Viable Embryos in IVF Do Not Always Lead to a Pregnancy: Identifying Misclassified Medical Images of Embryos Using Artificial Intelligence

A recent study by AI Healthcare company Presagen, in collaboration with US-based IVF clinic group Ovation Fertility, has shown that viable (good quality) embryos do not always lead to a pregnancy due to patient factors (e.g. Endometriosis). A unique AI algorithm developed by Presagen called UDC was able to detect when medical images of embryos classified as ‘non-viable’ because they did not lead to a pregnancy, were actually viable and did not achieve a pregnancy due to patient factors beyond the embryo. These medical images identified by the UDC are therefore misclassified and can be considered as inherent errors in the large medical dataset. Training AI on data with errors can negatively impact AI performance. It was shown that removing the UDC-identified data errors resulted in an increase in performance of the AI trained to assess embryo viability.

The study was presented at the American Society for Reproductive Medicine (ASRM) 2020 Scientific Congress, October 2020. The novel technique underpins Presagen’s Life Whisperer product, which helps IVF clinics select the most viable embryos to assist improve pregnancy outcomes for IVF patients. Life Whisperer is currently being used by IVF clinics globally. Life Whisperer has been shown to perform 25% better than manual visual assessment alone and can reduce time-to-pregnancy by 14%, saving patients’ money and increasing revenue for clinics.

The study also calls into question literature that claim to have trained AI to assess embryo viability with accuracies over 90%. With inherent errors well over 10%, accuracies of over 90% can only be achieved if these types of inherent data errors are adequately controlled for.

The poster presentation can be downloaded below.