P-140 Does assessment of time-lapse videos result in improved prediction of clinical pregnancy compared to assessment of single images of blastocysts?

Abstract

Listen

Translate article

Abstract Study question Is there a difference in performance between artificial intelligence(AI) algorithms that predict the likelihood of pregnancy from a single blastocyst image compared to time-lapse videos? Summary answer The performance of AI to predict pregnancy from a single blastocyst image was similar to the performance of an AI using time-lapse video. What is known already Vitrolife’s “iDAScore (iDA)” and Presagen’s Life Whisperer Viability (LW) are AI systems trained on videos (iDA) or images (LW) of blastocysts, with corresponding pregnancy outcomes (fetal heartbeat). iDA requires a dedicated time-lapse incubator and uses multiple images to predict pregnancy. LW requires only one image taken with a camera that meets image quality requirements. However, as tools for selecting viable embryos for transfer, these methods adopt different approaches: iDA relying on multiple morphokinetic timepoints to achieve prediction based on past embryo development, and LW relying on predictive analytics from morphological assessment at the decision point prior to transfer. Study design, size, duration 359 embryos cultured in EmbryoScope +/8 that reached blastocyst stage (Gardner classification ≥1) on Day 5 were utilised for first single vitrified blastocyst transfer (between 2018-2021). Pregnancy prediction was scored from EmbryoScope’s embryo images using iDA and LW, and a dataset was created with fetal heartbeat as the outcome. Participants/materials, setting, methods iDA data was classified into quartiles (Low, Medium, High, Very High) and LW into the manufacturers’ recommended four categories (Low, Medium, High, Very High). The trend of pregnancy outcome and respective scores were evaluated using a trend test (Cochran-Armitage test). Additionally, the Quadratic Weighted Kappa coefficient was used to evaluate the concordance of the categories. The Area Under the Roc Curve (AUC) for each prediction accuracy was also determined and compared using Delong’s test. Main results and the role of chance A trend of increasing pregnancy rate was observed for both methods as the score category increased (iDA: Low 35.2%, Medium 46.9%, High 52.1%, Very High 65.4%; LW: Low 36.0%, Medium 46.3%, High 60.5%, Very High 63.4%). The respective AUCs were iDA: 0.635 LW: 0.627, (not significantly different; P = 0.732). The Quadratic Weighted Kappa coefficient was 0.622, indicating that the categories of evaluation were consistent. Limitations, reasons for caution This is a retrospective study performed in a single clinic. In addition, the study was limited to transfer cycles of embryos where blastocyst formation was reached on day 5 of culture. Wider implications of the findings These results indicate that an AI that predicts pregnancy from a single blastocyst image can perform equally well compared to AI that predicts pregnancy from time-lapse videos. This suggests that tracking of morphokinetic timepoints during embryo development may be less important than end-point morphology for predicting pregnancy at blastocyst stage. Trial registration number not applicable