuscript Revised: 11 DEC 2008
- Manuscript Received: 30 SEP 2008
- urothelial carcinoma;
- ureteral cancer;
- renal pelvic cancer;
- prognostic factor;
- competing risks
Carcinomas of the upper urinary tract are uncommon tumors that usually occurred in elderly patients. Competing causes of mortality should be considered when treating these patients.
All patients with upper urinary tract tumors who were treated surgically at Centre Hospitalier Universitaire de Québec and affiliated hospitals from 1978 to 2001 were retrospectively reviewed. Clinical and pathologic variables were assessed from both the preoperative and postoperative periods of management, and clinical outcomes were tracked. Competing risks regression, Cox proportional hazards modeling, and multiple imputation were used to assess predictors of cancer-related and competing risks‒related mortality in both preoperative and postoperative settings.
Competing risks were responsible for 46% of deaths in this cohort of 168 patients. Preoperatively, the most important predictor of cancer-related mortality was a clinically invasive tumor (hazards ratio [HR], 3.97; P < .001), whereas increasing age (HR, 1.07; P < .001) was found to be the most important predictor of competing mortality. Postoperatively, tumor grade was the most important predictor of cancer-related mortality (HR, 3.92; P < .001) whereas constitutional symptoms (HR, 1.91; P = .015) and increasing age (HR, 1.06; P < .001) were found to be predictive of competing mortality.
In the current study, stage and grade were found to be the 2 most important independent predictors of survival in patients with tumors of the upper urinary tract and were highly correlated. Pain or weight loss was found to be a novel predictor of survival in this cancer. Although a survival disadvantage was not noted for women, nephron-sparing surgery, ureteral tumors, or older patients with respect to cancer, competing causes of mortality were found to be responsible for greater than one-third of observed deaths and age was the best predictor of this occurrence. Cancer 2009. © 2009 American Cancer Society.
Carcinomas of the upper urinary tract are relatively uncommon tumors, accounting for approximately 5% of renal and urothelial tumors and occurring at an annual incidence rate of 0.6 to 1.1 cases per 100,000 people at risk.1 Renal pelvic tumors are 2 to 3 times more common than ureteral tumors, with the distal ureter being involved much more frequently than the proximal ureter.2, 3 Importantly, the incidence-rate of upper urinary tract cancers appears to be increasing.1 Although the reasons for this increase have not been clearly elucidated, increased environmental exposure to urothelial carcinogens, improved diagnosis due to ureteroscopy,1 and improved survival of patients with urothelial carcinoma of the bladder have all been proposed as possible explanations.1, 4
Although several prognostic factors have been proposed for carcinomas of the upper urinary tract, some important general predictors of cancer survival have to our knowledge not yet been evaluated in this population. Furthermore, few of the identified risk factors have been tested in a large cohort using contemporary statistical methods. In the current study, we report what to our knowledge is the largest Canadian series published to date of upper urinary tract tumors and examine preoperative and postoperative factors that predict patient survival after the surgical treatment of these tumors. We also demonstrate the importance of competing causes of mortality in determining patient outcome.
MATERIALS AND METHODS
After institutional review board approval, we retrospectively reviewed the medical records of all patients who had undergone surgical treatment for carcinomas of the upper urinary tract at the Centre Hospitalier Universitaire de Québec and affiliated hospitals between January 1, 1978, and December 31, 2001. One hundred sixty-eight patients were identified who met these criteria, and pertinent clinical data were extracted from their medical records. In the case in which a patient had not been seen for a period of time exceeding 1 year, attempts to contact the patient, the patient's family, or the patient's referring physician were made. Date and cause of death was determined by review of death certificates, review of medical record, and a provincial death registry search.
