Interpretable Machine Learning Models for Overall Survival Prediction in Stage I-II Endometrial Cancer Patients Receiving Postoperative Adjuvant Chemotherapy: A SEER-Based Study

Why clinicians should know about it

• Picked for Pathology and Forensic Medicine (top studies of the week, 23 August 2026).

Abstract

PURPOSE: Patients with stage I-II endometrial cancer receiving postoperative adjuvant chemotherapy are clinically heterogeneous, and stage alone is insufficient for individualized prognostic assessment. We developed and internally validated interpretable survival models to predict overall survival in this population. PATIENTS AND METHODS: Female patients with primary endometrial cancer diagnosed from 2010 to 2021 were identified from the SEER database. Eligible patients were aged ≥18 years, had stage I-II disease, underwent surgery, and received postoperative systemic therapy/chemotherapy. Patients were randomly assigned to training and internal validation cohorts at a 7:3 ratio, stratified by event status and stage. Cox, LASSO-Cox, random survival forest, gradient boosting machine, and XGBoost survival models were developed using prespecified clinicopathological variables. Performance was evaluated using Harrell's C-index, time-dependent AUC, IPCW Brier score, calibration curves, decision curve analysis, and risk stratification. RESULTS: Overall, 9,118 patients were included, with 6,382 in the training cohort and 2,736 in the validation cohort. The median follow-up was 87.0 months. In the validation cohort, GBM showed the highest but moderate discrimination, with a C-index of 0.6745 and AUCs of 0.7216, 0.7197, and 0.6896 at 1, 3, and 5 years, respectively. Cox and LASSO-Cox showed comparable discrimination and better calibration. Important predictors included age, tumor size, histological subtype, stage, radiotherapy, lymph node surgery extent, race, number of regional lymph nodes examined, and time from diagnosis to treatment. GBM-based risk stratification separated patients with distinct overall survival outcomes. CONCLUSION: Interpretable survival models were developed and internally validated for this clinically heterogeneous population. GBM supported risk stratification, whereas Cox provided calibrated model-based survival estimates within the available clinicopathological framework. The moderate predictive performance and absence of external validation should be considered, and future studies should incorporate additional pathological and molecular predictors.

Abstract as published, via PubMed.