LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.

[INTRODUCTION]

[0:00:03] HC: Welcome to Impact AI. Brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven, machine-learning-powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research in computer vision for people in planetary health. You can sign up at pixelscientia.com/newsletter.

[EPISODE]

[0:00:34] HC: Today, I’m joined by guest, Daniella Gilboa, co-founder and CEO of AIVF, to talk about in vitro fertilization. Daniella, welcome to the show.

[0:00:44] DG: Hi, thank you for having me.

[0:00:45] HC: Daniella, could you share a bit about your background and how that led you to create AIVF?

[0:00:50] DG: Yes So, I come from deep within the IVF field. I’m a clinical embryologist, which is an IVF biologist. I live and breathe embryos, so it’s really from understanding it from the clinical aspect, as well as the epidemiology or the data aspect. So, I have two hats. One is clinical embryologist. Another is statistician. I study IVF from the field, from like the wet work of biology, as well as looking at the data.

[0:01:22] HC: What does AIVF do? Why is it important?

[0:01:26] DG: Right. So, during my time in, spending time in IVF labs, I think you really can understand that on the one hand, IVF as a technology is probably one of the most important ones that it was invented, I think in the last 50 years in medicine. But on the other hand, it’s just not good enough. It’s not optimized. It could be optimized, but it’s not there. I’ll explain what I mean. There is kind of like an uncertainty. When you start IVF, you really don’t know how long it will take you, whether you will succeed or not, how much it will cost you. You’re really not able to predict or plan your fertility journey. I think by 2024, the fact that we can plan something ahead is crucial.

So, with IVF, and you’ll hear it from so many patients, the fact that you really don’t know how you’ll come out of it, this is a huge problem. Then you see the statistics and you see it, you know that most of the cycles fail. You don’t succeed. Most of the cycles – the statistics is that you fail most of the cycles. It’s very hard. It’s agonizing. You really need to focus on that. It’s very hard doing a career or studying while going through IVF. All of this makes you want to – I think, the question is, can we do better? Or how can we optimize it? What’s the point or the point in time in the process where you say, “This is where it fails.”

This is how we started AIVF. I’ll start with a story, a real true story. We had one patient, this is a story that really, I think, changed my mind in the sense that this is what needs to be done. We had a patient and she was going through numerous IVF cycles. I would say even, she was close to – this was I think the 9th or 10th cycle that she went through and it was failure after failure. Her doctor came into the lab one day and he said to me, “Please take a look at this patient, the embryos. Do your magic. It’s her last chance. She’s not going to try another cycle. This is it for her. Let’s see what we can do.”

I looked at the file. It was a very thick patient file. It was like a book. I looked through the treatments and I couldn’t say why she was not succeeding. Everything kind of like looked okay. She had a good number of eggs and a high fertilization rate and she had – the embryos looked pretty much okay, according to the grading of the embryologist. You couldn’t really say why she was not succeeding. I called a friend who was working with me with the lab that day, and we looked at the embryos. Her embryos were incubated inside an incubator that has a camera inside, and so you’re able to look at the video of the embryo. We could really – not analyzing one single image, but rather the entire video of the embryo from fertilization to the blastocyst stage. That’s about five to six days in the lab.

She had like 12 embryos and this is what we did for like four hours. Going back and forth with the video, like editing the video, trying to understand or capture the different milestones that we are looking for in the developing embryo. I’ll give you an example. When was the time that the embryo was fertilized? When was the time the event where the embryo became a two-cell? And then a four-cell embryo? And then an eight-cell?

So, we were looking at the events that embryologists usually look for. This is what we did for four hours. You understand that it’s completely not scalable. We ended up with analyzing all the embryos, and we ended up with two very good-looking embryos. I remember that I said to her, “You have a very good chance of succeeding this time of becoming pregnant. There’s two good-looking embryos. They got the highest score, the highest grade. Let’s do it.” That’s it. We did the embryo transfer and I forgot all about it.

Then, two weeks after this doctor bumped into me and he says, this is what he said, exactly. I quote, “I don’t know what magic you did there, but she’s pregnant.” Then, it hit me that I didn’t do anything special. I just did what I had to do. This is what I was trained to do. Evaluate embryos. This is, I think, the core of what we do as embryologists. Understand embryos, evaluate embryos, diagnose embryos. This is where I said, “This needs to be the standard of care. But it’s completely not scalable. This is where we have to put AI into the system that will do exactly like humans, and even better.”