Definitions and Predictors
Overall survival (OS) was defined as the period of time survived after surgery and cancer-specific survival (CSS) was defined as the period of time survived after surgery without evidence of local recurrence or metastases; patients who died of noncancer-related causes were censored. Preoperative predictors were those variables generally available to the physician before any attempt at therapy and included patient age, sex, the presence of constitutional tumor-related symptoms (weight loss and flank pain), tumor localization (renal pelvis or ureter), and clinical tumor stage (superficial or invasive). Postoperative predictors were those variables generally available after surgical resection and included the TNM 2002 pathologic T classification, pathologic lymph node status, pathologic tumor grade, tumor histology (urothelial or nonurothelial carcinoma), and the type of surgical procedure (nephroureterectomy or nephron-sparing surgery). All surgical procedures were performed before the introduction of urologic laparoscopy at our medical center, and therefore all were performed using the open technique. A standardized lymphadenectomy was not universally performed.
Missing Data Analysis
Summary statistics indicated that some variables had missing data. Patterns and mechanisms of missing data were investigated using hierarchical clustering, rank correlation, and graphic methods.5 Complete data were available for all preoperative predictors in 122 (73%) patients, for all postoperative predictors in 161 (96%) patients, and for both preoperative and postoperative predictors in 118 (70%) patients. A comparison of covariates and outcomes between patients with complete data and missing data is shown in Table 1. No significant differences in survival outcomes were observed between patients with complete and incomplete data; therefore, an assumption of missingness occurring at random appeared justified. Multiple imputations were performed using a bootstrap-based expectation maximization method implemented in the Amelia package available for R statistical software (version 2.6.1; R Project for Statistical Computing). Fifty-one imputed datasets were created: 1 for model development and 50 for model evaluation. Identical models were fit to each of the 50 evaluation datasets and their results were combined using the methods of Rubin and Schenker.6
|Variable||All Cases n=168||Incomplete Cases n=50||Complete Cases n=118||P*|
|Median age, y||70||IQR, 64-77||70||IQR, 66-78||70||IQR, 64-77||.673†|
|Pain or weight loss||64||38%||16||32%||48||41%|
|TNM 2002 stage||.023‡|
|Year of surgery||.372§|
|Median follow-up, y||5.88||5.67||6.46||.885§|
|Survival outcome at 5 y|
|Overall survival||54%||95%CI, 46-63||43%||95%CI, 30-63||58%||95%CI, 49-69||.203§|
|Cancer-specific survival||69%||95%CI, 61-78||60%||95%CI, 45-80||72%||95%CI, 63-82||.278§|
The modeling process started with an examination of univariate Kaplan-Meier curves of the candidate predictors using the log-rank test for comparisons and with univariate competing risks cumulative incidence curves using the Gray test for comparisons.7 If the P value for any comparison was <.25, the predictor was considered potentially important and was included in future multivariate models. Two approaches were used for modeling predictors of patient survival multivariately. First, multivariate Cox proportional hazards models were constructed for OS and CSS, censoring competing deaths as needed. Second, multivariate Fine-Gray competing risks regression models were fit defining noncancer deaths as competing causes of mortality. Covariate functional form, the proportional hazards assumption, and the presence of influence were assessed using methods described in Therneau and Grambsch, including smoothing splines, residual plots, and the zph test.8 Differences between the endpoints and groups were considered statistically significant if the 2-sided P value was <.05.
Characteristics of the Patient Cohort
Preoperative patient characteristics are shown in Table 1. The most common presenting sign was hematuria (73%) followed by flank pain (32%), weight loss (11%), and an abdominal mass (6%). Clinically superficial tumors were noninvasive at final pathology in 81% of cases, whereas clinically invasive tumors were found to be invasive in the pathology specimen in 90% of cases. Tumor stage appeared to be somewhat bimodal with pTa lesions and pT3 lesions occurring in 33% and 32% of cases, respectively. Most patients with positive lymph nodes had pT3 (65%) or pT4 (24%) disease (P < .001) and had grade 3 tumors (82%; P < .001). Similarly, 86% of grade 1 lesions were staged as pTa, and 67% of grade 3 tumors were staged as pT3 or pT4. One hundred percent of pT4 tumors and 61% of pT3 tumors had a histologic grade of 3 (P < .001).
