Development and Validation of an Explainable Machine-Learning Model to Predict Cumulative Live Birth in PCOS-Associated Infertility

Abstract

Introduction: Polycystic ovary syndrome (PCOS) is the leading cause of anovulatory infertility, yet predicting the cumulative live birth rate (CLBR) after in vitro fertilization (IVF) is difficult because of marked clinical heterogeneity. Conventional linear models discriminate modestly and most machine-learning (ML) tools remain uninterpretable. We aimed to develop and internally validate an explainable ML model for CLBR in women with PCOS.

Methods: In a multicenter retrospective cohort at two tertiary reproductive-medicine centers in Indonesia (January 2018–December 2023), 1,245 women with PCOS (Rotterdam criteria) undergoing a first IVF/intracytoplasmic sperm injection cycle were analysed. Five algorithms (logistic regression, support vector machine, random forest, gradient boosting and eXtreme Gradient Boosting [XGBoost]) were trained (70%) and internally validated (30%) with 5-fold cross-validation. Discrimination used the area under the ROC curve (AUC) with DeLong confidence intervals (CI); SHapley Additive exPlanations (SHAP) quantified feature importance. Multivariable logistic regression, calibration and number-needed-to-treat (NNT) were also derived.

Results: Overall CLBR was 54.6% (95% CI 51.8–57.4%). XGBoost performed best (AUC 0.82, 95% CI 0.79–0.85; accuracy 76.4%; sensitivity 78.2%; specificity 74.1%) and significantly exceeded logistic regression (AUC 0.72; ΔAUC 0.10, p<0.001). SHAP ranked top-quality embryo number, maternal age, anti-Müllerian hormone and oocyte yield as dominant predictors. Each additional top-quality embryo raised CLBR odds (adjusted OR 1.85, 95% CI 1.66–2.06, p<0.001); ≥3 versus <3 embryos yielded an NNT of 3.3.

Conclusion: An explainable XGBoost model accurately predicts CLBR in PCOS-associated infertility and can support individualised counselling and cycle-management decisions. Prospective external validation is warranted before clinical deployment.