Endometrial cancer (EC) is the most common gynecologic malignancy in the United States, with over 66,000 new cases and 13,000 deaths recorded in 2023. Its incidence has been rising by about 1% per year among women aged 50 and older.
Most patients are diagnosed after experiencing postmenopausal bleeding (PMB), which prompts investigation. However, more than 15% of endometrial malignancies occur in postmenopausal women who have no bleeding at all - a group that is especially hard to identify and screen.
With wider use of ultrasound, many women are now found to have an incidentally thickened endometrium (the lining of the uterus) without any symptoms. This thickening occurs in 10%-17% of asymptomatic postmenopausal women, yet the actual cancer rate in this group is under 1%.
The problem is that applying the same biopsy threshold used for symptomatic women to all asymptomatic women would expose the majority of low-risk women to unnecessary invasive procedures, wasting healthcare resources and causing anxiety.
This retrospective study enrolled 971 postmenopausal women with asymptomatic endometrial thickening who underwent hysteroscopy (a procedure to look inside the uterus) between January 2017 and August 2024 at a hospital in Shanghai, China.
A total of 41 clinical variables were collected for each patient, including demographics (age, BMI, parity), ultrasound findings (endometrial thickness, Doppler blood flow signals), blood tests (CA125, D-dimer, inflammatory markers), and coagulation parameters.
The dataset was split into a training set (522 patients), an internal validation set (223 patients), and an external validation set of 226 patients from a second hospital. This multi-cohort design helps confirm that the model works beyond the original population.
LASSO regression (a statistical method that shrinks less important variables to zero) was used to narrow down the 41 features to the most predictive ones, helping prevent overfitting - a situation where a model performs well on training data but poorly on new cases.
Nine different supervised machine learning algorithms were trained and compared, including logistic regression (LR), random forest, XGBoost, LightGBM, support vector machine, gradient boosting, k-nearest neighbors, Gaussian mixture model, and decision tree.
Each model was evaluated using standard performance metrics: AUC (area under the ROC curve, measuring overall discrimination ability), sensitivity (ability to detect true cancer cases), specificity (ability to correctly exclude benign cases), and F1 score.
The SHAP (SHapley Additive exPlanations) method was applied to make model predictions interpretable. SHAP values show how much each input variable contributes to a specific prediction - both globally (which features matter most overall) and for individual patients.
A nomogram - a visual chart that converts patient data into a probability score - was also built based on logistic regression to help clinicians quickly estimate a patient's individual risk of malignancy at the bedside.
Across multiple analytical methods, five variables consistently emerged as the most important predictors of endometrial malignancy: parity (number of prior deliveries), Doppler flow signals (blood flow detected in the endometrium by ultrasound), endometrial thickness, CA125 (a tumor marker in the blood), and D-dimer (a coagulation marker).
Women in the malignant group (atypical hyperplasia or cancer) had significantly higher endometrial thickness (average 10.75 mm) compared to those with benign findings (average 8.00 mm). ROC analysis established a diagnostic threshold of 8.15 mm with a sensitivity of 75% and specificity of 68%.
Doppler flow signals indicating active blood flow within the endometrium and elevated CA125 and D-dimer levels in blood tests were independently associated with a higher risk of non-benign lesions, pointing to the role of inflammation and abnormal clotting in tumor biology.
Notably, lower parity (having had fewer pregnancies carried to term) was associated with higher cancer risk, consistent with the well-known protective effect of pregnancy on endometrial cancer risk through hormonal mechanisms.
Among all nine machine learning models tested, logistic regression (LR) showed the best overall performance, achieving an accuracy of 88%, a sensitivity of 78%, a specificity of 98%, and an AUC of 0.81 on the test dataset.
The high specificity of 98% is clinically important - it means the model correctly identifies nearly all benign cases as benign, avoiding unnecessary invasive procedures in the vast majority of low-risk women.
A nomogram built from the logistic regression model achieved consistent AUC values of 0.82 (training cohort), 0.82 (internal validation), and 0.81 (external validation), indicating robust generalizability across different patient populations.
The calibration curve - which checks whether predicted probabilities match actual observed outcomes - showed excellent consistency, meaning the model's probability estimates are trustworthy and not just directionally correct.
SHAP force plots allow clinicians to visualize how each specific variable pushes a prediction toward malignancy or benignity for an individual patient. This moves the model from a black box to a transparent decision support tool.
On a population level, SHAP summary plots rank variables by their overall importance. In this study, endometrial thickness and CA125 contributed most strongly to predictions across the entire cohort, while Doppler signals and D-dimer provided additional independent signal.
This explainability is crucial for clinical adoption. A doctor seeing a high-risk prediction can understand exactly why - for example, because of an 11mm thick endometrium combined with elevated CA125 and Doppler blood flow - rather than being asked to simply trust a number.
Combining AI interpretability tools with traditional clinical nomograms gives practitioners both a visual tool for counseling patients and a quantitative probability estimate to inform shared decision-making about the need for hysteroscopy or biopsy.
This study produced a validated, interpretable logistic regression model that can accurately identify which asymptomatic postmenopausal women with thickened endometrium are at genuine risk of precancer or cancer, and which can safely be managed expectantly.
By combining routinely available clinical data - ultrasound measurements and simple blood tests - into a structured risk score, the model could help reduce unnecessary hysteroscopic procedures in the large majority of women with benign thickening.
The model is particularly valuable given the current lack of consensus on optimal endometrial sampling thresholds for asymptomatic women. Rather than using a single thickness cutoff for everyone, a multivariable, individualized approach better captures true risk.
Future work should prospectively validate the model in larger, more diverse populations and explore how it integrates with emerging molecular biomarkers to further improve risk stratification in this growing patient group.