Patients with renal cell carcinoma nodal metastases can be accurately identified: External validation of a new nomogram

Authors


Abstract

Outcome of patients with renal cell carcinoma nodal metastases (NM) is substantially worse than that of patients with localized disease. This justifies more thorough staging and possibly more aggressive treatment in those at risk of or with established NM. We developed and externally validated a nomogram capable of highly accurately predicting renal cell carcinoma NM in patients without radiographic evidence of distant metastases. Age, symptom classification, tumour size and the pathological nodal stage were available for 4,658 individuals. The data of 2,522 (54.1%) individuals from 7 centers were used to develop a multivariable logistic regression model-based nomogram predicting the individual probability of NM. The remaining data from 2,136 (45.9%) patients from 5 institutions were used for external validation. In the development cohort, 107/2,522 (4.2%) had lymph node metastases vs. 100/2,136 (4.7%) in the external validation cohort. Symptom classification and tumour size were independent predictors of NM in the development cohort. Age failed to reach independent predictor status, but added to discriminant properties of the model. A nomogram based on age, symptom classification and tumour size was 78.4% accurate in predicting the individual probability of NM in the external validation cohort. Our nomogram can contribute to the identification of patients at low risk of NM. This tool can help to risk adjust the need and the extent of nodal staging in patients without known distant metastases. More thorough staging can hopefully better select those in whom adjuvant treatment is necessary. © 2007 Wiley-Liss, Inc.

Renal cell carcinoma (RCC) accounts for ∼3% of cancers in adults as well as 85% of all primary malignant kidney tumours.1 The 5-year survival rate for all stages of RCC has improved in the recent years due to an important stage migration, whereby the majority of patients are diagnosed with localized disease.2 However, nearly 25% of contemporary patients are diagnosed with advanced disease, which includes either distant or nodal metastases (NM).2 The prevalence of exclusive node-positive RCC ranges from 2 to 10%, across different cohorts. The reported survival rate of patients with exclusive NM demonstrates great variability, and ranges from 5 to 40%.3, 4, 5, 6, 7, 8, 9 Even after consideration of staging biases, these survival rates are in clear contrast with survival estimates from patients with localized disease (T1-3 N0 M0), where 5-year survival exceeds 85%. Therefore, patients with NM have clearly more aggressive disease than even the worst variant of localized disease. This possibly justifies more aggressive treatment in patients with NM, such as nephrectomy and resection of the NM or nephrectomy followed by immediate systemic therapy.3, 9 Better ability to identify patients with NM can be useful in pre-operative counselling, targeting patients for potential involvement in adjuvant therapy trials, and may also assist in surgical planning if more extensive or complete nodal dissections are contemplated.3, 9

In this manuscript we address the issue of accurate identification of patients at risk of NM and we develop a multivariable (MVA) logistic regression model-based nomogram for this purpose.

Abbreviations:

AUC, area under the curve; CT, computed tomography; MVA, multivariable analyses; NM, nodal metastases; PA, predictive accuracy; RCC, renal cell carcinoma; UVA, univariable analyses.

Material and methods

Patient population

Patient records were retrieved from institutional databases of 12 participating institutions and yielded 5,544 patients treated with either partial or radical nephrectomy between 1984 and 2001 (Table I). Of these, 665 (12.0%) individuals were excluded from analyses because of missing fields: 626 for missing symptom classification, 29 for missing tumour size, 9 for missing age and 1 for missing T stage. Finally, 4,658 (84.0%) individuals were included in all analyses.

Table I. Descriptive Characteristics of the Study Cohort (N = 4658)
VariablesIn development cohortExternal validation cohortp-value
Total2,522 (100.0%)2136 (100.0%) 
Centers  <0.001
 Urology Unit, “G. Rummo” Hospital, Benevento, Italy200 (7.9%) 
 St. Etienne University Hospital, St. Etienne, France483 (19.2%) 
 Graz Medical University, Graz, Austria1,074 (42.6%) 
Radboud University Nijmegen Medical Centre, Nijmegen, The Netherlands43 (1.7%) 
 Angers University Hospital, Angers, France209 (8.3%) 
 Grenoble University Hospital, Grenoble, France167 (6.6%) 
 Necker Medical School, Paris, France346 (13.7%) 
 Rennes University Hospital, Rennes, France872 (40.8%) 
 Henri Mondor University Hospital, Creteil, France324 (15.2%) 
 University of Verona, Verona, Italy638 (29.9%) 
 University of Padova, Padova, Italy252 (11.8%) 
 Brest University Hospital, Brest, France50 (2.3%) 
Age (years)  <0.001
 Mean (median)61.7 (63.0)60.4 (62.0) 
 Range20–8821–91 
Clinical T stage  <0.001
 T1a847 (33.6%)657 (30.8%) 
 T1b615 (24.3%)505 (23.6%) 
 T2277 (11.0%)321 (15.0%) 
 T3764 (30.3%)634 (29.7%) 
 T419 (0.8%)19 (0.9%) 
Symptom classification  <0.001
 Asymptomatic1,812 (71.8%)1,009 (47.2%) 
 Local487 (19.4%)832 (39.0%) 
 Systemic223 (8.8%)295 (13.8%) 
Tumor size (cm)  <0.001
 Mean (median)5.6 (5.0)6.2 (5.5) 
 Range1.0–26.00.5–23.0 
Nodal metastases107 (4.2%)100 (4.7%)0.5

