Endometrial cancer and its precursor lesion, atypical endometrial hyperplasia, are diagnosed by tissue biopsy. However, the decision to proceed to biopsy is made based on clinical and imaging findings, and not all clinical features are equally predictive of malignancy or precancerous change.
Gynecologists currently rely on a combination of factors -- patient age, menopausal status, body mass index, ultrasound findings, and bleeding symptoms -- to triage who needs urgent investigation. These assessments are often made qualitatively without formal risk scoring, leading to variability in clinical practice.
Machine learning offers a path to systematically identify the most predictive combination of clinical variables and build a reproducible decision-support tool that can be used at the point of care to objectively stratify risk before invasive investigation.
The study enrolled 1,221 patients who underwent hysteroscopy or endometrial biopsy at a tertiary gynecology center. Pathological results classified patients into three groups: malignant (endometrial carcinoma), precancerous (atypical hyperplasia), and benign (non-atypical hyperplasia, polyps, normal endometrium).
Seven different machine learning algorithms were trained and compared: random forest, logistic regression, support vector machine (SVM), gradient boosting, decision tree, K-nearest neighbors (KNN), and naive Bayes. Comparing multiple algorithms is best practice because no single method is optimal for all clinical datasets.
Candidate input features were collected from routine clinical assessment: age, body mass index (BMI), menopausal status, uterine bleeding symptoms, endometrial thickness on ultrasound, resistance index of uterine blood flow, gland density on biopsy, and cellular atypia grade.
Random forest was the top-performing algorithm, achieving an ROC-AUC of 0.98 on training data and 0.96 on the held-out test set. This near-perfect discrimination indicates the model can reliably separate malignant from benign and precancerous endometrial conditions.
The model's high performance persisted across all three diagnostic categories, not just the binary malignant versus benign comparison. Accurate identification of atypical hyperplasia -- the precancerous state -- is particularly valuable because prompt treatment can prevent progression to cancer.
Other algorithms also performed well -- gradient boosting achieved AUC 0.94, and logistic regression 0.91 -- confirming that the clinical features selected contain genuine predictive signal. Random forest's margin over the others was consistent and statistically meaningful.
Feature importance analysis from the random forest model identified the variables that contributed most to accurate classification. Cellular atypia on biopsy was the single strongest predictor, consistent with its known histopathological significance as the defining criterion for precancerous lesions.
Age and menopausal status were the next most important features, reflecting the well-established epidemiology of endometrial cancer as predominantly a disease of post-menopausal women. Endometrial thickness on ultrasound, a standard screening parameter, also ranked highly.
Resistance index -- a measure of blood flow pulsatility in uterine arteries -- contributed meaningful predictive value, consistent with evidence that malignant lesions generate abnormal vascular patterns detectable on Doppler ultrasound. BMI contributed as expected given obesity's role as a major endometrial cancer risk factor.
All seven input variables are collected in routine gynecologic consultation -- no additional tests are required. A clinician could input these values into a decision-support interface and receive an estimated probability of malignancy, precancerous change, or benign disease in real time.
High-risk patients flagged by the model could be prioritized for urgent hysteroscopy and biopsy, while low-risk patients might be managed conservatively with surveillance. This risk stratification would reduce the burden on diagnostic services by concentrating investigations on patients most likely to have significant pathology.
The model's accuracy with precancerous lesions is especially important: women with atypical hyperplasia who are identified early can be treated with hormonal therapy or surgery before cancer develops, representing a genuine prevention opportunity rather than just earlier detection.
This study demonstrates that routinely collected clinical variables, when analyzed by a random forest model, can identify endometrial malignancy and precancerous lesions with near-perfect accuracy. The performance advantage over simpler approaches like logistic regression was consistent and meaningful.
External validation in different clinical settings is needed before widespread deployment. A model trained at a single center may reflect local patient characteristics or clinical practices not fully representative of other populations. Multi-center validation studies are the logical next step.
Integrating molecular biomarkers -- such as serum CA125, circulating tumor DNA, or endometrial microbiome profiles -- into future model iterations could push performance further and enable diagnosis from non-invasive tests alone, moving toward a truly preventive approach to endometrial cancer management.