NEW ORLEANS — An artificial intelligence (AI) model did just as well as embryologists in selecting embryos for transfer, a researcher reported here.
In a retrospective, double-blind, randomized comparative reader study, the clinical pregnancy rate reached 61.0% for embryos selected by a group of five embryologists as compared with 62.1% for embryos selected by a machine learning model (P<0.001 for noninferiority), reported Oleksii O. Barash, PhD, laboratory director of the Reproductive Science Center of the San Francisco Bay Area.
Furthermore, in the 65% of cases when all five embryologists agreed on which embryo to transfer, the AI model selected the same choice 99% of the time, he said during a presentation at the American Society for Reproductive Medicine annual meeting.
In instances when AI came to a different conclusion than the embryologists, researchers analyzed the clinical pregnancy rates of each group. “In four out of five pairs, actually AI did a better job,” Barash noted.
Only one embryologist, who had more than 10 years of experience, did better than the AI model, with clinical pregnancy rates of:
- Embryologist 1: 54.7% vs 57.7% with AI
- Embryologist 2: 46.1% vs 58.8%
- Embryologist 3: 42% vs 66.2%
- Embryologist 4: 66.1% vs 41.9%
- Embryologist 5: 50% vs 56.3%
Barash pointed out that the AI model could potentially help with standardizing embryo selection across laboratories and networks. “The addition of artificial intelligence can potentially address all this inefficiency. And hopefully, one day we won’t have to double guess ourselves and ask, ‘did I choose the right embryo to transfer?'”
Morine Cebert, MSN, PhD, of Yale New Haven Health System in Bridgeport, Connecticut, told MedPage Today that in addition to standardization, this technology could be beneficial in terms of patient access.
“It’s just interesting the ways that we are helping to improve the quality of care using machine learning, especially in fertility where the goal is to actually do something that we don’t have a 100% success rate on,” she said.
“This is kind of like the next phase of the workforce,” she added, noting that a big question will be what jobs will be lost to AI. Still, she said she’s hopeful “in terms of the quality of care and how can this help certain groups who might not have access.”
Before the study, a machine learning model was developed that predicts the chance of clinical pregnancy by considering embryo morphology grades, determined by Gardner classification and day of development. The model itself was trained on 12,626 single-blastocyst transfer cycles from U.S. clinics. Barash noted that the model accounts for differences in success rates between sites and repeat patients and that the expected calibration error is 0.031. This algorithm is part of the software Embryo Assist, which was designed by Alife Health.
For the current study, researchers used data from 438 transferred embryos with known outcomes from a total of 10 clinics. Using this data, he created 1,257 simulated patient panels which had two to five embryos each. Each panel was matched by age, race, and preimplantation genetic testing status. Of the five embryologists who were human comparators to the AI, two had 1 to 3 years of experience, and three had more than 10 years of experience.
Barash is a scientific advisory board member at Alife Health.
Cebert has no conflicts of interest.
American Society for Reproductive Medicine
Source Reference: Barash O, et al “Clinical evaluation of a machine learning model for embryo selection: a double-blinded randomized comparative reader study” ASRM 2023; Abstract 0-1.