Clinical and pathologic evaluation

The symptom classification was defined as previously described.10 Briefly, asymptomatic patients were those with no local or systemic symptoms that could be attributed to RCC and their diagnoses were incidental. Local symptoms consisted of lumbar pain, hematuria or palpable mass. Finally, systemic symptoms consisted of anorexia, asthenia, weight loss or symptoms due to presence of metastasis. Patients were staged pre-operatively with computed tomography (CT) of the abdomen and pelvis, chest CT or chest X-ray, serum electrolytes and liver function tests. Tumour size was defined according to the greatest tumour diameter in centimeters observed on CT scans. Presence of NM was defined according to lymphadenectomy findings. In all cases a hilar lymphadenectomy was performed and included all lymph nodes on the ipsilateral side of the great vessels. In select cases, based on surgeon preference, more extensive lymphadenectomies were performed. In all cases, presence of NM was confirmed pathologically. Presence of distant metastases was confirmed based on radiographic findings and all patients with distant metastases were excluded from analyses.

Statistical analyses

The statistical methods consisted of 2 steps: (i) development of a MVA logistic regression model-based nomogram for prediction of lymph node metastases; (ii) external validation of the nomogram in an independent dataset.

The study sample (N = 4,658) was randomly divided into 2 subcohorts of 2,522 (54.1%) and 2,136 (45.9%) individuals, according to the referral centers. The subcohort of 2,522 patients from 7 centers was used to develop the nomogram. Age, symptom classification and tumour size were used as covariates in MVA logistic regression models addressing the presence of NM.

Subsequently, the new nomogram was externally validated in 2,136 patients from 5 other centers. This group represented the external validation cohort. The area under the curve (AUC) method was used to quantify its MVA predictive accuracy (PA). The PA was also tested in the external validation cohort for each individual predictor. Finally, a calibration plot was fitted to explore the extent of over or under estimation of the observed rate of NM. All statistical tests were performed using S-PLUS Professional, version 1 (MathSoft Inc., Seattle, Washington). All tests were two-sided with a significance level at 0.05.

Results

Demographic and baseline characteristics of the 4,658 individuals with RCC are summarized in Table I. In the development cohort, patient age ranged from 20 to 88 years (mean 61.7, median 63.0). Tumour size ranged from 1.0 to 26.0 cm (mean 5.6, median 5.0). Regarding the symptom classification, 1,812 (71.8%) patients were asymptomatic. Conversely, 487 (19.4%) had local and 223 (8.8%) systemic symptoms. Nodal metastases were detected in 107 (4.2%) individuals.

In the external validation cohort, age, symptom classification and tumour size characteristics were similar, but not statistically the same (all p < 0.001) (Table I). For example, the average age of patients in the development cohort was 61.7 vs. 60.4 years in the external validation cohort (p < 0.001). The clinical stages demonstrated minor variation between the 2 cohorts. For example, in the development cohort, T1a stage accounted for 33.6 vs. 30.8% in the external validation cohort. Conversely, in the development cohort T2 stage accounted for 11.0 vs. 15.0% in the external validation cohort. Finally, mean tumour size was 5.6 cm in the development cohort vs. 6.2 cm in the external validation cohort (p < 0.001). The rate of NM was virtually the same in both cohorts, namely 4.2 vs. 4.7% in the development and external validation cohort (p = 0.5), respectively.

Table II shows univariable (UVA) and MVA logistic regression models predicting the probability of NM in the development cohort (N = 2,522). In UVA analyses, relative to asymptomatic patients, the presence of systemic symptoms was associated with a 6.1-fold increase in the rate of NM (p < 0.001) vs. a 3.5-fold increase for the presence of local symptoms (p < 0.001). The effects of age and tumour size were modeled as cubic splines. The effect of each predictor variable is shown in Figures 1a1c. Age had a bimodal effect on the rate of NM, where the probability of NM was highest in the youngest and in the oldest individuals (Fig. 1a). However, this effect failed to reach statistical significance (p = 0.1). Therefore, from a practical perspective, age had no effect on the rate of NM. The effect of symptom classification (Fig. 1b) paralleled the odds ratios shown in the univariable analyses (Table II). The effect of tumour size displayed in Figure 1c, demonstrated a statistically significant (p < 0.001) and steep increase in the probability of NM up to a breakpoint of 8 cm. From 8 cm onwards the increase persisted but at a substantially lower rate.

