Not Boring Founders: Emi Gal, Ezra

0:00Cancer is a really huge problem. like 40 % of the world's population will be diagnosed with cancer during their life. And the worst part about that is that half of the people who find cancer will find cancer late. And only two out of 10 people who find cancer late survive. Now, if we could find cancer early for those people, eight out of 10 would survive. I strongly believe that the cure for cancer is early detection. And so if we fast forward 10 years, about 10 million people today, every year, find cancer. Cancer late. At Ezra right now, we find cancer in about 10 % of our members, possible cancer.

0:39It's not often confirmed or it's pre-cancer. So for us to find cancer early in those 10 million people, we probably need to screen like 100 to 200 million a year. to give ourselves the best chance at finding those people who are from the scan will really add a ton of value because we found cancer. This story is about working hard even when it seems silly. And for the people trying to make crazy things happen.

1:11Not boring is for the optimist. Take a little shot of optimism. Take a little shot of Optimist.

1:27Let's just zoom out and take a little shot of Optimist. Emmy, welcome to Not Boring Founders. Thank you, Chris. We're here. So this is a fun and a new one for us. And I think I want to do this more as kind of follow-ups, but wrote a piece about Ezra a couple of weeks ago. The audience reception was phenomenal. Like it's very clear that people want a way to beat cancer. And I think that the main feedback that I got was one, like amazing that this technology exists, but two, the price point is actually very affordable, particularly with what you've done with something like this, when it's anything medical and I go into it, I'm always worried that people who know what they're talking about way more than me are going to dunk on it.

2:08And the reception was like 99.9 % positive, which was cool to see. So I'm excited to dig in even deeper with you today. You think me too? So to get started, I mean, in case people haven't read the 8 ,000 words that I wrote on Ezra, what do you do? So, very tender, CEO of Ezra, what we do is we're trying to beat cancer through early detection. And the way we do that is through a full-body MRI that takes about 30 minutes now. We've got an FDA clearance for a new AI that's enabled us to decrease scan time. So the scan takes 30 minutes. It's powered by AI on the scanning front, on the interplayation front, on the report generation.

2:45and then you get their scan and 20 days later we tell you whether you have cancer or not across the entire body that's the scan in a nutshell and it costs just over a thousand dollars yeah i did my scan and i did the 60 minute one because i think i missed it by about two weeks i had to do it ahead of the piece and the piece came out kind of when one flash was announced puja did it afterwards my wife and did the 30 minute scan and that it was like the easiest thing in the world like kind of in out luckily cancer free and it's just a really nice peace of mind one of the things that i was thinking is like does it seem too fast like are people like are you gonna miss something here if you do it like how do you build that trust what's feedback been yeah it's been great actually so we have about 30 000 members on our mailing list who are people who signed up for an extra scan maybe didn't get a scan in the end but kind of they're in our database and we sent out an email blast to everyone when we launched as a flash and i was looking at the responses from folks and everyone was like amazing i wanted to get us to scan but the price point was a problem, excited to get a scan.

3:47And we launched it June 1st, and we grew 70 % month over month in the month of June as a result of launching that scan. So the reception has been great. We have gotten a ton of questions around, hey, is it the same scan? Is it of the same quality? Is there anything different about it? And so on. And so we've had to do a bit of work to emphasize that thanks to this AI that was developed, the image quality is identical to the 60-minute scan. The only significant difference is that we remove spine in the 30-minute scan in order to kind of condense scanning time even further. And because spine is not really that relevant when it comes to cancer screen.

4:26Like you don't have cancer in the spine unless you have cancer in the primary organ and the flash scan does cover all the primary organs. So reception has been fantastic. We're now on our way working to make it even faster, probably 15 minutes in a year. Amazing. The fact that you're using AI for something that is so potentially beneficial for people, obviously been working on this for a while. Now everybody's adding AI into their product. What was the process to get here? How long have you been working on this? How did you figure out that AI was going to help here? Talk to me about your relationship with Dan Sedgson because I think that's really interesting.

