P-208 Embryo assessment at the click of a button is now possible: evaluation of a deep-learning algorithm integrated directly with the time-lapse platform

Abstract

Abstract Study question Is the embryo scoring function based on deep learning of specific time-lapse systems clinically useful for classifying human blastocyst? Summary answer Blastocyst grading according to iDAScore® is directly associated with conventional morphology and implantation potential, at least in treatments without preimplantation genetic testing for aneuploidy (PGT-A). What is known already The conventional approach of embryo evaluation in the time-lapse systems is for embryologists to manually annotate a number of morphological and/or morphokinetic parameters. These values are then used in models of clinical outcome prediction. Embryo selection can be automated by using artificial intelligence (AI) to predict morphokinetic and morphology parameters. In this case, AI is employed in an indirect way to optimize the daily workflow based on existing traditional parameters. Finally, the most innovative approach is the use of AI to directly predict pregnancy, implantation or even live birth by using only time-lapse images. Study design, size, duration A retrospective cohort study including 518 patients who underwent IVF treatments and whose embryos (n = 3,406) were cultured in EmbryoScope Plus® time-lapse systems. Blastocysts were routinely evaluated by senior embryologists according to the ASEBIR morphological criteria. Then, embryos were scored using the iDAScore algorithm whose values range from 1 to 9.9. Participants/materials, setting, methods Embryo evaluation was performed automatically by iDAScore with the use of deep learning and a neural network. The algorithm was developed considering the entire embryo development to rank embryos according to likelihood of implantation. Embryo score was compared with conventional morphological quality, euploidy rate and the subsequent implantation outcome of 567 single blastocyst transfers. Then, we quantified the contribution of the automatic embryo score to implantation with multivariate logistic regression analysis in different patient populations. Main results and the role of chance The comparison between the embryo score provided by the iDAScore and the morphological category (A, B, C or D) assigned by embryologists showed a direct association*. The mean and standard deviation was 9.2 ± 0.4 for A; 8.2 ± 1.2 for B; 6.9 ± 1.6 for C and 4.0 ± 1.8 for D. The euploidy rate increased when embryos showed higher automatic scores*: 45.9% for score ≤ 8.0 (n = 354), 55.6% for score 8.1-8.8 (n = 169) and 62.8% for score >8.8 (n = 180). The implantation rate increased as the embryo score improved*: 37.8% for score ≤ 7.8 (n = 127), 50.9% for score 7.9-8.9 (n = 163), 65.7% for score >8.9 (n = 277). The logistic regression analysis of iDAScore took into account possible confounding factors: oocyte origin (donated vs. autologous); type of embryo transfer (fresh vs. frozen); oocyte age; patient body mass index; PGT-A (tested vs. non-tested embryos) and day of embryo transfer (fifth vs. sixth day of embryo development). iDAScore value was related to the odds of implantation in the oocyte donation program (OR = 1.61; 95%CI [1.19-2.19]; p < 0.001; n = 265) and in conventional treatments with autologous oocytes (OR = 1.52; 95%CI [1.22-1.90]; p < 0.001; n = 192). There was no significant association of embryo score with implantation in treatments involving PGT-A (n = 110). *p<.05 Limitations, reasons for caution This study is limited by its retrospective nature. The single-center design should be taken into account when considering the universal application of the model. Although our clinic was not involved in the development of iDAScore, and therefore this study should be considered as an external validation. Wider implications of the findings Our findings confirm that embryo evaluation can be performed automatically allowing embryologists to utilize their time more efficiently in other tasks of the IVF process that are more complex and require high precision and attention. Trial registration number This research has been funded by a grant from The Ministry of Science, Innovation and Universities FIS (PI21/00283) awarded to M.M.