Familial cancer histories were available for 148 patients and revealed that 4 patients had a family history of bladder cancer, whereas 1 patient had a family history of kidney cancer. Detailed personal cancer histories were available for 159 patients and indicated that the most common preoperative tumors were bladder cancer (23%), uterine cancer (6% of women), breast cancer (6% of women), colorectal cancer (4%) and prostate cancer (1% of men). Seven patients had previously undergone cystectomy for bladder cancer. Postoperative tumors were also very common, with bladder cancer found in 34.5% of patients, followed by uterine cancer (2% of women), prostate cancer (1% of men), and colorectal cancer (1%). Eight patients required a cystectomy for postoperative bladder cancer.
The vast majority of the tumors were urothelial carcinomas (92%), but squamous cell carcinomas (7%) and adenocarcinomas (1%) were also noted. Although the surgical technique varied, a large majority (80%) of patients underwent nephroureterectomy. The other surgical procedures used were nephrectomy (4%), total ureterectomy (1%), distal ureterectomy (11%), and transurethral resection of the intramural ureter (4%). One patient who had initially received a distal ureterectomy went on to require an ipsilateral nephroureterectomy for residual cancer. Six patients (4%) developed contralateral upper urinary tract tumors. Of these, 2 were treated with contralateral nephroureterectomy and dialysis, 2 with distal ureterectomy, 1 with complete ureterectomy, and 1 with a transurethral resection of the intramural ureter.
The median patient follow-up time was 5.9 years, and 82 patients had died by the date of last follow-up (49% of the cohort): 44 from their upper urinary tract cancer and 38 from unrelated causes. Competing risks therefore accounted for a substantial proportion of overall mortality in patients with upper urinary tract tumors (Fig. 1). An additional 13 patients had a tumor recurrence and had not yet died of their cancer at the time of last follow‒up. Three of these patients had an isolated ipsilateral local tumor recurrence and 10 had developed metastases. Systemic chemotherapy was administered neoadjuvantly in 2 (1%) patients, adjuvantly in 8 (5%) patients, and in a salvage setting for metastases in 6 (4%) patients.
Preoperative Predictors of Survival
Univariate survival modeling identified 4 potential preoperative predictors: clinical tumor stage, smoking status, the presence of constitutional symptoms, and increasing patient age at surgery. Sex did not predict mortality in any univariate model. Competing risks cumulative incidence plots of cancer mortality by preoperative variables are shown in Figure 2. These clinical variables were then modeled together in a multivariate Cox regression model that censored competing risks (Table 2) and in a multivariate Fine-Gray competing risks regression that explicitly models the competing causes of mortality (Table 3). In both the competing risks and Cox regression frameworks, the most important predictor of cancer mortality was found to be the presence of a clinically invasive tumor, which quadrupled the risk of dying from the upper urinary tract malignancy. Increasing age was the risk factor most linked to noncancer causes of mortality, with every 5 years of increasing age resulting in a 40% increased risk of noncancerous death. Although the presence of constitutional symptoms was found to be a significant predictor of OS and a borderline predictor of CSS and mortality due to competing risks, it did not predict cancer mortality in the competing risks regression model. Cigarette smoking was not found to be a significant predictor of mortality in any of the models tested.
|Parameter||Overall Survival||Cancer-specific Survival|
|Preoperative Model||HR||95% CI||P||HR||95% CI||P|
|Pain or weight loss||1.89||1.19-3.00||.007||1.76||0.94-3.32||.080|
|Invasive clinical stage||1.87||1.10-3.19||.021||3.96||1.83-8.58||<.001|
|Pain or weight loss||1.91||1.18-3.09||.013||1.76||0.93-3.34||.122|
|Competing Risks Mortality||Cancer Recurrence and Mortality|
|Preoperative Model||HR||95% CI||P||HR||95% CI||P|
|Pain or weight loss||1.92||0.96-3.82||.064||1.28||0.72-2.28||.402|
|Invasive clinical stage||0.30||0.12-0.77||.012||3.97||2.08-7.58||<.001|
|Pain or weight loss||1.91||1.02-3.56||.015||1.08||0.58-2.01||.798|
Postoperative Predictors of Survival
Univariate modeling of the postoperative predictors identified additional predictors of mortality: the pathologic TNM stage, tumor grade, and the presence of positive lymph nodes. Given the very strong correlations between lymph node positivity and tumor grade and stage, this variable was not modeled in multivariate models. An additional variable, the anatomic site of the tumor (renal pelvis, ureter, or both) was found to be significant in univariate competing risks analyses for mortality but induced significant collinearity when modeled with other variables and was, therefore, removed from multivariate models. Neither the type of surgical procedure nor the tumor histology predicted death due to cancer. Competing risks cumulative incidence plots of postoperative predictors of cancer mortality are shown in Figure 3.
