Objective
Our objective was to develop a nomogram based on pathological hysterectomy characteristics to provide a more individualized and accurate estimation of lymph node metastasis in endometrial cancer.
Study Design
Data from the Surveillance, Epidemiology, and End Results database for 18,294 patients who underwent hysterectomy and lymphadenectomy were analyzed. A multivariate logistic regression analysis of selected prognostic features was performed, and a nomogram to predict lymph node metastasis was constructed. A cohort of 434 patients was used for the external validation.
Results
The nomogram showed good discrimination with an area under the receiver operating characteristic curve of 0.80 (95% confidence interval, 0.79–0.81) in the training set and 0.79 (95% confidence interval, 0.78–0.80) in the validation set. The nomogram was well calibrated.
Conclusion
We developed a nomogram based on 5 clinical and pathological characteristics to predict lymph node metastasis with a high concordance probability.
Endometrial cancer is the most common malignancy of the female genital tract and the seventh most common cause of death from cancer in women in Western countries. The vast majority of women with endometrial cancer are diagnosed with early-stage tumors that are associated with a good prognosis; however, a subgroup of women have more aggressive neoplasms and are at increased risk of relapse and death.
The current surgical approach for the treatment of endometrial cancer is still debated. In particular, indications for lymph node (LN) dissection have not been established and include omitting lymphadenectomy in patients with presumed early-stage and low-grade disease, performing lymphadenectomy only in patients who are at high risk for nodal metastases and performing a complete lymphadenectomy in all uterine cancer patients irrespective of grade and depth of myometrial invasion.
Several authors have suggested that complete lymphadenectomy may be associated with improved survival outcomes, particularly for patients with LN metastases. However, the retrospective nature of most of these studies has rendered their results equivocal. In contrast, the results of 2 recent randomized clinical trials showed that lymphadenectomy did not provide an overall or recurrence-free survival benefit in the early stages of disease. These trials have been criticized for the following reasons: a limited effort with respect to the extent of dissection and lymph node evaluation, too many low-risk patients, and no direct decision on adjuvant therapy based on lymphadenectomy result.
Sophisticated imaging techniques (eg, positron emission tomography/computed tomography) offer a less invasive and morbid means of evaluating LN status. However, the sensitivity of such techniques is more limited than that of lymphadenectomy, which is considered the most accurate way to detect the presence of LN metastases. An evidence-based algorithm for surgical treatment decisions could be helpful, especially when the preoperative health status of the patient is unfavorable in terms of anesthetic status or comorbidities. Clinical and pathological variables (eg, myometrial invasion, histological type and grade) have been reported to be associated with the risk of LN metastasis. However, individually, none of these characteristics can be used to identify a subset of patients for whom LN resection is unnecessary. Prognostic tools such as nomograms that use statistical models to combine variables to obtain the most accurate and reliable predictions have been adopted in several oncologic disciplines.
The aim of this study was to develop a nomogram to predict LN status for endometrial cancer by combining selected clinical and pathological risk factors using a multivariate model. It could be used when hysterectomy has been performed and lymphadenectomy omitted.
Materials and Methods
Study population
The mathematical model was developed using data from the Surveillance, Epidemiology, and End Results (SEER) database. All data were publicly available, deidentified, and exempt from institutional review board review. We identified patients with histologically proven endometrial cancer diagnosed between 1988 and 2007. Case listings were generated using codes specific for both clinical (age at diagnosis and race) and tumor characteristics (primary organ site, extent of disease, histologic subtype, grade differentiation, and number of regional LN examined). We included women who underwent primary operations for endometrial cancer with at least hysterectomy and extensive lymphadenectomy (more than 10 regional lymph nodes removed).
