Machine Learning-Based Molecular Classification of Endometrial Cancer Using Clinicopathological and Immunohistochemical Features: A Chinese Population Study

Abstract

The Cancer Genome Atlas (TCGA) molecular classification of endometrial cancer (EC) into POLEmut, MMRd, p53abn, and NSMP subtypes has substantially improved prognostic stratification. However, molecular classification typically depends on sequencing, whereas surrogate clinical workflows combine immunohistochemistry (IHC) with targeted POLE testing, limiting accessibility in many clinical settings. This study developed, evaluated, and interpreted machine learning models for automated EC molecular subtype classification using routinely available clinicopathological and IHC features, and assessed the relative contributions of molecular and clinical predictors through a controlled feature-ablation analysis. A cohort of 229 Chinese patients with EC was analyzed. Twenty-one features were engineered from clinical and molecular variables, and six ML algorithms (Random Forest, Gradient Boosting, Logistic Regression, Support Vector Machine, Naive Bayes, and MultiLayer Perceptron) were evaluated using both 5-fold and 10-fold stratified cross-validation. Class imbalance was addressed using SMOTE applied exclusively within the training folds. On the full feature set, Gradient Boosting achieved 96.1% accuracy in 5-fold cross-validation and 97.0% in 10-fold cross-validation (macro $\mathrm{F} 1 = 9 7. 0 \%)$, exceeding the performance of a transparent rulebased classifier (93.9% accuracy). Feature importance analysis identified POLE (26.7%), TP53 (20.5%), and mismatch-repair (MMR) markers (20.5% combined) as the dominant contributors to classification. Feature-ablation experiments showed that models trained using only molecular features maintained performance comparable to the full-feature models, whereas removing molecular markers reduced accuracy to 52-60% (macro $\mathrm{F} 1 \leq 0. 4 4)$, approaching the majority-class baseline (63.8%). These findings indicate that the predictive performance is primarily driven by molecular biomarkers, while routine clinicopathological variables alone provide limited discriminatory power. The proposed framework provides an accurate, automated, and interpretable approach for implementing IHC-based molecular classification of endometrial cancer in routine clinical practice.