Can machine learning predict endometrial cancer in patients with postmenopausal uterine bleeding?

Abstract

Introduction: Postmenopausal bleeding (PMB) is a common clinical symptom, with endometrial cancer (EC) accounting for about 10% of cases. This study aimed to develop machine-learning models to predict EC in women with PMB, supporting risk stratification and optimizing diagnostic pathways.

Material and methods: We retrospectively analysed 617 women who underwent hysteroscopy for PMB at the Meir Medical Center between 2014 and 2023. Demographic, clinical, laboratory, and imaging data were collected. Three supervised machine-learning algorithms - Random Forest, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) - were trained and evaluated. Synthetic minority oversampling technique addressed class imbalance. Model performance was assessed using accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve.

Results: Seventy-two women (11.7%) were diagnosed with EC. eXtreme Gradient Boosting achieved the highest sensitivity (80%) for detecting EC, despite moderate accuracy (65%). Random Forest and LightGBM showed higher accuracy (85% and 84%, respectively) but much lower sensitivity. The most influential predictors in the XGBoost model were tamoxifen use, hormone therapy, age, hypertension, gravidity, parity, endometrial thickness, diabetes, and duration of bleeding.

Conclusions: eXtreme Gradient Boosting provided the best clinical performance by minimizing missed diagnoses. Machine-learning models may enhance decision-making by identifying women at high risk for EC who require further evaluation. Larger datasets and additional clinical features are needed to improve specificity and reduce false positives.