Figure 1.

(a, c, d, f) The solid line represents the relation between predicted probability and log odds of nodal metastases. The dotted lines above and below the solid line represent the upper and lower 95% confidence interval. (b, e) Solid squares point the estimate and the solid line above and below each square represent the upper and lower 95% confidence interval.

Table II. Univariable and Multivariable Logistic Regression Models Predicting the Probability of Nodal Metastases at Nephrectomy
PredictorsUnivariable analyses or; p-valueMultivariable analyses
or; p-valueor; p-value
  • 1

    Coded as cubic splines

Age1− ; 0.1– ; 0.2
Tumor size1– ; <0.001– ; <0.001– ; <0.001
Symptom classification– ; <0.001– ; <0.001– ; <0.001
 Local vs. asymptomatic3.5; <0.0012.0; 0.0042.0; 0.004
 Systemic vs. asymptomatic6.1; <0.0012.8; <0.0012.9; <0.001

In MVA analyses, local, as well as systemic symptoms and tumour size achieved independent predictor status (all p ≤ 0.004). Individuals with local symptoms demonstrated a 2-fold higher rate of NM (p = 0.004), relative to individuals without symptoms, when all other variables were held constant. Similarly, patients with systemic symptoms had a 2.8-fold increase in the rate of NM (p < 0.001) vs. asymptomatic patients. The effects of age and tumour size were virtually the same as in the UVA analyses. As in UVA analyses, age failed to reach independent predictor status (p = 0.2). In consequence, from a practical perspective, after accounting for the effect of symptom classification and tumour size, age does not affect the rate of NM. The graphical display of the effect of these variables on the rate of NM is shown in Figures 1d1f.

External validation of the MVA nomogram demonstrated 78.4% accuracy. Despite its non-statistically significant effect in either UVA or MVA models, age improved the discriminant properties of the MVA model by 0.6%. Based on this consideration, age was kept among the nomogram predictors. Figure 2 shows the MVA nomogram, where age, symptom classification and tumour size findings define the risk of NM in RCC patients. In this nomogram, tumour size is associated with the highest number of risk points, where a 26-cm tumour contributes to 100 risk points. Presence of systemic symptoms contributes to 20 risk points. Finally, advanced age (88 years) contributes to 22 points. Figure 3 shows the nomogram calibration plot, where the nomogram predicted probabilities are shown on the x-axis and the observed rate of NM is shown on the y-axis. Perfect predictions are represented by the 45° line. Interestingly, the nomogram predictions are virtually overlapping with what would be considered as perfect predictions.

Figure 2.

Multivariable logistic regression model-based nomogram, where age, symptom classification and tumour size findings define the individual probability of nodal metastases at nephrectomy. Nomogram instructions: To obtain nomogram-predicted probability of nodal metastases at nephrectomy, locate patient values at each axis. Draw a vertical line to the “Points” axis to determine how many points are attributed for each variable value. Sum the points for all variables. Locate the sum on the “Total Points” line to assess the individual probability of nodal metastases at nephrectomy on the “Probability of nodal metastases” line.

Figure 3.

Multivariable logistic regression model-based nomogram calibration plot predicting nodal metastases at nephrectomy, where the x-axis denotes the predicted probability and the y-axis denotes the observed fraction. The 45° line represents ideal predictions, while the solid line represents the external logistic calibration (validation). The predictive accuracy (AUC) was 78.4%.

Discussion

RCC patients with exclusive NM represent a small albeit high-risk group which deserves particular interest in the context of novel adjuvant or systemic treatment modalities, such as the kinase inhibitors.

Recently published data clearly indicate that the presence of NM in RCC patients has a dismal effect on survival.11 The prognostic implication of NM in RCC3 appears comparable to those of prostate12 or bladder13 cancers, where NM drastically worsen the prognosis. The dismal survival of patients with NM may be improved with the advent of kinase inhibitors, such as sorafenib and sunitinib.14 These agents demonstrated progression-free survival benefits in patients with distant metastases. Their effect on patients with NM is under investigation within ongoing randomized clinical trials.11 While awaiting the results of these trials, it might be postulated that the same or similar progression-free survival benefits will be seen in patients with exclusive NM, if tyrosine kinase inhibitors are used.

