IVF can be a painstaking process. Could AI make it more precise?

Artificial intelligence is gaining new ground in nearly every corner of the clinic — including, increasingly, in vitro fertilization.

IVF has helped millions of families conceive. But it’s also expensive, often emotionally and physically grueling, and requires many patients to go through numerous rounds to get pregnant. Researchers are looking to see if AI can more precisely pick the most viable embryo to implant — and in turn, improve success rates, reduce the risk of pregnancy loss, and bring IVF costs down.

“There is room for AI to push the science and go beyond what is already known,” said Hadi Shafiee, a researcher at Brigham and Women’s Hospital in Boston who studies clinical applications of AI.


But experts say there are still ethical issues that need to be grappled with, including how such tools should be used and what happens when a model’s prediction is wrong. There’s also a looming question about whether there’s enough evidence to be rolling the technology out in the clinic just yet. Even so, a smattering of clinics around the globe have already started adding AI tools into the process.

An IVF cycle typically takes three weeks and, in the U.S., can cost between $12,000 and $25,000. It starts with stimulating the ovaries with hormones, followed by retrieving eggs and fertilizing them with sperm in a lab before transferring the resulting embryo or embryos to the uterus. To decide which to transfer, fertility clinics grade each embryo based on various features. However, those assessments can vary widely.


“It’s a very subjective process,” Shafiee said. “It really depends on the experience of the embryologist and even things like if the person is tired.”

Some experts hope AI could one day be akin to the most experienced, sharpest embryologist every time, making the process more objective and standardized. In recent research, Shafiee and his colleagues trained a type of AI called a convolutional neural network on images of embryos, which were labeled by experts with a score. In a test on 97 different embryo cohorts — which included a total of 742 embryos — the system was able to pick the highest-quality ones accurately 90% of the time.

The researchers also trained a system to directly assess a particular embryo’s implantation potential using data from 97 embryos from five clinics. That system scored an embryo accurately 73% of the time, compared to 67% among 15 trained embryologists.

“AI can see other things and bring in other factors that can improve the procedure,” Shafiee said. “That’s the beauty of this approach.”

The research is still in early stages, but the technology is already catching on with some companies. The IVF startup Life Whisperer is working with clinics in five countries to use AI to help select embryos for implantation. Its model was trained on photos of 20,000 embryos and their outcomes, submitted by 12 clinics from around the world.

Michelle Perugini, Life Whisperer’s co-founder and chief executive, said the company’s technology can lead to pregnancy 15% faster than traditional selection methods, results that the company presented at an American Society for Reproductive Medicine conference, but is not yet a journal publication. Speeding up the process is not only a cost-saver for clients, who end up doing fewer cycles, but also a boon for clinics, which charge less for second and third rounds of IVF and can instead focus on new clients.

Another part of the process that Life Whisperer and others are looking into is using images of embryos to detect chromosomal abnormalities, which have been tied to implantation failures and pregnancy loss.

Better predictions might not only be able to improve the process for patients, but could also improve a clinic’s calculations when it comes to offering specific payment plans for IVF clients, such as the services offered by Univfy.

Clinics sign up with Univfy, which feeds a potential client’s data on factors like maternal age, a partner’s fertility, and hormone levels into a model trained to determine pregnancy chances. Univfy then provides an easy-to-understand report to help patients decide how to proceed.

Many of the clinics Univfy works with provide insurance based on the predictions made in those reports. For example, a patient with an 80% chance of getting pregnant might pay $26,000 for three cycles upfront. If the patient gets pregnant from the first cycle, they’ve spent more than they would have otherwise, but if the three attempts don’t succeed, they would get $20,800 back.

A number of IVF clinics already offer refund programs. But Mylene Yao, the CEO and co-founder of Univfy, said that relying on AI-powered models allows clinics to offer more tailored payment or refund options to a broader pool of patients.

Each of the 50 clinics Univfy works with — which span 17 states and Canada — gets a bespoke model trained only on its own data. Univfy said that allows each center’s model to be as predictive as possible for its particular pool of patients, which might look different than the makeup of patients at a different clinic.

“The patient populations are all so different,” said Yao. “Not only from state to state, but from metropolitan to suburbs, and even from suburb to nearby suburb.”

Yao said the company plans to work on ways to add genomic and therapeutic data for more personalized recommendations.

“There are so many ways that fertility could be advanced with current day technology,” Yao said.

A baby conceived with the guidance of Life Whisperer’s model is expected to be born any day now in Australia. But such a development likely won’t happen any time soon in the U.S., given that Life Whisperer hasn’t secured approval from the Food and Drug Administration. The FDA has approved a handful of AI products in different medical fields and proposed a framework to evaluate AI, but fertility tools using the technology have mostly languished.

“They’ve struggled to get their head around these AIs,” Perugini said. From her perspective, that regulatory roadblock is “ limiting the access to technology that could really help.”

But other experts say there’s still a need for more evidence on not only how accurate the models are, but whether they can consistently improve outcomes. They caution that these predictions are another datapoint to consider, rather than a replacement to human expertise.

“I don’t think at the moment that we’re to the point where AI will completely override our morphological grading scheme,” said Matthew VerMilyea, the vice president of scientific advancement at Ovation Fertility, a network of IVF and genetic testing labs. “The plan is to use this as another tool in our toolbox to help give us more confidence that we’re selecting the right embryo.”

But there remain unanswered questions around how to add AI ethically into the fertility process, particularly in situations where patients struggling with infertility might be led to pin too high of hopes on a technology that is still not well-understood.

“There is a lot of pressure on a lot of couples trying to conceive,” said Junaid Nabi, a public health and bioethics researcher at the Aspen Institute. “The problem here is it is a matter of life.”

Most of these systems are black boxes, which means the data or calculations behind them aren’t publicly disclosed. And like with nearly any experimental tool, they run the risk of providing inaccurate results. Researchers caution that the AI they build should only be used as one component in the decision-making process, but there could still be legal repercussions if an embryo selected with the help of AI turns out to have an abnormality the system didn’t catch.

Life Whisperer’s Perugini is careful to point out that any algorithm her company would use would be to detect disorders in service of picking the embryo most likely to result in a healthy baby, and not about weeding out specific qualities.

There are no guidelines that can help doctors and patients deal with the ethical issues that AI creates specifically when used in IVF, according to Nadi. He warned that there’s a need to navigate those issues before the technology becomes more widely used.

“They can keep working on the algorithms and improve them,” Nabi said. “But relying on algorithms isn’t a solution.”

This is part of a yearlong series of articles exploring the use of artificial intelligence in health care that is partly funded by a grant from the Commonwealth Fund.

Source: STAT