Study finds AI better than a doctor at crucial stage of IVF

Will AI replace fertility doctors? Why computers are the only ones that can end the agony of failed IVF cycles, miscarriages, and risky multiple births

• Two new papers being presented today will reveal huge success with AI at selecting viable embryos
• It could cut the risk of failed IVF cycles, risky multiple births and miscarriages
• Here, we explain the benefits and downfalls of the technology and other innovations 

Around 200,000 couples a year try IVF. Two thirds of couples who try IVF experience failure at least once.

Fertility is a murky area of science, and it is likely that we don’t even know what questions to ask to find the answers to our current questions.

But what is clear is that a significant proportion of failed cycles are down to issues with the embryo – and computers may be able to fix that.

Two new papers being presented later today at this week’s American Society for Reproductive Medicine conference will reveal claims of unprecedented success at selecting viable embryos using artificial intelligence.

In doing so, they could dramatically lower the risk of failed cycles, and miscarriages (70 percent of which are caused by embryo abnormalities). They could also put an end to the gung-ho (but, for now, necessary) approach of implanting multiple embryos to maximizes chances, which often leads to twins or triplets – and the risk of preterm birth, preeclampsia, and costly childbirth complications.

Now the two teams – one from the US and one from Australia – are racing to patent their algorithms to enter the $7.5 billion AI market as the first barrier to human life.

But there are some areas it still cannot cover… 

Many are investing hope in technology to get around another human fallacy in fertility: judgment

WHAT IS THE PROBLEM WITH THE CURRENT SYSTEM OF SCREENING EMBRYOS?  

IVF and egg freezing now allow couples the chance to defy human biology to have babies despite fertility issues, or past the mother’s 40th birthday.

But the crucial step of picking the right blastocysts – which develop into embryos – is still left to the human eye.

Within the parameters of what is clearly normal, and what is clearly abnormal, every embryologist has their own way of deciding whether they fancy the chances of a blastocyst to make it all the way.       

It’s not a perfect art, so, naturally, hopes are high that technology could help us get around another human fallacy: judgement.

WHAT DETERMINES IF AN EMBRYO WILL SURVIVE?

There are many things that can derail the IVF process, including many factors that we don’t even know about yet.

When it comes to something like miscarriages – painfully common in IVF patients – embryos are the focus.

Embryo abnormalities account for 70 percent of miscarriages, while the rest is down to other factors, such as the mother’s response to hormones, uterine disorders, hormone levels, etc.

From an embryologists perspective, there are three main areas to be concerned about:  

1. Does it look normal? This is where AI could play a role. 

Embryologists decide which blastocysts to implant by assessing whether they look normal, and whether they are developing steadily (growing a little bit every 10 to 12 hours, rather than irregular bursts). 

Nobody knows what makes a blastocyst look normal or abnormal, or grow steadily. But the two new studies (described more below) showed that by training a computer to understand what looks normal and abnormal, we could achieve very reliable results. 

2. The genetic factor: Is it chromosomally normal? This is where AI cannot help – yet. 

But other new tech can. AI is trained to learn what normal development looks like. Just like the humans that trained it, the computer cannot see whether the blastocyst has any genetic defects by sight alone. Zaninovic hopes to incorporate DNA screening into the AI model one day, but does not yet know how that would work. 

There are other tools, though, which could be used in conjunction with it.  

Last month, Columbia University’s Dr S. Zev Williams announced a new DNA sequencer, a tiny handheld screening tool, between the size of a small chocolate bar and a USB, which can deliver verdicts within minutes, rather than weeks as current testing methods do. 

Dr Williams, who was once a fellow in the Cornell lab that developed one of the new AI techniques, said that this gadget could work ‘synergistically’ with the AI (if the AI does prove effective in more trials). 

‘What you’ll end up saying is: within this group of normal embryos, which looks the healthiest? Or within this group of healthy-appearing embryos, which one has a regular number of chromosomes?’

A whole other area of concern is an issue called ‘mosaicism’, when an embryo is borderline. It’s an area we still don’t fully understand, but the crux of it is that genetic testing has improved to such an extent that we don’t just see normal or abnormal embryos, we see some in the middle. The problem is, that decision has to be put to the couple, who are already riding a whirlwind of emotions. Many couples undergoing fertility treatment yield few viable embryos. If their only one is mosaic, it can feel like an agonizing gamble. 

The goal is to be able to (a) determine how much of a problem mosaicism really is, and (b) distinguish between risky mosaic embryos and safe ones. But we’re nowhere near that yet. 

3. The mother factor: Will it implant? No technology can help in this regard yet, but pills might. 

Once an embryologist has selected the perfect embryo, which looks normal and is genetically sound, there is still a chance that it won’t implant. Again, as with everything in fertility, it is not black and white. Sometimes it’s down to undiagnosed endometriosis (a disorder that affects the womb lining), or a blood clotting issue, or even that they have ‘natural killer’ cells that react to the synthetic hormones as if they were a virus. 

