We identified a conceptual discrepancy between common machine learning-based and clinical IVF embryo selection and demonstrated that this discrepancy can lead to major errors in preference of obviously visually defective embryos over high-quality successfully implanted embryos. Our results establish the benefit of harnessing the extensive data and reliable labels in discarded embryos to improve ML-based embryo selection and provide two practical recommendations: (1) exclusive focus on embryo intrinsic features for ranking, (2) reduce negative label ambiguity. These guidelines will shift machine learning-based IVF toward optimizing embryo ranking and implantation prediction independently - an approach better reflecting the clinician’s decision that involves the evaluation of all the embryos in the context of its cohort.
This is the lab's first endeavor to machine learning in the domain of in vitro fertilization.
Project conceived and led by Itay Erlich, who recently submitted his Ph.D. dissertation!
Twitter summary: https://twitter.com/AssafZaritsky/status/1446181283776942084