Development of the nomogram
To develop a well-calibrated and exportable nomogram to predict the metastatic LN risk, we built a logistic regression model (LRM) using a training cohort (ie, a training set) of 18,294 patients, which was extracted from the SEER database using the previous criteria, and we validated the model with an independent validation cohort (ie, a validation set). Univariate and multivariate logistic regression analyses were used to test the association between the metastatic LN risk and clinicopathological characteristics. The complexity of the model was controlled using Akaike information criteria. A P < .05 was considered significant. The following variables were included in the analysis: age at diagnosis; race (white, African American, and other); histological subtype (adenocarcinoma, papillary serous, clear cell, and carcinosarcoma); grade differentiation (well = 1, moderate = 2, poor = 3-4); and primary site tumor characteristics (endometrium, 50% or less myometrial invasion, greater than 50% myometrial invasion, cervical stromal invasion).
The predictive accuracy of the model was assessed in terms of its discrimination and calibration. Discrimination is the ability to differentiate between women with positive metastatic LN and women with negative metastatic LN, and it is measured using the receiver operating characteristic curve and summarized by the area under the curve (AUC). An AUC of 1.0 indicates perfect concordance, whereas an AUC of 0.5 indicates no relationship. Calibration (ie, the agreement between the observed outcome frequencies and the predicted probabilities) was studied using graphical representations of the relationship between the observed outcome frequencies and the predicted probabilities (calibration curves). A calibration curve can be approximated by a regression line with intercept α and slope β. These parameters can be estimated in an LRM with the event as the outcome and the linear predictor as the only covariate. Well-calibrated models have α = 0 and β = 1. Therefore, a sensible measure of calibration is a likelihood ratio statistic testing the null hypothesis that α = 0 and β = 1. The statistic has a χ 2 distribution with 2 df (unreliability [ U ] statistic). We also evaluated the average (E average) and maximal errors (E maximal) between predictions and observations, which were obtained from a calibration curve.
Validation
An internal validation of the accuracy estimates was performed with 200 bootstrap resamples to obtain relatively unbiased estimates. Bootstrapping allows for the simulation of the performance of the nomogram if it was applied to future patients and provides an estimate of the average optimism of the AUC.
For external validation, the model was applied on a sample of 434 patients referred to as the validation set, which was developed from a single database that recorded patient data from 4 institutions: Tenon Hospital (Paris, France; 116 patients), Bichat Hospital (Paris, France; 43 patients), E3N cohort (Paris, France; 200 patients), and Creteil Hospital (Paris, France; 75 patients). Patients were included if they had available data for the components of the nomogram. In addition, to study the predictive accuracy of the model in terms of calibration, patients were clustered into deciles according to their nomogram score. For each decile group, we calculated the difference between the predicted and the observed LN metastasis probability. The mean error between the predicted and the observed LN metastasis probability was the sum of the differences for each decile group divided by 10.
Other statistical tests
The categorical and numerical variables were analyzed using the χ 2 test and the Student t test, respectively. Differences were considered significant at a level of P < .05. All analyses were performed using the R package with the Design, Hmisc, Design, Presence/absence ( http://lib.stat.cmu.edu/R/CRAN ).
Results
Patient population
The overall data from the 18,294 patients in the training set and the 434 patients in the validation set were analyzed. Patient characteristics are summarized in Table 1 . The populations were significantly different for all characteristics studied by the nomogram. The metastatic LN frequencies for the training and validation sets were 7.89% (1443 of 18,294) and 13.13% (57 of 434), respectively.
