Abstract Study question Can an objectively trained AI system complement embryologists in their ranking of embryos and minimize time to pregnancy, when only poor-quality embryos are available? Summary answer An objective AI algorithm could have decreased cycle to pregnancy by 18% in cohorts of embryos where only poor-quality embryos were available. What is known already Morphokinetics is a key criterion for embryologists to rank embryos as good, fair, or poor quality, with the latter frequently being discarded. Although poor-quality embryos can still lead to a pregnancy, selection of a viable embryo remains challenging to embryologists due to their depleted morphological characteristics. Unlike embryologists, an AI can be trained to see beyond embryo morphological quality, on data that includes many poor-quality embryos that resulted in a pregnancy. As such, AI has the potential to recognize viable poor-quality embryos, which becomes especially crucial in cases where good or fair embryos are unavailable from the embryo pool. Study design, size, duration Data from a French center was acquired through a MIRI® (ESCO) time-lapse system and uploaded on EMBRYOLY, between January 2022 and March 2023. For a total of 204 patients (women’s average age 35 y.o.), embryologists were faced with having to choose an embryo to transfer among 4.3 ± 1.7 poor-quality embryos either because the cohort was composed of only poor-quality embryos or because higher quality embryos had already been transferred. Participants/materials, setting, methods A previously described deep learning algorithm trained on videos of developing embryos with known pregnancy outcome was used to score embryos according to their chances of leading to a clinical pregnancy (fetal heartbeat between 6-9 weeks). A 2D ResNet model was trained, validated and tested on 8138 images of day 5 embryos annotated by senior embryologists to identify those of poor-quality (Gardner grade ICM and/or TE = C, corresponding to 44.7% of the dataset). Main results and the role of chance The association between video predictions and pregnancy was assessed with a logistic regression. A McNemar test was used to compare cycle to pregnancy (CTP) with and without EMBRYOLY’s hypothetical impact. EMBRYOLY’s AI was able to rank poor-quality embryos, as illustrated by a 2-7% relative increase in likelihood of clinical pregnancy observed with an increase in the video score (logistic regression with unit level +0.02, OR = 1.04, 95CI = [1.02-1.07], N = 261 transferred embryos, p < 0.05). EMBRYOLY’s first choice was concordant with the embryologists’ first choice in 74% of the cases that led to a pregnancy (N = 42 embryos). For patients who had two poor-quality embryos transferred with at least one pregnancy (N = 14 cycles), the results show that if the embryologist had first transferred the highest EMBRYOLY-ranked embryo, the number of cycles to pregnancy would have been reduced by 18% (from 1.7 to 1.4, McNemar, p < 0.05), corresponding to a median of 2.3 months. Limitations, reasons for caution The endpoint used was clinical pregnancy as measured by fetal heartbeat, and does not indicate the probability of live birth. The current investigation was performed with retrospective data from a single center. In the future, a multi-centre evaluation should be performed to generalize our findings. Wider implications of the findings Our findings support that an objectively trained AI can help embryologists not miss out on poor-quality embryos that can lead to a pregnancy, which is all the more important to target when good or fair embryos are unavailable from the embryo pool. Trial registration number not applicable
Read full abstract