Embryo classification beyond pregnancy: early prediction of first trimester miscarriage using machine learning

Abstract

Purpose: First trimester miscarriage is a major concern in IVF-ET treatments, accounting for one out of nine clinical pregnancies and for up to one out of three recognized pregnancies. To develop a machine learning classifier for predicting the risk of cleavage-stage embryos to undergo first trimester miscarriage based on time-lapse images of preimplantation development. Methods: Retrospective study of a 4-year multi-center cohort of 391 women undergoing intra-cytoplasmatic sperm injection (ICSI) and fresh single or double embryo transfers. The study included embryos with positive indication of clinical implantation based on gestational sac visualization either with first trimester miscarriage or live-birth outcome. Miscarriage was determined based on negative fetal heartbeat indication during the first trimester. Data were recorded and obtained in hospital setting and research was performed in university setting. Results: A minimal subset of six non-redundant morphodynamic features were screened that maintained high prediction capacity. Features that account for the distribution of the nucleolus precursor bodies within the small pronucleus and pronuclei dynamics were highly predictive of miscarriage outcome as evaluated using the SHapley Additive exPlanations (SHAP) methodology. Using this feature subset, XGBoost and random forest models were trained following a 100-fold Monte-Carlo cross validation scheme. Miscarriage was predicted with AUC 0.68 to 0.69. Conclusion: We report the development of a decision-support tool for identifying the embryos with high risk of miscarriage. Prioritizing embryos for transfer based on their predicted risk of miscarriage in combination with their predicted implantation potential is expected to improve live-birth rates and shorten time-to-pregnancy.