5:01Tell me everything about incorporating AI here. Great question. So we've actually been at this for five years. I can't believe it's in the song, but I started Ezra in 2017. we incorporated and raised our seed rally in 2018. And we have since been working on AI before kind of AI was even cool to a large extent. The thesis behind Ezra from day one was that MRI as an imaging modality is incredible for screening because it doesn't expose it to radiation. It's even resonance because it's high sensitivity, relatively high specificity. It's just a really amazing modality, but it is a slow type of scan.

5:42Therefore, it's an expensive modality. And so the thesis from day one was, hey, I think we can use AI to make MRI super fast, super affordable. And that's when I started. I thought that was the thinking. It took us two years to build our first AI and get it cleared by the FDA. That was our prostate AI that assisted radiologists to when they read MRI scans of the prostate. It then took us 12 months to build and deliver our second AI, which is Ezra Glass in production. And I think our third AI that we're working on right now, it's going to take us about eight months. So we're kind of getting faster and faster and faster at building, getting FDA clearance and shipping AIS.

6:22And is the FDA on their end getting faster at it as well? Are they just seeing a lot more AI-based products? Are they more familiar and just kind of understanding? Like, why is that piece getting faster? The FDA has actually, I've been very impressed with the people at the FDA. They have really good biomedical engineers who review the type of AIs that we develop and submit. And they've been relatively fast at kind of accepting our applications, telling us whether they need more data or information, analyzing data, and then issuing clearance. Generally, the key is making sure that when you submit the data to the FDA to show that your medical device is safe and effective, that you submit data that is comprehensive.

7:00Like in our case, RAI is used on a screening asymptomatic population. So we had to have U.S. census representation of the population in the data set that was submitted. RAI enhances images. Some of those images will be images with cancer in there. So we had to submit data to show that in people who had cancer, we enhanced the images and we didn't erase the quality or what. And then RAI is applied across multiple manufacturers of MRI machines. We had to have a distribution of data that spans across the life manufacturers. There's a lot of work in AI development in healthcare that goes into just the data part, more so than the models itself.

7:44Will the AI replace radiologists? Like when someone's going to Ezra right now, could you just put me in the tube and then like the AI takes over and tells me whether or not I have cancer? Like where do you see those lines and where do you see that evolving over time? Yeah, so for now, at least, I think radiologists will be replaced by radiologists who use AI as opposed to just AI. I don't see a near future in which AIs are fully replacing radiologists, not as much because of a kind of technical limitation, but because of regulatory and kind of compliance limitations. Where does malpractice fall if you do the wrong job, if it's just an AI analysis?

8:20If it's an AI plus radiologist, it's a little bit more clear. So I think that technology will get good enough to be able to automate most of the work that a radiologist does. I'm not sure the healthcare system is prepared to embrace a fully automated kind of AI. And so our focus is to take each of the cost centers for an ASA scan, the scan itself, the radiology time, and the report generation times and internal operations time, and use AI to decrease the or core space across those core centers and then pass cost savings to consumers. I do envision a future in which AI can be used without a radiologist in the loop for change map detection.

9:03So for example, if you back, get done a scan and then you commit to another scan in a year, I see a future in which we're just using AI to analyze whether there are any significant differences in your subsequent scans. And if there are, that's when we do a kind of more high resolution scan and then put in a forward radio. When I talked to Dan Sedgson, the way that he described it, it was kind of this idea of induced demand that you just be able to do so many more screenings and AI could help with a lot of that stuff that never would have happened before. And then when it gets serious, you bring it to the human radiologist.

9:34And so it just increases the amount of screening that you can do, speeds up the time, lowers the cost, but that human is still there. Absolutely. And actually we're working on this$500 scan that we think will be able to launch in about two years. And this scan will be a 10-minute scan that is focused on just change map protection where the change will be fully done by AI, like the change analysis. So to do it in 10 minutes, are you, like I've gotten my full scan and then I come in and it just looks at areas that were problematic before or is it just lower, it's just doing everything but at even lower resolution, the AI fills in the gaps?

10:10So actually what got us to 30 minutes is not what's going to get us to 15. And so when you now go into an imaging facility to get a scan, as a scan or otherwise, the scanner has no prior information. It starts from zero. It acquires whatever protocol was designed for that particular scan. What we're working on is a concept of memory. So when you're going into an MRI machine, because we have priors about you, we can use that as memory in order to only acquire the data that is necessary for us to be able to detect a change do a diff essentially between the subsequent scans. That means that, Anne likes to say, Dan Sodexson, we're escaping the tyranny of image quality because if we're just focused on a change map, we're not going to acquire sequences that radiologists will look at.