The postoperative variables were then combined with the preoperative variables in regression models predicting postoperative survival outcomes (Tables 2 and 3). Given the extremely high correlation between clinical and pathologic stage, the clinical stage was not included in any postoperative model. In both competing risks and Cox regression models, tumor grade was found to be the most important predictor of cancer mortality, with patients with high-grade tumors experiencing a 4-fold to 6-fold increase in the probability of dying of their malignancy. Whereas the postoperative models did not find pathologic tumor stage to be an important predictor of survival, their high correlation with tumor grade (Spearman rho = 0.59; P < .001) explains the important attenuation of its univariate predictive effect observed in the multivariate models. Increasing age and the presence of constitutional symptoms both predicted noncancer mortality.
Predicting survival in patients with tumors is one of the most important tasks of physicians who treat cancer. Accurate assessment of the risk of death impacts greatly on the choice of treatment, particularly when extirpative and potentially morbid surgical procedures are being considered. In the case of upper urinary tract tumors, several prognostic factors have been proposed over the past 40 years. The most commonly cited predictors of survival for these tumors include the following: tumor stage,2, 9-21 tumor grade,10, 11, 13-15, 21 positive lymph nodes,22, 23 tumor DNA ploidy,13, 24 tumor localization (ureter vs renal pelvis),11, 12 tumor multifocality,2, 11, 13, 25 coincident bladder cancer,26 type of surgical procedure,2, 9, 11, 17 performance status,27 patient age,2, 9, 20 and female sex.2, 21 Although each of these variables may appear to be an important predictor of survival in univariate analyses, it is more appropriate to consider groups of covariates for prognostic purposes. Multivariable statistical modeling can aid greatly in this regard by identifying the most critical predictors and modeling them in a manner that is both meaningful and practical.
In the current study, we evaluated predictors of survival in a large cohort of upper urinary tract tumors. We attempted to separate the predictors into 2 categories: those that are known preoperatively and those that are known postoperatively. Preoperative risk factors are important because they help to identify patients at high risk for disease recurrence and, in theory, these patients may benefit from an additive form of treatment before surgery, such as neoadjuvant chemotherapy. Unfortunately, to our knowledge, there are no randomized studies evaluating neoadjuvant treatments in patients with upper urinary tract tumors. This may be in part because high-risk patients have traditionally been difficult to identify preoperatively. Our data suggest that clinically invasive tumors carry a very poor prognosis and present a population worthy of targeting for additional therapy. Given the mounting evidence supporting neoadjuvant chemotherapy for urothelial carcinomas of the bladder, it may be sensible to consider high-risk patients with urothelial carcinomas of the upper urinary tract for similar treatment.28, 29
Postoperative predictors have the advantage of being generally more accurate than preoperative predictors in identifying patients at high risk of dying of their cancer. We found that histologic tumor grade and pathologic tumor stage were the most important predictors of mortality from upper urinary tract cancer. From a clinical standpoint, these 2 factors could be used to help identify patients who might benefit from adjuvant therapy. There is evidence to suggest, for example, that adjuvant radiotherapy may benefit patients with high-risk disease,20, 30, 31 although not all groups agree with this finding.32 Similarly, adjuvant chemotherapy has been shown to be of some benefit to high-risk patients.30, 33-35 In the current cohort of patients, if only grade 3 and pT3/T4 tumors were selected, approximately one-third of patients could have potentially been offered some form of early adjuvant therapy in the hope of improving outcomes. However, given that the loss of a renal unit causes a drop in global renal function that may result in the avoidance of cisplatin, the most active agent in treating urothelial carcinomas, chemotherapy is most likely best given neoadjuvantly for upper urinary tract tumors.