[0:06:50] HC: So, AI can look at that data, and a whole lot more efficiently, and perhaps in more detail than a single person could.

[0:06:57] DG: Exactly.

[0:06:59] HC: How do you apply machine learning to this? AI is a great buzzword. But how does this work in practice?

[0:07:04] DG: So, we started with collecting huge amounts of data, and this is the data that I was talking about. Videos of embryos, and for each video, we had the outcome, whether it was implanted and born, even the gender, boy or girl, whether it’s normal or abnormal. So, we had all different kinds of outcomes, and we said, let’s start with a basic AI model that is able to, I would say, scan or look at the videos of the embryos. Look at the embryos. Understand the different features in the developing embryo and correlate it to the outcome.

So very simple. Exactly like the human. Let’s become as good as the embryologist, this was the very first step. I think later in the process, we were able to show these models are better than humans. We conducted different studies showing the embryologist versus the machine, the AI, and what we actually showed, and this is interesting that it’s the best performance you get when the embryologist is completely aligned and synced with the AI, rather than doing it by himself. So, it’s the AI and the human that gets the best performance, versus the human or the AI by itself, which is great because it’s the interaction of the AI and the human that gets the best performance.

Anyway, the models are machine learning models that are basically video analysis. It’s analyzing videos, identifying the different features, and integrating it into one model that is able to predict the outcome. This prediction is the output is a score from 1 to 10. That is from 1 to 9.9 that is correlated with the actual outcome. So, if an embryo that was analyzed by our score gets, I would say, an example, gets a score of a nine, I can actually tell you what the probability of success is. This is great, because for a patient to really be able to predict or plan this journey, and looking at the embryos and saying, “Okay, I have five embryos and all of them gets great scores. This means that I have very good chance of becoming pregnant and I’ll transfer one, I can freeze four, and I know what I’m dealing with.” This is really reducing the uncertainties.

[0:09:54] HC: Then, training models based on these embryo videos. What kind of challenges do you encounter?

[0:09:58] DG: Many. First you need data, basically from all over the world, because the issue of heterogeneity of the data. So, the fact that you have data from one clinic, is these models could be great for, I would say, academic, or scientific publications, but they can never become a real product that is robust and strong enough to be working in a clinical setting. So, you need huge amounts of data from different places around the globe, different patient demographics, age groups. You need to really be able to collect a lot of data. Make sure that each embryo has the outcome. The fact that you have just huge data of just embryos, and you don’t know what happened to each and every embryo, you can’t run a model on these. So, this is one aspect.

The other is really how do you collect such data, such huge amounts of data? What kind of partnerships you do with clinics? Most of these data come from clinics. So, you have to have some partnerships with these clinics. What kind of studies and validations you need to do? Once you have these models, you really have to validate them first on retrospective analysis, and then put them in clinics, and really see how they work in real clinical settings. It’s like back and forth, back and forth, and really understanding that the validation studies that you do are crucial, really important for – at the end of the day, the model, the product, becomes a product needs to be robust, needs to be strong, and needs to be something that patients and embryologists and physicians really rely on.

[0:11:46] HC: How do you go about validating these models? You mentioned some of the complexities, but what’s your strategy?

[0:11:52] DG: First of all, I think it’s really showing it retrospective analysis. I think this gives you some sense of understanding where you are. This is one thing. The other thing is, it’s like two different languages. You have a great group. We have a great group of AI researchers and they come out with KPIs or metrics such as performance or accuracy of the algorithm. Then, you need to translate it into something that is in clinical usage. What do you mean that the algorithm has an accuracy of X? What does it mean – tomorrow morning in a clinical setting, what does it mean?

So, I think that these two languages, totally different languages was, I think, what we needed to really understand and understand how to translate it. We started with putting these models into real running them in real clinics. These were the first clinics that we had. We’re kind of like a design partner. So, we just asked them to try it. It was all under a clinical study design. The patients were either we asked for their consent, or the clinic asked for their consent, but it was all under a clinical study. So, we just tested it. The model versus the embryologist. Or the performance of the model versus the performance of the lab. Something like that. We just showed, for each and every clinic, we showed that if you use AI as a tool for the embryologist. We showed it, each and every time, increased the success rates. Increased shorter time to pregnancy. The decision-making is faster, more accurate. You freeze less embryos because each embryo you freeze is accurate. You don’t just freeze for the sake of freezing. You freeze the right embryos that have the highest chances of becoming a baby.

