P-137 A deep learning model using time-lapse imaging data to predict successful clinical pregnancy and embryo euploidy

Abstract Study question Are the deep learning models helpful in selecting embryos with a high probability of clinical pregnancy and euploidy from time-lapse imaging during embryo growth? Summary answer Our deep learning models using time-lapse imaging data predicted the probabilities of successful clinical pregnancy and euploidy of embryos with high accuracy. What is known already Recent studies have attempted to select the most viable embryo using deep learning of time-lapse imaging during incubation. However, most studies included embryos that were not transferred due to low evaluation by the embryologist as embryos with unsuccessful pregnancies, or they mixed single and multiple embryo transfer cases in the analyses. Some studies’ algorithms rely on humans’ subjective interpretation of morphological and morphokinetic features, leading to embryologist-to-embryologist and within-embryologist variabilities. Some reports, including the study predicting embryo aneuploidy, used a snapshot image for prediction. These approaches have restricted the potential of deep learning in embryo selections. Study design, size, duration Retrospective analyses of time-lapse videos of 2,192 embryos with clinical pregnancy outcomes, defined as confirmed gestational sacs in utero, and 1,467 embryos with karyotype outcomes examined by preimplantation genetic testing for aneuploidy (PGT-A). The data were obtained from two different in-vitro fertilization clinics between November 2017 and December 2021. All transfers were performed as a single embryo transfer. Participants/materials, setting, methods The deep learning model was trained using time-lapse videos with known clinical pregnancy outcomes for predicting the probability of clinical pregnancy with given time-lapse video sequences and patient ages (CP model). Another deep learning model was trained using time-lapse videos with known karyotype outcomes by PGT-A for predicting the probability of euploidy (AP model). The models’ predictive powers were measured using the average area under the curve (AUC) of the receiver operating characteristic curve. Main results and the role of chance The CP model predicted clinical pregnancy with an AUC of 0.693. The AP model predicted embryo aneuploidy with an AUC of 0.702. Limitations, reasons for caution The CP model is based on the embryos that were transferred because the embryologist judged them as having a high probability of pregnancy. Clinical pregnancy may be associated with maternal factors besides embryo factors. Wider implications of the findings Our two deep learning models, the CP and AP models, will be helpful for selecting the embryos to be transferred. The CP model may increase pregnancy rates in patients undergoing assisted reproductive technology. The AP model may reduce the miscarriage caused by embryo aneuploidy. Trial registration number not applicable