Tell us about yourself. What are you working on right now?

I’ve been a clinical embryologist for almost 25 years, and I currently work at a university hospital in Valencia, where we do IVF. I’m very involved in reproductive societies, both here in Spain and all around Europe, and as a university hospital we of course also do a lot of research, and training of new professionals—doctors, embryologists, nurses—as well as treatment of patients.

Some of our readers won’t be familiar with how IVF works—could you explain it for us?

Humans are very bad at reproduction—and within reproductive health we have different kinds of treatments, with IVF as one of the more complex. 

Normally, every month a woman develops a lot of oocytes [eggs] inside their ovary, but usually only one will mature, ovulate, and, in the best case, be fertilized. In IVF, we use hormones to stimulate the ovary to develop more than one oocyte in a single month, extract them all, and then inject them with sperm. Of those only 60-70 percent will fertilize, and then [after five days on average] only around half of those will have further developed into the earliest form of embryo, a blastocyst. Only some blastocysts will be of good enough quality to be transferred into the uterus, only some of those transferred will actually implant, and only some of those that implant will lead to pregnancy.

That’s why it’s so important that we check the embryos along every stage of their development. In every phase there are markers we look for—for example, ones signifying that an embryo is aneuploid [has an incorrect number of chromosomes], which makes non-implantation or miscarriage more likely. Checking the embryos at every stage increases the probability of selecting the best one to transfer, which in turn increases the probability of a baby.

It’s very important to note that IVF is not about pregnancy—the goal is a healthy, live birth.

How accurate are those embryo assessments with current technology?

It depends on your tools, and the information you can extract from your embryos. For example, if you culture your embryos in a conventional incubator you can’t see them except under a microscope. But every time you take an embryo out of an incubator you’re exposing it to changes in light, temperature, and atmosphere—we use gasses to mimic the environment of the uterus—so every examination will have an impact. In that situation embryologists only check the embryos once a day, but then your best-case scenario is you only have five or six single moments of information. You don’t know what happened across each 24-hour day, at moments when maybe some other specific markers of viability were visible. 

In our laboratory, almost 10 years ago, we started to use a timelapse system in our incubators—these take a photo every five minutes, in seven to nine different focal depths. This lets us track more markers than just embryo morphology—for example, the specific times when things happen are also markers for implantation chance and possible aneuploidy. Nowadays we can also analyze the medium we’re using to culture the embryos to detect any release of metabolites or DNA, and there’s PGT-A testing, too, where we biopsy an embryo and sequence its genome. 

The more markers we can track and analyze, the more accurate we can be about an embryo’s quality—and we now have more information than ever before.

How are you applying AI to this process?

With our time lapse system, we were generating a huge amount of information per embryo—but when I was selecting embryos I was only checking a few key parameters, because my brain couldn’t possibly process and score all of the images and other data that was recorded. It would take too much time.

Then I met J. C. Riveiro, one of the founders of Cercle, and we started work on using machine learning to automate the process of watching embryos. AI makes it possible to analyze everything, and the system could provide a better breakdown of each embryo’s quality.

But when we started to analyze our results, J. C. pointed out that I was always complaining about its recommendations—I was saying, “well, that really depends on the age of the woman,” or, “that depends on whether the patient has a history of smoking,” or, “that depends on sperm quality.” We realized that analyzing embryos alone wasn’t enough, so, instead of putting into the market a product that only made the analysis process more comfortable, we realized that an approach which helps everyone in healthcare—clinicians and patients alike—was more important. 

We think that personalized healthcare, specific to each patient, is best—and not something that just tells an embryologist (or any other doctor) what to do. We have to provide the professionals with all the data that’s important for each specific patient, and advise them according to that, not to an opinion or to current trends. For example, in IVF, 10 years ago the trend was for a certain amount of [reproductive hormone] gonadotropin for stimulation of the ovaries, but in another 10 years there will be a different trend, and we don’t know if an older protocol might actually still be more effective for a new patient. We decided to make another kind of AI assistant to help with that.

How does this broader AI system work?

