AI result calls into question genetic testing of embryos during IVF

This surprising and confusing AI result calls into question the “Gold Standard” PGT-A genetic testing of embryos in IVF.

PGT-A genetic testing is conducted to assess whether an embryo is normal (euploid) or aneuploid (abnormal). It is accepted in the industry that normal euploid embryos improves pregnancy and live birth success rates. The PGT-A test itself is expensive and invasive – requiring a biopsy (cutting off a piece) of the embryo which could eventually become a baby.

AI was trained to predict the genetic outcome just using images (morphology). Completely non-invasive and instant. When compared to outcomes using PGT-A, the AI appears to be less “accurate”. 80% accuracy when AI is very confident that an embryo is normal – ‘according to PGT-A’.

However, in recent study reported at the Pacific Coast Reproductive Society meeting this year (PCRS 2024), Indian Wells, CA, the AI demonstrated additional predictive power for pregnancy and live birth on embryos that were already classified as normal (euploid) by PGT-A: https://www.fertstert.org/article/S0015-0282(24)00433-3/fulltext. How is this possible for the “less accurate” AI?

When the AI was initially trained on embryos images with PGT-A outcomes, an algorithm (UDC) was run on the data to assess whether there were any likely errors in the data. What we found surprised us: over 20% of the embryos that had a PGT-A outcome of abnormal, the algorithm was convinced were mislabelled and were actually normal. This led us to conduct a literature search which provided further evidence to support our observation: (1) normal embryos could be PGT-A tested as abnormal, and thus unnecessarily discarded (groups citing up to 40% variability in some cases, particularly for embryos originally classified as abnormal); (2) ‘abnormal’ embryos (according to PGT-A) that were used have led to healthy live births; and (3) some randomized controlled trials have failed to show improved clinical outcomes with the use of PGT-A.

To remove doubt, we excluded these potentially “erroneous” datapoints in the training dataset to create the AI to predict genetic outcomes. As a small test to what we observed above, we ran the AI on a very small (not statistically significant) sample of 16 embryos that were PGT-A tested and then re-tested a second time. About 20% of the embryos switched classification! The second PGT-A test was in greater agreement with the AI assessment. Further details in our published paper here: https://doi.org/10.1093/humrep/deac131.

Unfortunately, over the last couple of years IVF clinics globally have been reluctant to conduct a larger PGT-A re-test study so we can better understand the true accuracy of the AI.

Fortunately, regardless of the true AI accuracy for detecting the genetic integrity of the embryo, we have shown that the use of artificial intelligence, which is non-invasive, instant and low-cost, has additional benefits over and above PGT-A tested embryos and can lead to improved pregnancy and live births outcomes. And this is all that matters to IVF patients.

Thanks Tex, USF and Ovation, who have collaborated with us for years on developing this great technology, and are co-authors of this great poster at PCRS.