Characterizing impact of positive lymph node number in endometrial cancer using machine-learning: A better prognostic indicator than FIGO staging?

Abstract

Background: Number of involved lymph nodes (LNs) is a crucial stratification factor in staging of numerous disease sites, but has not been incorporated for endometrial cancer. We evaluated whether number of involved LNs provide improved prognostic value.

Patients and methods: Patients diagnosed with node-positive endometrial adenocarcinoma without distant metastasis were identified in the National Cancer Database. We trained a machine-learning based model of overall survival. Shapley additive explanation values (SHAP) based on the model were used to identify cutoffs of number of LNs involved. Results were validated using a Cox proportional hazards regression model.

Results: We identified 11,381 patients with endometrial cancer meeting the inclusion criteria. Using the SHAP values, we selected the following thresholds: 1-3 LNs, 4-5 LNs, and 6+ LNs. The 3-year OS was 82.0% for 1-3 LNs, 74.3% for 4-5 LNs (hazard ratio [HR] 1.38; p < 0.001), and 59.9% for 6+ LNs (HR 2.23; p < 0.001). On univariate Cox regression, PA nodal involvement was a significant predictor of OS (HR 1.20; p < 0.001) but was not significant on multivariate analysis when number of LNs was included (HR 1.05; p = 0.273). Additionally, we identified an interaction between adjuvant therapy and number of involved LNs. Patients with 1-3 involved LNs had 3-year OS of 85.2%, 78.7% and 74.2% with chemoradiation (CRT), chemotherapy, and radiation, respectively. Patients with 6+ involved LNs had 3-yr OS of 67.8%, 49.6%, and 48.9% with CRT, chemotherapy, and radiation, respectively (p < 0.001).

Conclusion: Number of involved LNs is a stronger prognostic and predictive factor compared to PA node involvement.

Keywords: Endometrial cancer; Machine learning; Staging.