Perioperative prediction of adnexal malignancy by an interpretable machine learning model for guiding ovarian preservation in premenopausal endometrial cancer.

BACKGROUND

Ovarian preservation in premenopausal patients with endometrial cancer remains challenging due to the potential presence of concurrent adnexal malignancy. To support surgical decision-making, we developed an interpretable machine learning model for the perioperative identification of high-risk patients, thereby facilitating personalized ovarian preservation strategies.

METHODS

We conducted a retrospective analysis of endometrial cancer patients treated at our institution between 2010 and 2024. After feature selection via multicollinearity analysis and LASSO regression, eight machine learning algorithms were trained to predict coexisting adnexal malignancy. Model performance was evaluated using ROC analysis, accuracy metrics, and Brier score calibration. SHapley Additive exPlanations (SHAP) were applied to interpret the contribution of key features in the optimal model.

RESULTS

Among 296 included patients, 29 (9.8%) had coexisting adnexal malignancy. Sixteen predictive features were selected from clinical, imaging, serum biomarker, and histopathological domains. The Naive Bayes classifier achieved superior performance with an AUC of 0.92 (95% CI: 0.86-0.97), accuracy of 91.0%, and well-calibrated predictions (Brier score: 0.11). The SHAP further elucidated the contribution of each variable to the model's predictions, emphasizing the importance of factors such as Cancer Antigen 125, Estrogen Receptor status, Human Epididymis Protein 4.

CONCLUSION

The Naive Bayes model exhibits high discriminative accuracy for the perioperative prediction of concurrent adnexal malignancy. This decision support tool shows strong potential for clinical application, promoting individualized surgical management and aiding in ovarian preservation for premenopausal endometrial cancer patients.