| Variables | Training set (n = 18,294) | Validation set (n = 434) | P value |
|---|---|---|---|
| Age at diagnosis, y | |||
| Median (mean) | 62 (62.24) | 66 (64.99) | < .001 |
| Range | 18-97 | 38-87 | |
| Race | < .001 | ||
| White | 15,801 (86.4%) | 419 (96.5%) | |
| African American | 929 (5.1%) | 2 (0.5%) | |
| Other | 1564 (8.5%) | 13 (3.0%) | |
| Tumor grade | < .001 | ||
| 1 | 6512 (35.6%) | 221 (50.9%) | |
| 2 | 7092 (38.8%) | 153 (35.3%) | |
| 3-4 | 4690 (25.6%) | 60 (13.8%) | |
| Histologic subtype | < .001 | ||
| Adenocarcinoma | 16,998 (92.9%) | 399 (91.9%) | |
| Clear-cell | 307 (1.7%) | 19 (4.4%) | |
| Papillary serous | 577 (3.1%) | 14 (3.2%) | |
| Carcinosarcoma | 412 (2.3%) | 2 (0.5%) | |
| Primary site tumor characteristics | .024 | ||
| Endometrium | 3936 (21.5%) | 82 (18.9%) | |
| ≤50% | 9033 (49.4%) | 198 (45.6%) | |
| >50% | 4039 (22.1%) | 120 (27.6%) | |
| Cervical stroma invasion | 1286 (7.0%) | 34 (7.8%) | |
| FIGO 2009 stages | < .001 | ||
| I | |||
| IA | 12,578 (68.8%) | 148 (34.1%) | |
| IB | 3378 (18.5%) | 193 (44.5%) | |
| II | |||
| II | 960 (5.2%) | 43 (9.9%) | |
| III | |||
| IIIA | — | — | |
| IIIB | — | — | |
| IIIC | 1378 (7.5%) | 50 (11.5%) | |
| Extent of lymphadenectomy | < .001 | ||
| Median (mean) | 18.00 (20.85) | 10.00 (11.26) | |
| Range | 10-90.0 | 0-80.0 |
A nomogram for the prediction of metastatic lymph nodes
Table 2 summarizes the multivariate logistic regression analyses. The metastatic LN risk was independently associated with age at diagnosis, race, tumor grade, histologic subtype, and the primary site invasion. The nomogram corresponding to the model is shown in Figure 1 . For each patient, points were assigned for each of these 5 clinical variables, and a total score was calculated from the nomogram. The total points corresponded to a predicted metastatic LN probability. The prediction model had an AUC of 0.80 (95% confidence interval, 0.79–0.81) in the training set before the bootstrap technique was applied.
| Variables | OR (95% CI) | P value |
|---|---|---|
| Age at diagnosis, y | 0.99 (0.98–0.99) | < .001 |
| Race | .028 | |
| African American | Referent | |
| White | 0.74 (0.59–0.94) | |
| Others | 0.89 (0.66–1.20) | |
| Tumor grade | < .001 | |
| 1 | Referent | |
| 2 | 1.70 (1.44–2.00) | |
| 3-4 | 2.62 (2.21–3.11) | |
| Histological subtype | < .001 | |
| Adenocarcinoma | Referent | |
| Carcinosarcoma | 1.347 (1.01–1.86) | |
| Clear cell | 1.89 (1.33–2.66) | |
| Papillary serous | 2.84 (2.21–3.64) | |
| Primary site tumor characteristics | < .001 | |
| Endometrium | Referent | |
| <50% | 6.67 (4.35–10.00) | |
| >50% | 31.33 (23.48–41.00) | |
| Cervical stroma invasion | 46.13 (35.57–58.50) |
Validation of the nomogram
Discrimination
First, the nomogram was internally validated using the bootstrap correction technique. The 200 repetitions of bootstrap sample corrections provided an estimated concordance probability of 0.80 (range, 0.78–0.82) ( Figure 2 , solid line ). In the validation set, the discrimination accuracy of the model was 0.79 (range, 0.78–0.80) ( Figure 2 , dotted line ).
Calibration
The predicted probability obtained from the bootstrap correction and the actual probabilities of metastatic LN in the training set are shown in the calibration plot ( Figure 3 ). The corresponding predicted and actual probabilities of LN metastasis in the validation set are shown in the calibration plot ( Figure 4 ). There was no difference between the predicted probabilities and the observed rate of metastatic LN ( P = .142). Among the 10 decile groups, the difference between the predicted and observed probabilities of LN metastasis never exceeded 9% except for patients in the 10th decile group (nomogram score above 120) in which the difference reached 45% (the predicted and observed probabilities of LN metastasis were 31.0% and 76.0%, respectively) ( Table 3 ).
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