Unlike other studies, we did not find that the female sex,2, 21 the type of surgical procedure,2, 9, 17 increasing age,2, 9, 20 tumor localization,11, 12 or tumor multifocality2, 11, 13, 25 carried an independent risk of mortality from upper urinary tract cancer, once competing risks were considered. The finding that many seemingly important covariables were excluded from our model can be explained in several ways. First, we only selected variables with a univariate P value ≤.2 for inclusion in our multivariable models. Although this is believed to be a conservative cut point that maximizes the chance of retaining truly important covariables, it is possible that important covariables were excluded.36 Second, our cohort of patients is relatively small for predictive modeling and, as such, may not be truly representative of this disease. However, given the rarity of upper urinary tract tumors, the only realistic solution to this problem is data-sharing to validate our model in a larger cohort of patients, something we will be looking forward to in the future. Third, pathologic stage and grade appear to be highly correlated with many of the covariables not included in the final model. For example, 82% of lymph node-positive patients have grade 3 tumors, and these are stage pT3 or pT4 in 88% of cases. Because stage and grade are very powerful predictors of survival and also strongly influence the probability of positive lymph nodes, entering lymph node status into the model results in collinearity and does not improve model predictions. It is interesting to note that although the female sex is not correlated with stage or grade, it is highly correlated with the presence of symptoms, another important predictor of survival.
The type of surgical procedure performed in our cohort of upper urinary tract tumors was a poor predictor of survival. Because 80% of patients underwent open nephroureterectomy, patient selection could be masking any true deleterious effect of nephron-sparing surgery. However, we believe that this is unlikely, and evidence exists that corroborates our results. For example, in low-grade low-stage disease, endoscopic surgery has been shown to be equivalent to open surgery in terms of survival.37-42 This suggests that, when patients are selected for nephron-sparing surgery, the primary determinants of survival are grade and stage, not the surgical procedure.
A key observation of the current study is the critical importance that competing mortality plays in patients with upper urinary tract tumors. In fact, with a median follow-up of nearly 6 years, appoximately one-third of all patient deaths in the current study were directly attributable to noncancer causes. Age is a decisive factor to consider when counseling patients with upper urinary tract tumors, because it has a dramatic effect on the risk of competing causes of mortality. Even after considering their cancer burden, we found that every year that our patients aged increased their risk of a noncancer death by approximately 7%. This is perhaps best explained by noting that the median age of our patient cohort was 70 years, an age at which the comorbidity load increases rapidly. Unlike other studies, we did not find that age predicted cancer mortality.2, 9, 20 Although one might surmise that an under-representation of young people could explain why age was not an important predictor of cancer-specific survival, we are aware of no biologic reason why older patients should have more aggressive tumors than younger patients. However, there is ample reason to believe that increasing age is associated with a rising prevalence of significant medical comorbidities and that these represent the true underlying reason that elderly patients with cancer are more likely to die from competing risks.
Stage and grade are the 2 most important independent predictors of survival and appear to bee highly correlated. In the current study, the presence of pain or weight loss were identified as a novel predictor of survival in this cancer. We did not find a survival disadvantage for women, patients who underwent nephron-sparing surgery, patients with ureteral tumors, or older patients with respect to cancer. Competing causes of mortality are responsible for greater than one-third of deaths in patients with upper urinary tract tumors and that age is the best predictor of this occurrence. Hence, elderly patients with upper urinary tract carcinomas are at high risk of competing deaths, suggesting that medical comorbidities play an important role in determining outcome in this patient population and ought to be considered in treatment decision making.
Conflict of Interest Disclosures
The authors made no disclosures
- 5Regression Modeling Strategies. New York: Springer-Verlag; 2001..
- 6Multiple Imputation for Nonresponse in Surveys. 1st ed. Vol 1. New York: John Wiley & Sons; 1987..
- 7Competing Risks: A Practical Perspective. Hoboken, NJ: John Wiley & Sons; 2006..
- 8Modeling Survival Data: Extending the Cox Model. New York: Springer; 2000., .
- 36Applied Survival Analysis: Regression Modeling of Time to Event Data. 1st ed. New York: Wiley-Interscience; 1999., .