Several investigators suggested that lymphadenectomy should be considered in patients who are believed to be at a non-negligible risk of NM (EAU Guidelines update 2006/www.uroweb.org). However, as of today there are no clear guidelines regarding exactly who should be subjected to a staging lymphadenectomy with the intent of identifying those with NM. For example, Blom et al. demonstrated that in T1-3 N0 M0 RCC patients the rate of unsuspected lymph node metastases is as low as 3.3%.15 Such data indicate that the majority of RCC patients will be free of NM. This is indeed the case in our series, where respectively 4.2 and 4.7% of patients had NM in the development and external validation cohorts. This rate corroborates our previous findings and those of others.9, 16, 17 Based on this low rate of baseline NM prevalence, it is clear that not all individuals with clinically non-metastatic RCC should be subjected to a lymph node dissection. On the contrary, it appears that the great majority of patients do not benefit of a staging lymphadenectomy, as ∼95% will not harbour NM. In consequence, risk stratification tools, such as nomograms, are needed to avoid unnecessary lymphadenectomies in low NM risk patients. Based on this consideration, our analysis attempted to identify readily available pre-operative variables that could help risk stratify RCC patients with exclusive NM.

Our data demonstrated that tumour size and symptom classification represent strong, independent predictors of the probability of NM. Despite its lack of statistical significance, age improved the discriminant properties of these 2 variables and was included among the nomogram predictors. Although it is tempting to state that youngest and oldest patients are at higher risk of NM and that RCC diagnosis at an early (<30 years) or at an older (>80 years) age are associated with a higher prevalence of NM, the lack of statistical significance within our data would make such assumption at best speculative. Moreover, the clinical nature of our data precludes any valid attempts at defining the aetiology of age-specific distribution of NM. The combined contribution of age, tumour size and symptom classification for prediction of NM resulted in 78.4% PA, in the external validation cohort. This implies that the nomogram can accurately predict the probability of NM in ∼8/10 patients. Risk stratification can help identifying individuals at either low or high probability of NM. In consequence, we believe that it can help reducing the number of unnecessary lymph node dissections that demonstrate surgically negative lymph nodes in the absence of NM.

For example, if a nomogram threshold of 2% is used and lymph node dissections are not performed in all patients below this cut-off, then 665 (31.1%) individuals from the external validation cohort of 2,136 patients would fall below the cut-off value. In the external validation cohort the negative predictive value of the nomogram cut-off of 2% would be 99.7% and 2/100 patients with positive lymph nodes would be missed. The use of a higher cut-off, such as for example 4%, would result in avoidance of 1,064 (49.8%) lymph node dissections, a negative predictive value of 99.3% and 8 missed patients out of 100 with positive lymph nodes.

The benefit of more accurate selection of lymphadenectomy candidates are 2-fold. First, adverse effects associated with lymph node dissections could be reduced. Although data on lymph node dissection in RCC are scant, it could be postulated that the performance of lymph node dissections can add to the morbidity of nephrectomy.15 Second, accurate risk stratification may entice urologic surgeons to perform staging lymph node dissections in high-risk patients. Currently, many surgeons omit a lymphadenectomy at nephrectomy and possibly understage at least 4.2% of patients.

Several limitations apply to our study. First, our analyses are based on hilar lymph node dissections. Hilar lymph nodes do not represent a sentinel landing zone for RCC. Lymph node metastases may be situated in the inter-aorto-caval region as well as contralateral to the great vessels. Skip areas may exist, so that hilar nodes may be exempt of metastases but inter-aorto-caval or contralateral nodes may be involved. Therefore, our study needs to be interpreted as a prediction of the likelihood of invasion of hilar lymph nodes. Despite the limitation in the extent of lymph node dissection, the resection of hilar nodes still represents the standard of care, when a lymph node dissection is performed in RCC (EAU Guidelines update 2006/www.uroweb.org).

Second, our nomogram rests on age, symptom classification and tumour size. These predictors, except for tumour size are not specifically related to the likelihood of NM. In an ideal situation, our nomogram would include specific NM predictors. Unfortunately, there are no reliable biomarkers that can be used for this purpose yet.

Third, our population exclusively originates from European centers. It is possible that patients with genotypically or phenotypically different RCC may have a different prevalence of NM. Moreover, it is possible that the association between the predictors and NM may differ from what we observed. Therefore, ideally this nomogram should be externally validated in other regions of the world prior to its adoption into clinical practice outside of Europe.

Conclusions

Our nomogram can contribute to the identification of patients at low risk of NM. This tool can help to risk adjust the need and the extent of nodal staging in patients without known distant metastases.

Acknowledgements

Pierre I. Karakiewicz is partially supported by the Fonds de la Recherche en Santé du Québec, the CHUM Foundation, the Department of Surgery and Les Urologues Associés du CHUM.

Ancillary