P-199 Ensuring no clinical risk: a cohort analysis on the agreement between the embryo selected by the embryologist and the embryo selected by artificial intelligence

Abstract Study question Is there any clinical risk in selecting the embryo to be transferred using a deep learning-based model? Summary answer Most embryos selected by senior embryologists match automatically best scoring embryo, and outcomes are better when this coincidence occurs. What is known already In vitro fertilization (IVF) techniques have changed over time with aim of improving results. Today, embryology facing change common most areas medicine, introduction automation. The use automated systems IVF laboratory happening, for example, electronic witnessing ranking according their implantation potential. Relying on type system not easy operators, as they must ensure that treatments will damaged. design, size, duration This single-center cohort study including 5,411 patients who underwent treatments. Their were cultured EmbryoScope® time-lapse (Vitrolife, Denmark) routinely evaluated ASEBIR morphological criteria. Then, scored iDAScore v2 algorithm. Participants/materials, setting, methods A total 7,178 transfers analyzed. was retrospectively learning algorithm from 1 9.9. Finally, we performed analysis agreement between embryologist v2. relative (or incidence implantation) (RR), reduction (RRR) absolute (ARR) calculated. Main results role chance general, considering all (fresh frozen) outcome (n = 7,178), rate higher matched highest-scoring (58.37% vs. 55.49% p 0.014). Embryos single transfer fresh cycle 2,783) top-scoring 63.50% time. An ongoing pregnancy achieved 57.96% patients. Out 36.54% had an did maximum iDAScore, 553 frozen embryo. group, first devitrified coincided highest 44.67% 247). 62.34% However, correspond 55.27%. RR showed each patient 1.13 times more likely become pregnant if score. RRR stated coincidence, probability increased 12.79%. ARR (0.07) suggested every 100 match, 7 would implant than without match. Limitations, reasons caution limited its retrospective nature. Furthermore, design should considered generalizing results, although our clinic involved development specific model. Wider implications findings high embryologist's decision artificial intelligence's comfort assisted reproduction professionals also justifies realistic models based intelligence perform selection procedure good experienced embryologist. Trial registration number Not applicable