So, it really changes the way the lab works and it just optimizes everything.

[0:14:02] HC: You mentioned earlier that the best solution is the embryologist working with the AI. For that to work and for the embryologist to really trust the AI, is explainability of the model a factor here? If so, how do you make your models explainable?

[0:14:17] DG: Yes. That’s a good question. Actually, it was, I think, these systems, not only in IVF, but I think in medicine, in general. So, they’re considered decision support. Because at the end of the day, it will always be the decision of the embryologist. But we provide a tool. So, instead of just relying on the way we grade embryos, we embryologists grade embryos, there is now another tool for me to grade embryos that I think better than any system that we had. But I think for me, as an embryologist, just using all these new technologies, at the end of the day, provide me with a better way of doing my job.

We used all kinds of messaging here. We did some webinars. I was going from one stage to another. We presented all the work at scientific conferences and I was going from one lab to another, and just showing them that it’s not taking their job. It’s just making them – it empowers them, and it’s just making them better at what they do. Part of it is explainability. So, I’ll give you an example. If I’m an embryologist, and I’m looking for specific features, and this is how I make decisions, then I expect the AI to be looking at the same features.

The black box of AI doesn’t really work in medicine, and I think physicians and embryologist, and all caregivers, we really want to know that this AI is real. So, part of it is explainability. It was very complicated and challenging for the AI team to provide explainability. But we did that. We took a lot of energy here, a lot of resources to really come up with explainable models and say, “This is what the AI is looking for, just for like a sanity check for the embryologist until they rely on it, and until I think they have enough trust.” But this is the way to go in medicine, and we did that.

Then, later on, we actually showed that the AI is capable of identifying features that cannot be seen by the human eye, features that are highly correlated with different outcomes. So, this was just amazing. I think by the time that we showed this, people said, “Okay, I got it.” But this was just unbelievable. But this is really the beauty of developing AI models.

[0:16:53] HC: So then, any AI could teach the embryologist what these new features are and what they might look for as well?

[0:16:59] DG: Yes. When we first presented it, I called it computational embryology. Because really, we learn new stuff, new things. We learn – it’s like becoming a new field of not the, I would say, conventional or the classic embryology. But now it’s computational embryology where we use machine learning or we use AI to understand embryonic development from different angles. We couldn’t have done it five, seven years ago before. It’s very exciting for us and we have relationships with different academic centers, and some master’s degrees, and Ph.D on exactly this computational embryology. So, it’s something that is becoming deep inside the academic world and it’s very interesting for us.

Another part of developing AI is regulation or the regulatory process and the ethical process. So, you have so many other – so many, I would say, angles to this product that’s supposed to optimize IVF and really, really help patients at the end of the day. But you have to look at the science part and the regulatory part and the ethical considerations here. So, it’s just very, very interesting to be involved in such an exciting field.

[0:18:28] HC: Why is now the right time for all of this? You mentioned a bunch of stuff going on, but why now and not five years ago, or five years from now?

[0:18:35] DG: That’s a good question, actually. When we started, that was back in 2018. So, I remember talking to so many IVF experts, and I said, “There might be a better way.” They said, “No. No. We have great embryologists,” one of them said to me, “I have great embryologists. I really don’t need any outside system.” But I think what happened is, little by little AI became something that was just there way before ChatGPT and was actually being able to see the change like a year after the scientific conferences suddenly accepted abstracts about AI. Then, more and more people presented their science, their AI science in our field. It wasn’t only epidemiological studies being presented or basic science, but suddenly you had computer scientists standing on the stage explaining why their model is predicting embryonic development. So, it was very interesting to see that change.

Then, came ChatGPT. This was just – this exploded once ChatGPT was there. So, I think people really understood what AI can do for them, and it became a household name. Suddenly AI is just there. It’s all over the place.

