Endometrial cancer is the most common gynecological cancer, affecting approximately 25.7 women per 100,000 each year. Despite better treatments being available, mortality rates have increased by 21% over the past two decades - a troubling trend that points to gaps in early detection.
The current approach to catching endometrial cancer early depends on patients recognizing symptoms like abnormal uterine bleeding and seeking medical care. But research shows that Black women and women from minority groups are less likely to recognize or act on these symptoms, leading to later diagnosis and worse outcomes.
There is no routine screening test for endometrial cancer equivalent to the mammogram for breast cancer or the Pap smear for cervical cancer. This gap makes AI-based risk prediction tools particularly valuable for identifying high-risk women before symptoms appear.
This study focused on predicting not just endometrial cancer itself but also endometrial intraepithelial neoplasia (EIN) - a precancerous lesion of the uterine lining that can progress to cancer if untreated. Identifying EIN early allows intervention before cancer develops.
The goal was to build an AI model using simple clinical features that are readily available without specialized testing - features like age, BMI, endometrial thickness on ultrasound, and symptoms. This approach makes the tool usable in primary care settings, not just specialist clinics.
A particular motivation was reaching women in rural and minority communities who may not have access to advanced diagnostics. A risk prediction tool based on widely available clinical information could flag high-risk women for further evaluation even in resource-limited settings.
The study included 564 patients from Suleyman Demirel University in Turkey. Clinical features collected included age, menopause status, bleeding history, obesity, hypertension, diabetes, smoking, endometrial thickness, and history of breast cancer.
Six machine learning methods were compared: Random Forest, logistic regression, multilayer perceptron (a type of neural network), CatBoost, XGBoost, and Naive Bayes. This comprehensive comparison allowed the researchers to identify which approach worked best for this particular prediction task.
Because endometrial cancer and EIN are relatively uncommon (7.9% prevalence in this cohort), the training data had a class imbalance problem - many more negative cases than positive ones. The synthetic minority oversampling technique (SMOTE) was used to balance the training data, helping the models learn from both classes equally.
The Boruta algorithm - a feature selection method that identifies which input variables are truly important versus redundant - was applied to the full set of 9 clinical features. It identified just 3 features as essential: age, body mass index (BMI), and endometrial thickness.
This three-feature model is clinically appealing because endometrial thickness can be measured with a standard vaginal ultrasound, and age and BMI are always available from routine clinical records. No blood tests or biopsies are needed to generate a risk prediction.
Each model was optimized using 5-fold cross-validation on the training set (451 patients), then evaluated on a held-out internal validation set (113 patients). Hyperparameter tuning and boosting were applied to maximize model performance.
The best performing model was the multilayer perceptron (neural network), achieving an AUC (area under the ROC curve) of 0.938 for predicting endometrial intraepithelial neoplasia and endometrial cancer. This is a strong result, meaning the model correctly distinguished high-risk from low-risk patients nearly 94% of the time.
Overall accuracy was 0.94 (94%), reflecting the model's ability to correctly classify the majority of patients. However, precision was 0.71 and recall was 0.50, indicating that while the model rarely raises false alarms, it misses about half of true positive cases in the validation set.
The three-feature model (age, BMI, endometrial thickness) performed comparably to or better than models using all 9 features, confirming that the feature selection step successfully identified the most informative predictors without unnecessary complexity.
A key strength of this model is that it does not depend on menopausal status or specific symptoms. Current clinical pathways for endometrial cancer screening are largely triggered by postmenopausal bleeding, which means asymptomatic women or those who fail to recognize or report bleeding may be missed entirely.
By predicting risk from age, BMI, and endometrial thickness - all obtainable from routine health assessments - the model could be applied during standard gynecological or primary care visits to flag women who should be evaluated more thoroughly.
The authors specifically highlight the potential benefit for Black women and minority populations, for whom symptom-based detection has historically been less effective. An AI tool that does not require symptom recognition could help close this disparity in endometrial cancer outcomes.
This study demonstrates that AI can predict endometrial cancer and precancer risk with high accuracy using only three widely available clinical features: age, BMI, and endometrial thickness. The multilayer perceptron achieved an AUC of 0.938, making it a promising screening tool.
The simplicity of the required inputs - obtainable in any clinical setting with an ultrasound machine - means this tool could be implemented far more broadly than specialized molecular tests, reaching patients in rural, low-resource, and minority communities.
Future work should focus on external validation in diverse populations, addressing the relatively low recall for positive cases (50%), and exploring integration with telehealth platforms that could allow AI-guided risk assessment in settings without specialist gynecological care.