Machine learning-based prediction of two-year overall survival in metastatic endometrial cancer: a SEER population-based study

Metastatic endometrial cancer carries a poor prognosis. Accurate prognostic tools are lacking. Machine learning (ML) may improve survival prediction using available clinicopathological and sociodemographic data. We analyzed 11,720 patients with metastatic endometrial cancer from the SEER database. The primary outcome was two-year overall survival (OS). Thirteen ML classifiers were trained (70/30 split) and evaluated using AUC, accuracy, and F1 score. SMOTE and calibration analysis were applied to the best-performing model. Kaplan–Meier, Cox regression, and SHAP-based interpretability analyses were performed. XGBoost achieved the highest AUC (0.726). Following SMOTE, F1 improved from 0.446 to 0.516 (AUC 0.715), with a calibration intercept and slope of -0.097 and 0.941, respectively. SHAP analysis identified histologic subtype and chemotherapy receipt as the most influential predictors, followed by tumor grade, age, and race/ethnicity. On multivariate Cox regression, Grade IV (HR 2.08), undifferentiated histology (HR 1.97), age ≥ 80 years (HR 1.71), and Non-Hispanic Black race (HR 1.35) were independently associated with worse OS; chemotherapy receipt was associated with improved observed survival (HR 0.43). Kaplan–Meier analysis confirmed significant differences in overall survival by histologic subtype and chemotherapy status (log-rank p < 0.001 for both). ML models using registry-based data predict two-year OS in metastatic endometrial cancer with modest but consistent performance. Histologic subtype, chemotherapy, tumor grade, and age are important predictors of survival. Racial disparities in survival persisted after adjustment, underscoring the need for targeted interventions to address healthcare inequities.