Could an artificial intelligence (AI) algorithm predict fetal heartbeat from images of vitrified-warmed embryos? Applying AI to vitrified-warmed blastocysts may help predict which ones will result in implantation failure early enough to thaw another. The application of AI in the field of embryology has already proven effective in assessing the quality of fresh embryos. Therefore, it could also be useful to predict the outcome of frozen embryo transfers, some of which do not recover their pre-vitrification volume, collapse, or degenerate after warming without prior evidence. This retrospective cohort study included 1109 embryos from 792 patients. Of these, 568 were vitrified blastocysts cultured in time-lapse systems in the period between warming and transfer, from February 2022 to July 2023. The other 541 were fresh-transferred blastocysts serving as controls. Four types of time-lapse images were collected: last frame of development of 541 fresh-transferred blastocysts (FTi), last frame of 467 blastocysts to be vitrified (PVi), first frame post-warming of 568 vitrified embryos (PW1i), and last frame post-warming of 568 vitrified embryos (PW2i). After providing the images to the AI algorithm, the returned scores were compared with the conventional morphology and fetal heartbeat outcomes of the transferred embryos (n = 1098). The contribution of the AI score to fetal heartbeat was analyzed by multivariate logistic regression in different patient populations, and the predictive ability of the models was measured by calculating the area under the receiver-operating characteristic curve (ROC-AUC). Fetal heartbeat rate was related to AI score from FTi (P < 0.001), PW1i (P < 0.05), and PW2i (P < 0.001) images. The contribution of AI score to fetal heartbeat was significant in the oocyte donation program for PW2i (odds ratio (OR)=1.13; 95% CI [1.04-1.23]; P < 0.01), and in cycles with autologous oocytes for PW1i (OR = 1.18; 95% CI [1.01-1.38]; P < 0.05) and PW2i (OR = 1.15; 95% CI [1.02-1.30]; P < 0.05), but was not significantly associated with fetal heartbeat in genetically analyzed embryos. AI scores from the four groups of images varied according to morphological category (P < 0.001). The PW2i score differed in collapsed, non-re-expanded, or non-viable embryos compared to normal/viable embryos (P < 0.001). The predictability of the AI score was optimal at a post-warming incubation time of 3.3-4 h (AUC = 0.673). The algorithm was designed to assess fresh embryos prior to vitrification, but not thawed ones, so this study should be considered an external trial. The application of predictive software in the management of frozen embryo transfers may be a useful tool for embryologists, reducing the cancellation rates of cycles in which the blastocyst does not recover from vitrification. Specifically, the algorithm tested in this research could be used to evaluate thawed embryos both in clinics with time-lapse systems and in those with conventional incubators only, as just a single photo is required. This study was supported by the Regional Ministry of Innovation, Universities, Science and Digital Society of the Valencian Community (CIACIF/2021/019) and by Instituto de Salud Carlos III (PI21/00283), and co-funded by European Union (ERDF, 'A way to make Europe'). M.M. received personal fees in the last 5 years as honoraria for lectures from Merck, Vitrolife, MSD, Ferring, AIVF, Theramex, Gedeon Richter, Genea Biomedx, and Life Whisperer. There are no other competing interests. N/A.
Read full abstract