This is one aspect. The other aspect is patients. Patients are asking and looking for it. You can’t expect to be living the 2024, very advanced technology, and go have your – go through any medical process or journey and feel like 50 years ago. So, you’re actually looking for it. Patients were starting started saying, “Do you apply any new technologies in your lab? Do you use AI? There is something, I hear there’s something, are you using it?” I think there’s now, we’re seeing a momentum of clinics and physicians and embryologists actually actively looking for these new technologies and solutions. I think by now, we all understand that it empowers the way we do embryology and we make it much, much more accurate embryology than ever before.

[0:21:12] HC: How do you measure the impact of this technology to – it’s clearly a positive impact for the women and their fertility journey. But how do you measure that change?

[0:21:21] DG: Over time. I could tell you now that the time to pregnancy with AIVF is 1.6 cycles, which is a significant, significant reduction versus the average, which is above 3. It really depends on the clinic. But this is the average. So, it’s significant. I think, over time, you see there’s more and more and more clinics using AI and you see increasing success rates. You see, again, the different, the way they do, the way they manage their day-to-day work, the way they make decisions, faster decisions, more accurate decisions. We save time. A lot of time. Part of it is, I think, a holistic approach of, imagine a physician sitting in front of a patient and being able to explain stuff easily with real numbers.

So, the language is changing. It’s no more something like, “You have nice looking embryos, you have high chances or good chances,” which really tells you nothing. I think the language now is, “These are your embryos. This is the score. This means that you have 80% or 90%, we suggest you freeze one, or you freeze all, or we suggest you do a fresh transfer and the rest of is like,” the way you interact with the patient and you consult the journey ahead is changing and it’s more accurate. It just allows you, with the patient, to make more informed decisions. This is the right way of doing medicine, by the way. It needs to be data-driven, rather than subjective human analysis.

We see that our patients and doctors are, how would I say, it’s less computer time, more patient time. So, because of this very informed interaction, it’s an interaction that is intelligent. It empowers not only the physician, but also the patient. It’s all part of it. It’s really changing the way we do medicine.

[0:23:32] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:23:36] DG: I love the question. Yes. So, resilience. I think it’s very a complicated field, because you have an amazing innovation, where you need to validate it so many times and in so many aspects. There’s a very long process of regulation, and you need to make sure that you do it right. I think resilience. I think loving the field. I think the science here is crucial. So, everything is science. You talk science. The product is science. The development is something that you always do. You never stop developing. The models keep on training and developing by themselves. So, you become better and better and better and it’s always there. The science and publishing it, and writing peer-reviewed papers. Part of the marketing is science. So, it’s just really different in AI in the medical world, rather than other AI companies that are not medical. We’re considered a medical device. So, how do you regulate AI? This is something that’s a huge topic that everyone’s concerned with. It’s just something that the company develops with time and the field is developing. And I think it’s just being more patient and things take more time, and it’s a bit slower.

Again, AI in medicine, but it’s worth it because the impact is just huge. The impact is amazing for us, seeing patients and then talking to patients to tell us that they’re pregnant, and they’re having a baby because of us. So, the impact is unbelievable. And I actually say to my, to the team, at AIVF. I say, every line of code you write is translated tomorrow morning to more happy people.

[0:25:46] HC: What do you say for the future of AIVF? Where do you see the impact of your company in three to five years?

[0:25:51] DG: I think, AIVF needs to become the standard of care. So, it needs to be incorporated in IVF in general, in all clinics. I see us becoming more involved with patients and advising patients. It’s not only a system that is facing the clinic, but also a system that is facing the patient and really helping the patient understand, and do a more informed fertility journey and be there in the different stages of the developing, of the fertility journey.

[0:26:31] HC: This has been great, Daniella. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:26:39] DG: Thank you so much. Aivf.co. Look for AIVF. Google us, look for us, aivf.co, and I’m always happy to talk with anyone who’s interested. So, look me up, LinkedIn, Facebook, I’m there.

[0:26:57] HC: Perfect. I’ll link to all of that in the show notes. Thanks for joining me today.

[0:27:01] DG: Thank you so much.

[0:27:02] HC: All right, everyone. Thanks for listening. I’m Heather Couture and I hope you join me again next time for Impact AI.

[OUTRO]

[0:27:13] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share it with a friend, and if you’d like to learn more about computer vision applications for people in planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]