We take all the information about an embryo and place it in a graph which Cercle’s fertility tool can then analyze using deep learning—but that graph also contains all the other information from outside the embryos. Machine learning can tell you that an embryo is a 10/10, with the best chance of implantation, but if you don’t include everything about the patient too then your information is biased and you won’t have an accurate system.

After all, even if I have the best embryo ever it isn’t going to matter if I’m placing it into an endometrium [uterus lining] which won’t accept it. The probability of implantation is actually zero. Cercle’s system would instead show you the final results from similar situations, and give you the opportunity to better advise [the patient] about the real options in each situation.

It offers possible treatments which are specific to each patient, but based on final outcomes for comparable patients. If I have a patient needing ovarian stimulation, the system won’t tell me, “you have to prescribe 200 mcg [of a drug]”—instead, it’ll provide me with profiles of similar patients from our records, which protocols they were given, and what their results were. For example, if you have a patient who wants to increase their chances of becoming pregnant then you’ll be shown completely different protocols than would be given to patients who are instead interested in just freezing their eggs, where you want to harvest as many oocytes as possible but don’t have to also prepare for embryo transfer within the same cycle. 

It also becomes more accurate the more information we give it. Let’s say a woman is 38 years old and blonde—if I put that into the system, it’ll give me a million patients with a similar profile and just as many possibilities for treatment. But if I also introduce her hormone profiles, her treatment history, the sperm donor’s blood tests, and more, then the system will narrow down. It’s a completely different machine learning approach than systems which use generative AI, because everything it produces is linked to real data, not supposition. However, we don’t actually process or use any personally identifiable data directly—our graph stores medical information on specific cases but nothing that would identify any specific patient, and that’s how we ensure there’s no possibility of personal information being accidentally leaked or revealed. Our team is also fully trained in healthcare data protection. We’re 100 percent HIPAA compliant.

How much does this rely on good record-keeping practices? Does it integrate with existing medical records systems?

Everything relies on it. In healthcare all medical information must be recorded, but we have lots of different formats—embryo photographs, radiological images, PDFs that patients email in with results from tests performed by their family doctor, Excel spreadsheets from other hospitals, and even handwritten notes. All that information is recorded digitally, and the system connects to all our databases so it can populate our graph.

While future records are updated automatically now everything is connected together, one of the more complex aspects has been retroactively analyzing all of our older information. That’s how you get maximum profit from this tool— processing all of that work you’ve been doing for years and years and years. I’ve been doing IVF for more than two decades, and there are patients whose histories are still stored here at the hospital, but I’ll never examine their embryos again. Maybe I’ll remember a patient from a few months ago, but I can’t remember the details from a case 20 years ago—I can’t even remember which protocol was in fashion back then, but there’s a chance it might actually be more effective for a patient I’m seeing now than the more modern protocol. 

We’ve had to understand all of the different types of information from each clinic and how they’re all different, but it now means that, with AI, I can keep reaching back to all of the patients that I’ve treated throughout my life, as can every other embryologist here.

Has AI led to improvements in outcomes for patients?

We developed the embryo selection system first—it was my field, and I wanted to improve the workflow in my lab—and our engineers had that ready after five months of development. We’ve been using it for the last two years, and, while we haven’t improved outcomes for patients with good prognoses, we have managed to increase the chances of pregnancy in patients with poorer prognoses. 

I’ve realized that I should have expected this—every embryologist can tell that a good embryo is good, but the difficult part is making a decision when you don’t necessarily know which “bad” embryos are less-bad than others, and you need to identify the least-bad option if it could still be viable for transfer. Most interestingly of all we’ve decreased the miscarriage rate, which I think is remarkable. As I mentioned, IVF patients come here to have a baby, not a pregnancy, and when you’ve been waiting for so long a miscarriage is really hard. 

Reducing the miscarriage rate in those patients was incredible, for me and the rest of the team.

Do you have a hot take about AI?

The best thing about AI is that it creates the possibility of managing data in a way that means we give better advice to patients, make better decisions, and ultimately give our patients better outcomes.

Management of data is the key—not necessarily automation, though that’s also good because it can save time and reduce subjectivity—but it’s all about managing data so that we can make better decisions. Data is the clue. I’m sure for people in tech their answer would be completely different, but for me, in healthcare, that’s the clear benefit.