Machine learning prediction of live birth after IVF using the morphological uterus sonographic assessment group features of adenomyosis

Predicting live birth after the first IVF/ICSI treatment is challenging, as many factors may interact to affect IVF/ICSI outcomes. Adenomyosis is one factor that impacts live birth rates. Machine learning algorithms have been shown valuable for detecting complex dependencies and predicting outcomes in different clinical settings. We aimed to develop a prediction model for live birth after IVF/ICSI treatment, using the Extreme Gradient Boosting (XGBoost) algoritm and incorporating the revised Morphological Uterus Sonographic Assessment (MUSA) group features of adenomyosis. We used a machine learning model based on data from 1037 women undergoing their first IVF/ICSI treatment between January 2019 and October 2022. The importance of each variable on the model was illustrated with the Shapley additive explanations algorithm (SHAP) variable importance. The prediction model was presented with the area under receiver operating characteristics curve (ROC). The proposed XGBoost model had a test AUC of 0.66 and accuracy of 0.59. S-AMH was the best variable for predicting live birth with a mean SHAP of 0.21, followed by a regular junctional zone as the best ultrasonographic variable, mean SHAP 0.13. The predictive ability of MUSA features in relation to live birth was limited. Additional variables should be included in future prediction models.