Predictive models for live birth outcomes of FET: Improving clinical decision-making based on machine learning analysis

Abstract Purpose: This study used multiple machine learning algorithms to predict live births from frozen embryo transfers (FET) based on patient demographics, laboratory test results, and parameters associated with the FET cycle. Methods: Data from 33,915 cycles of frozen-thaw embryo transfer performed at Chengdu Xinan Gynecological Hospital between January 2015 and December 2021 were used. The dataset was randomly divided into a training set (70%) and a test set (30%). Features were ranked for importance based on the random forest model, and features with the top 25 contribution values were used to develop logistic regression models, random forest models, support vector machine models, and XGBoost models. Shapley was used to interpret the results of the best-performing models. Receiver operating characteristic curves (AUC) under area and calibration curves were to be assessed for the performance of machine learning prediction models. Results: Ranking the importance of features based on the stable random forest algorithm showed that the most predictive features included AMH, Basal PRL, Basal T, Basal FSH, etc. The XGBoost model had the highest AUC (0.750, 95% CI 0.746-0.755). The XGBoost-based SHAP summary plot indicated that patients with lower age, shorter years of infertility, and D5 embryo type for transfer had a greater likelihood of live birth outcome after freeze-thaw embryo transfer. Conclusion: The XGBoost model performed best in predicting the outcome of freeze-thaw embryo transfer. The algorithm combined with the interpretability of SHAP summary plot can assist clinicians in the decision-making process of freeze-thaw embryo transfer.