10:59We'll acquire sequences that the AI will look at in order to determine whether the change is significant. If the change is significant, we will then bring you back in for a high-res typical scan that a radiologist can read. If the change is not significant, And we'll just be able to tell you, okay, nothing to be concerned about, see you in the year. And in that case, what happens if I've developed a new cancer somewhere else that you can't do change management detection on? Does it just not catch that? Like, will I get a false sense of security because it's only focused on the areas where there might have been a problem before?

11:29What's the right way to think about that? Yeah, so that is actually the key question. And it's the key area to solve for in doing kind of a change for architecture. And so to solve for that, what we've been doing from day one is we created this scoring system. We call it an e-score, an Ezra score, where every single finding in every single member from day one, I've ranked it on a score of one to five. One being just informative, five being emergent cancer. You need to do something about it immediately. And we've created this kind of really rich data set of e-scores. We then create a study, a clinical trial, where we call every person for whom we find custom cancer to see whether it was indeed cancer and what the results were.

12:12And then we enrich our data set with that. So our ability to build AIs that do change map detection will essentially be our ability to predict an e-score change from essentially two to three, three to four, four to five. By doing that, we become kind of agnostic to ORGA. We're literally just looking at this person across the entire body. Have they had a change from an E-score 1 or 2 to an E-score 3, 4 or 5? And if yes, that's how we're going to push you towards the next step, which is to get a higher risk. That's fascinating. So I guess talking about maybe false negatives brings up the big question that everybody has around false positives.

12:49We went into it a bunch in the piece, but can you explain in kind of like Eli 5 terms why the false positive issue is such a big issue and how you're addressing it? Absolutely. So I'll actually take one step back and talk a little bit about how you think about accuracy in a screen. And there are two measurements that he used. He used a measurement called sensitivity, which is how sensitive is this test for the disease that it's looking at? So a test with 100 % sensitivity will mean that if 100 people with cancer get the test, 100 people will have a positive result in the test. specificity is the second measure and it's the measure of how specific is the exam to the disease you're looking at so in other words i found something but is this something that i have found the disease that i'm looking for in this case i can't and so a good way to look about it to think about it isn't an alarm for a car you know if it's really sensitive you will counter a burglar but you also go off in the motorcycle.

13:50You know? And so when you tune sensitivity out, specificity goes down and vice versa. So designing a screening test is a function of finding the right blend of sensitivity and specificity. So with MRI, sensitivity is really, really high. It's like we're talking mid-90s sensitivity. It catches everything. And not just this, but MRI in general. That comes at the cost of specificity. cell specificity in MRI across the organs we look at, probably around 85%. So it has an inherent 15 % false positive rate. And so when you speak with doctors, they go like, well, 15 % false positive rate that will lead to unnecessary follow-up, anxiety, and so on.

14:37And so we've done a lot of work mitigating that. And we do a number of things in order to address risk of unnecessary follow-ups. The first one, and probably most important, is the Ezra scan has inherent follow-up on any filing because it's a direct consumer membership with a scan that you should get every year. So 90 plus percent of Ezra filings don't need any kind of follow-up because you have implicit follow-up and you get a scan again next year. Sorry to interrupt you, but this is one of the things that blew my mind the most, that you could have certain kind of like very early stage cancers, know about it.

15:11And even if that's the case, you're totally fine waiting a year to do the next screen to see if it's changed and moved. Like, can you unpack that a little bit? Because that blew my mind. I thought that if you found anything, it's like, all right, go in, get it taken out. I thought so too before I got into this space. But actually, a couple of interesting examples of prostate cancer in men. Most times, prostate cancers are really, really slow growing. And you are better off not doing anything about it, just monitoring. because the man is 65 and has prostate cancer, something else might kill him than the prostate cancer before the prostate cancer does.

15:46So the concept of active surveillance has become kind of quite big in prostate, especially because the prostate cancer shouldn't be touched most. They should just be monitored. Before Ezra, there was no way to easily monitor prostate cancer. You had to do a prostate biopsy, a PSA blood test, and so on, very unpleasant procedures. With SRA, you're in and out 20 minutes, the prostate and MRI, you can track it over. So those cancers, you definitely can wait and see and see how they progress. Another example is IPMNs, or interpapillary mucosal neoplasm, or essentially cysts that appear in the pancreas.

