P-267 Artificial intelligence (AI) can successfully predict pregnancy despite of the visual differences of fresh and frozen-thawed embryos

Abstract

Abstract Study question Can AI accurately predict pregnancy from fresh and frozen-thawed embryo images that may exhibit visual differences? Summary answer AI was able to distinguish between fresh and frozen images based on visual differences, but these differences did not affect the accuracy of pregnancy prediction. What is known already Frozen-thawed embryos may experience changes to their structure and composition, though these changes are difficult to detect with the human eye. AI has been shown to predict pregnancy by analyzing embryo images in IVF cycles. However, there has been no evidence to suggest that the morphologic changes caused by cryopreservation affect AI's pregnancy predictions. In this study, we developed an AI model to distinguish between images of fresh and frozen-thawed embryos and evaluated if the visual differences affected pregnancy prediction. Study design, size, duration We performed a retrospective study of single static images of 2,237 Day 5 blastocysts from two in vitro fertilization (IVF) clinics between February 2001 and December 2021. The images were collected from standard optical light microscopes and matched with metadata such as pregnancy outcomes, cryopreservation and assisted hatching information. We defined a positive pregnancy indication as the presence of a gestational sac (G-SAC). Participants/materials, setting, methods We constructed two CNN models to verify two hypotheses. The first model was a classification model that utilized day 5 images of fresh and frozen-thawed embryos from two IVF clinics. The proportion of fresh to frozen-thawed at each clinic was 603 to 667 and 402 to 565, respectively. The second model was a CNN designed to predict pregnancy and its performance was compared after incorporating the cryopreservation label through internal validation. Main results and the role of chance The first AI model classified frozen-thawed and fresh embryos with high AUROCs (0.848 and 0.912) and accuracy (0.846 and 0.912) at each clinic. The Grad-CAM images revealed that the AI learned to differentiate the two types of images based on features mainly in the embryo's zona pellucida region. The AUROCs of the second AI model for pregnancy prediction were 0.730 and 0.663 at each clinic, respectively. Adding the cryopreservation label did not significantly change the AUROCs, which remained at 0.734 and 0.650. The study found that the visual differences between fresh and frozen-thawed embryos had no effect on the performance of the pregnancy prediction model. It's important to note that assisted hatching was used in 88% of the frozen-thawed cycles in the study, which may have compensated for the changes to the zona pellucida caused by cryopreservation. Limitations, reasons for caution This study was validated using data from two IVF clinics, and a larger dataset from multiple centers is needed for external validation. Further research on non-assisted hatching (AH) cycles is necessary to confirm the role of AH in frozen-thawed cycles. Wider implications of the findings The study found that AI can accurately differentiate between fresh and frozen-thawed embryos by analyzing the zona pellucida region. The visual differences between the two types of embryos did not impact the accuracy of pregnancy prediction. Assisted hatching may also mitigate any negative effects on frozen-thawed embryos. Trial registration number not applicable