One theory is that, for some women, the uterus spontaneously contracts in reaction to the procedure and the hormones, which prevents the embryo from implanting. 

In a bid to tackle – and capitalize on – this mysterious problem, a Swiss company, ObsEva, has developed a pill called Nolasiban which can be given to the patient four hours before implantation to balance her hormones and prevent these contractions. In a paper which won a prize from SART (the Society for Assisted Reproductive Technology), they showed it improved the rate of embryos implanting by 32 percent. 

ObsEva CEO Ernest Loumaye, a gynecologist by trade, told DailyMail.com they expect to price the drug at around $3,000 per patient, though they are vying for reimbursement in Europe where IVF is subsidized by most states. It’s an eye-watering cost but Loumaye defends it as ‘much cheaper to insurance companies than the price of twins or triplets. They are more expensive during pregnancy, birth, and carry more risks’.

• 
  Male fertility crisis: Obesity, smoking and exposure to…

  
  EXCLUSIVE: Are women being sold an expensive fertility lie?…

Share this article

HOW DO THE NEW A.I. SYSTEMS WORK – AND HOW WELL?

Both of the studies being presented at this week’s ASRM conference trained the AI network using time lapse images or videos, showing how an embryo developed over time.

The first, led by Nikita Zaninovic at Cornell University, used 18,000 images to train the computer (‘we call it The Beast,’ Zaninovic told DailyMail.com), then they gave it another 32,000 to make a judgement on.  

First, it had to standardize the embryos. Currently, each embryologist has their own way of categorizing embryos, which changes from lab to lab – and can even change from embryologist to embryologist. Cornell’s AI was almost 100 percent successful (97.52 percent) at categorizing the embryos into groups of good, moderate and poor quality.

It then had to select which embryo had the best chance of making it to live birth. Comparing the computer’s decisions to their live birth data from patients, they found the computer accurately selected a viable embryo 85 percent of the time. 

The other AI was developed by Aengus Tran, a medical undergraduate at the University of New South Wales, working with his brother, a business undergraduate also at UNSW.

They developed a system, which they called Ivy, which was trained using time lapse videos then given to eight labs in four different countries.

Those labs used the technique to screen 1,603 patients aged 22 to 50, with the results double-checked by an embryologist.

It worked 93 percent of the time.  

These aren’t the first AI algorithms designed to test embryos. 

Eeva was approved by the FDA in 2014, which can screen abnormalities, but it is not a dynamic system. It cannot react to new information like these ‘deep-learning’ AIs can. 

Beyond that, the internet is littered with online companies claiming to do the same.

Zaninovic calls them ‘cowboys’. 

‘I haven’t seen their data. They don’t come to conferences like this [ASRM] to share their data with the rest of the industry, or publish their work.’

What’s more, he says, ‘they make some bombastic claims’.

One company in Australia, Life Whisperer, has made headlines with the claim that they can spot Down syndrome using AI, a genetic disorder affecting chromosome 21. 

As far as we know, it’s something that would need to be spotted in genetic screening, rather than on pure aesthetics. But we could be wrong. It could be that genetic abnormalities could be spotted using AI, and we just don’t know it yet. 

The problem, Zaninovic warns, is that AI learns using data. It only understands what’s normal and what’s not by first seeing examples of what’s normal and what’s not. 

‘For AI to detect Down syndrome you would have to show it lots of examples of Down syndrome embryos. And let me tell you, there aren’t many. That’s why I’m skeptical,’ he says.  

WILL THIS PUT EMBRYOLOGISTS OUT OF A JOB?

Zaninovic insists it won’t.

‘People say I’m shooting myself in the foot,’ he said.

‘This isn’t replacing embryologists, it’s just that your job will change.’

He’s probably right. In fact, a study by MIT last year compared human-only teams, robot-only teams and human-robot teams. They found that, by far, the most efficient category was the human-robot combination, which was 85 percent more efficient than the other two groups.

The biggest change, he says, will be that hospitals and clinics that want to implement this technique will need a full team of IT support. 

Ultimately, the opportunity to relieve pain and suffering is so great that it shouldn’t matter, he says. 

Dr Williams agrees.

‘Most people after a loss they blame themselves,’ he says.

‘They lifted something to heavy, they got into an argument at work, they got stressed. So they’re really upset with themselves, they feel a lot of self-blame and guilt, inappropriately. They shouldn’t, but it’s very natural of humans to do so. 

‘We have this ability to test and say no it wasn’t because of these things it was because, from the moment that embryo was formed, it was destined to result in this, and laying on your back for 40 weeks would have made zero difference. It’s very powerful to be able to show that.’ 

‘Things like miscarriage and infertility are some of the most ancient diseases,’ he adds. ‘I think there’s something poetic about solving that with some of the newest technology.’

Source: Read Full Article