16:26And there's some literature to show that people with IPMNs are more likely to develop pancreatic cancer. but if you found it sound like pmn you don't yet have pancreatic cancer you are just more likely to have pancreatic cancer in the future and pancreatic cancer is especially bad because you generally only find it when you're symptomatic and if you're symptomatic it means that it expands the liver maybe other organs but kind of it's metastatic if you're a natural membrane we find an ipmn then you can do a scan every six months to make sure that you catch pancreatic cancers early as it happens so that you can have a better chance at treating it.

17:03There are some types of cancer where, you know, it is really important to do something about it as soon as you find it. Ovarian cancer is a prime example of that. Ovarian cancer can go from stage zero to stage four within a year. And so part of what we're doing at Ezra is developing these single organ scans that are more affordable so that a woman who has the BRCA gene mutation who has a higher risk of ovarian cancer can do a$700 scan every six months and thus giving herself the best ability to find ovarian cancer. So going back to your original question, you know, incidental findings, longitudinal monitoring is the key to not following up on all of these things.

17:49On the flip of that, we do a number of things. We use AI to generate reports to explain that these things, you don't have anxiety when you drop the report in your YSR dashboard. The AI also gives a score, an e-score to these findings that I talked about earlier, and we only recommend follow-up on e-scores 4 and 5, e-scores 3 just monitored, e-scores 1 and 2 are just observed over time. And then finally, our MRI scanning protocol is what's called a multi-parametric scanning protocol that uses different types of sequences in order to increase the specificity of the scan. And so all of these things brought together in that we've mitigated most of the incidental filings.

18:31And as a result, our false positive rate is actually really low. We're going to publish a study in Q4 of this year on our specificity, but it's looking like it's going to be lower than the typical screening procedures that you might already have approved with guidelines in the market. That's amazing. So if that's, if the false positive one is kind of the one big hesitation, the other that you're working actively on bring down is price, but I guess the other kind of lever that you have there is getting insurance companies to pay for this for people. What's the process like to get to payers paying for screens?

19:03I would imagine if they can be convinced that it catches things early, it saves them a ton of money down the line. Absolutely. It saves money and it improves outcomes. So payers generally are fans of screening modalities like pairs, cover mammographs, colonoscopies, kick checks, et cetera. To get there, you need to have the right price point. And based on our kind of calculations, the right price point is$500. And then you need to have the data to show that your test has high sensitivity, reasonably high specificity, so that you show that you're able to find cancers without too many full spotters.

19:41And so from day one, we've been collecting data with the goal of ultimately having a$500 scan and a sensitivity and specificity paper that we can go to pairs with and be like, you know, it's pulled together all of our organs. The incidence is really high across all these organs for cancer. We have a$500 test scan that addresses these organs. We have high sensitivity and specificity. Let's work together to pursue populations with a focus initially probably on a high risk. Got it. That makes a lot of sense. How do you figure all this out? So, I mean, I think your story is fascinating, but like to say that you're collecting the data from day one, that you knew you needed to get the price down to$500.

20:23You started as a software entrepreneur. You have made the transition into like one of these huge, hard, world-changing problems. How'd that transition happen? Like what got you into this and what lessons are there for other people who over the past decade in the bull market have made a bunch of money in software and want to go solve something really, really hard. The background, very quickly, I emerged from Romania. I studied applying mathematics and computer science in university. While at uni, I started a software company, very limited space than Ezra, in the advertising technology space, became one of the largest video ad servers in Europe, was acquired by an app tech company in New York City, and that ended up in the US.

20:59My motivation behind Ezra is that I'm personally at high risk of cancer, myself, for skin cancer. I've had cancer in my family. Sadly, my mother passed away from cancer, or something. It was because she found cancer late. And so I've always been very interested in the space. And many years ago, this is I think 2015, I was volunteering for a nonprofit in Romania who built hospices that care for cancer patients. And in being involved with them, I realized that the main reason why people end up in a hospice due to cancer is because they have cancer. And so decided to try to, you know, solve that problem.

