AI improves IVF success rates

Using AI to select human embryos in fertility treatment
13 November 2020

Interview with 

Hadi Shafiee, Harvard University


Human embryos produce by IVF


Infertility is a major problem in developed countries worldwide. Just between 2017 and 2018, in America it fell by 2%. This means that, increasingly, couples are resorting to assisted conception methods, like IVF, to help them to conceive. This usually involves collecting eggs from the woman and sperm from the man, fertilising one with the other, allowing the resulting embryos to develop for a short period in the dish, and then choosing the best looking embryos to put back into the uterus to - hopefully - trigger a pregnancy. But this is a very subjective technique; the success rates are still relatively low at around 30%, and even among those couples who succeed, they usually need to go through several heart-rending and exhausting rounds of this process. So Hadi Shafiee, from Harvard, wondered whether an AI system could help, as he told Chris Smith…

Hadi - One of the most important factors is the ability to select the highest quality embryo and the current methods to do that are either manual, highly subjective, expensive, and time-consuming. So we wanted to tackle this problem by developing a machine learning-based, fully automated approach, to help embryologists to pick the best quality embryo.

Chris - Normally, an embryologist would look down a microscope at a selection of developing embryos, and they would say, "I reckon that one looks like the best prospect, followed by that one". And those are the ones they would cherry-pick to put back into the woman, aren't they, you're saying, can we use computers to help us do it even better?

Hadi - You got it right. Actually, we performed a very simple but very important experiment before this publication that we had at eLife. We recruited several embryologists with different backgrounds, you know, from two years to eight years of experience, and asked them, by providing them like hundreds of images of embryos with known outcomes, basically asked them to perform two simple clinical decisions that they normally do in their practice. And after they made their decision, we just rotated the images and asked them again to perform the same process on the same images. And there was huge variation between the decisions that they made.

Chris - So in other words, looking at the same embryo twice, they didn't make the same decision all the time, every time, showing there's enormous variation when they do it by eye?

Hadi - That is correct. So one application of using AI is to literally help them to have to at least reduce the subjectivity of the process. So basically how the system works is that we provide thousands of embryo images with the known outcomes. And then the machine intuitively kind of figures out what kind of features it had to emphasize in order to make the right decision.

Chris - If you do a trial and compare its performance against what an embryologist could achieve, how much better is the AI system?

Hadi - Very, very good question. So in fact, we did that study by recruiting 15 different embryologists from five different fertility centres across the country. And we provided images of embryos with really high quality. When you look at them morphologically they look amazing and we asked them to make their decision based on its implantation outcome. And we asked the same question from the machine. And we found that the machine actually did a better job, statistically, compared to the embryologist. 75% versus 60-something% on average.

Chris - That's quite a big difference. Do you actually know what it's looking for?

Hadi - At this stage it is literally a black box. We literally don't know what kind of features actually the machine was looking at.

Chris - Does the machine then make any surprising decisions? Does it ever disagree with you where it chooses an embryo and you'd think well in a million years, I'd never put that in?

Hadi - Very good question. Yes. In some cases it does surprise. You know, we literally saw embryos that actually, if you show it to an embryologist, the embryologist wouldn't transfer it. But then the machine picked that and then it ended up with a successful pregnancy. In some of the cases, it actually does surprise you in an interesting way, rather than in a negative way.

Chris - So if this were applied to your average IVF clinic, where rates are currently running at about a 30% success rate, what sort of success rate would this return? All other things being equal?

Hadi - In order to get a very confident answer, you literally need to perform a prospective clinical trial where you have a randomised study. But when you look at the retrospective data we have done, our machines showed that we were doing better even compared to MGH fertility centre, which is literally among the top 10% facility centres across the country.


Add a comment