Development and External Validation of Machine Learning Model to Predict Live Birth Following Assisted Reproductive Technology in Women with Ovarian Endometriomas: A Decision-Support Tool

Yifei Sun,1– 7,* Zijing Wang,1– 7,* Jiayi Zhou,1– 7 Linlin Cui,1– 8 Huidan Wang1– 7 1State Key Laboratory of Reproductive Medicine and Offspring Health, Center for Reproductive Medicine, Institute of Women, Children and Reproductive Health, Shandong University, Jinan, Shandong, 250012, People’s Republic of China; 2National Research Center for Assisted Reproductive Technology and Reproductive Genetics, Shandong University, Jinan, Shandong, 250012, People’s Republic of China; 3Key Laboratory of Reproductive Endocrinology, Shandong University, Ministry of Education, Jinan, Shandong, 250012, People’s Republic of China; 4Shandong Technology Innovation Center for Reproductive Health, Jinan, Shandong, 250012, People’s Republic of China; 5Shandong Provincial Clinical Research Center for Reproductive Health, Jinan, Shandong, 250012, People’s Republic of China; 6Shandong Key Laboratory of Reproductive Research and Birth Defect Prevention, Jinan, Shandong, 250012, People’s Republic of China; 7Research Unit of Gametogenesis and Health of ART-Offspring, Chinese Academy of Medical Sciences (No. 2021RU001), Jinan, Shandong, 250012, People’s Republic of China; 8Center for Reproductive Medicine, The Second Qilu Hospital of Shandong University, Jinan, Shandong, 250012, People’s Republic of China*These authors contributed equally to this workCorrespondence: Huidan Wang, Email [email protected]: Ovarian endometriomas damage the ovarian structure, alter ovarian inflammation, and impair ovarian reserve. Given the conflicting results, determining an optimal reproductive strategy for women with endometriomas―whether expectant management, medication, surgery, or assisted reproductive technology (ART)―remains challenging.Objective: This study aims to preliminarily develop and validate clinically applicable decision-support tools by training, testing, and validating an automated machine learning (ML) model to predict the likelihood of live birth following ART in women with endometriomas.Methods: The derivation and testing cohort included 1705 women, and the external validation cohort included 1475 women with ovarian endometriomas following ART retrospectively. Two ML models were developed and validated to predict the probability of live birth. Model performance was evaluated using the area under the curve (AUC), accuracy, sensitivity, specificity, F1 score, and Brier score. The SHapley Additive exPlanations (SHAP) method was employed to interpret feature importance.Results: Comparing seven ML algorithms, the Extreme Gradient Boosting (XGBoost) demonstrated superior predictive performance both in model-1 and model-2, achieving an AUC of 0.90 [95% confidence interval (CI): 0.88– 0.92] and 0.88 (95% CI: 0.86– 0.89) in test-datasets and 0.80 (95% CI: 0.76– 0.83) and 0.69 (95% CI: 0.65– 0.73) in external validation cohort. The SHAP analysis revealed that the age and features associated with ovarian reserve had strong predictive power and the ovarian endometriomas had limited predictive power.Conclusion: Model-2, which uses only pre-ART variables, can support reproductive strategy selection prior to ART initiation. Conversely, Model-1 is designed to support embryo transfer strategy option after oocyte retrieval, incorporating post-ART data. Although both models show promise as decision-support tools for personalizing infertility treatment in women with endometriomas, their clinical implementation awaits confirmation from prospective, multicenter validation.Keywords: machine learning model, assisted reproductive technology, ovarian endometriomas, live birth, SHAP interpretation