21:38And I spent about 18 months doing research, but I cycled through about a dozen ideas on how you could go about screening for cancer from like liquid biopsies to DNA-based nanobots that you put in your bloodstream to all sorts of like crazy ideas. And I very much took the scientific method. I would just like read something, kind of hypothesis, read the papers about that topic, meet with scientists. And 11 times out of 12, but the full-body moral idea was literally the 12th idea. I would meet with scientists and they would be like, oh yeah, that's not going to work. Or it's going to work in like 10, 15, 20 years.

22:16Don't go down that path. And I just ran this process full times. And actually the idea, I don't know if I've ever mentioned, but the idea for Ezra came on my honeymoon. I remember reading papers. My wife was right next to me, and I was reading this paper that was comparing MRI with CT, ultrasound, and other types of imaging modalities for the purpose of screening. And I think it was for prostate cancer. And it was concluding overwhelmingly that MRI is just the better modality. And so I turned to my wife, and I was like, hey, if I could do a full-body MRI that was relatively affordable, would you get one?

22:52And it's great for cancer everywhere in the body. And she was like, absolutely, do that. And so that was the seed idea for Ezra. And then the rest was just figuring out whether AI could be employed to get there and then just going through the process of executing on. What was the moment that you kind of got the confidence? So you had the MRI idea on your honeymoon, dug in, dug in on the AI. Like, what was the point of which you're like, all right, cool. Everything makes sense. I'm going to go do this. Yeah, so I actually built a prototype for Prostate where there was a public data set put together by the NIH called the ProstateX data set.

23:28And I literally downloaded the ProstateX data set and started building a model to see whether you could find prostate lesions using AI within that data set. And this was maybe 2017 or so. And it worked so well that I was like, you can do this. and 2018, I let T and so prove and bought my company and raised our seed round with the prostate AI prototype as the proof that this can work. We then launched the scan, I think, prostate exam in January of 2019 and then we launched our full body late. Do you think, and this is probably a case by case and there's no right one answer, but in your specific case, do you think that you had an easier or harder time starting Azure because you came from the outside and weren't a radiologist or an oncologist or something from kind of within the medical community?

24:23I think it was easier to start as an outsider because I didn't have any preconceived notions of what is possible or not or doable or not. I think people within the industry would have said, oh, MRI, false positives, don't do it. It's going to be so bad. It's a lot. That's what we're, you know, four or five years later, probably going to have a full positive rate that's better than what is expected from a screen without and so i think people in the industry would have just taken it for granted that it kept up so and so i think it's actually useful to come from the outside provided that if you're doing something as technical as what we're doing at Ezra you have the kind of ability to build something to prove that it can work an interesting counter example is my very first idea with Ezra when I was going through the 12 idea 12 approaches was liquid biopsies like using DNA sequencing of DNA fragments and blood to identify cancer and I had this kind of concern that the sensitivity would never be very high because if you have a 2mm lesion in your body it's going to shed so little DNA in the bloodstream that you're going to literally be trying to find a needle in a haystack And so I met with a scientist at Caltech who was like, I don't think we're ever going to get the sensitivity to be what you need it to be for screening.

25:47Now we're seeing that fast forward. Some companies like Grail and others, pan cancer screening liquid biopsies are not really showing great promise from a sensitivity standpoint. They might be great from a specificity standpoint, so paired well with a full body MRI, but not as standalone screening. these things. And so I think it just takes a lot of effort to go down the right path. And that effort is all about kind of reading all the research, doing the work, and then hoping that you chose the right direction because this stuff is very. Yeah. Now, four or five years later, you have a great team, a killer scientific advisory board, like kind of everything that you want expertise wise.

26:27I guess like if you have the drive and you've done the research and you have that outsider's perspective that's core, but you also need people who have experience around you. How do you go about building that team? How do you go about figuring out what you need when you're not coming from inside the community and who you need? What's that process like? Yeah, so we actually have an incredible team and our chief scientist, Dr. Dan Sodexson, is the inventor of parallel imaging and MRI. The main reason why MRI scans are passed today is because in the 90s, Dan invented what's essentially, I think you said it in your article, GPUs for scanning, you know, as opposed to CPUs for scanning.