"I just want AI to replace me as a scientist" | Diagnostic Robotics' co-founder predicts the future

In three months we did something that used to take 20 years. And if you ask me where the world is going with AI, I think the industry that is going to be influenced the most is those traditional chemistry and biology industries.

I'm going to be a scientist who tries to make science go faster.

What I envision specifically in healthcare is almost bots talking to bots passing information.

I'm not trying to automate everything. Like when you say automate drug discovery, I'm not going to discover everything. I just want to accelerate it, which comes back to my childhood dream. I just didn't want to do it myself. I just want AI to replace me as a scientist. That's it.

Well, my story goes way back. So I will start from my parents. As you can hear very quickly from my accent, I wasn't born in the states. I was born in Kiev of Ukraine and I was born in a time where Ukraine was part of USSR and in general being part of USSR means that you have very little freedom in almost everything. We come from a Jewish family and if you're Jewish it means there are certain universities you cannot get accepted to. There's certain limitations on the type of works that you can do. And after the nuclear reactor has exploded and of course the government didn't tell the people about that. They had to hear it from the radio from Norway. My mom was I think in the last train that was allowed to leave Kiev because they were afraid from radiation and they didn't want people to run away. And around '89 and '90s USSR started falling apart and there was a lot of pressure coming in from the United States to at least let the Jewish community leave. And the USSR went out with an announcement saying that if you're Jewish, you are allowed to leave. And the main message was you were never citizens to begin with. So, you're not immigrating, you're repatriating. In other words, coming back from wherever you came from. And my family was one of those families that was allowed to leave. Everybody knew there's going to war going to start in Israel. Remember the Gulf War in 1990?

And my mom was trying to predict almost a year in advance when to come to Israel. So, we wouldn't come during a war. Well, my mom was very precise. She almost predicted the time of the war. But unfortunately for us, we came two weeks before instead of two weeks after. So I always say that predicting the right timing has been always a big pain point in our family. And since I can remember myself since I was I think five, my dream was to be a scientist. That's all I wanted. I wanted to progress the knowledge of humanity just a little bit forward. I thought for many years I'm going to be doing it in healthcare. And when I was 14, I found a program in the Technion. It's one of the largest universities in Israel where they're allowing students to join PhDs in their research and I joined research mostly in cancer, prostate cancer, not because I selected the topic it's just what they were studying and part of my work was feeding those cells every day and they had one request when you feed them don't touch with the pipet, don't touch the cells you can get infection in. Unfortunately like a week in or like two weeks in I touched it. Then it got infection and they sent us to one of the hospitals to get them new cells. But since that point they would never allow me to do any biological experiments. Nothing. Don't touch anything. And what they let me do is instead the analysis. You have to count how many cells you have. I'm talking about counting till 50,000. I said I'm not doing that. That's going to be the worst summer of my life. And I think at that point I understood that I'm going to be a scientist, but I'm going to be the scientist who tries to make science go faster as opposed to the person who actually does the work itself. And I was really excited about everything that's going on. And I thought about like, hey, wouldn't that be cool to take all of those data that you have and try to predict future events? And I started building a system that tries to take 150 years of news articles using natural language processing to read it, identify patterns and predict future events. And that's the work I was doing in Microsoft.

Well, I wanted to do my PhD in natural language processing. And while I was playing with this, I was thinking, hey, it was almost like a joke. Okay, at the beginning we're thinking about building semantic analysis looking at whether an AI system can know if two words are similar or not. But then I was like hey why do boring things why not predict the future? It's almost like a joke and then you know I built like a very simple statistical system to do it and I was like it's working. So, one of the things I was trying to predict, there was like thousands of birds in Beebe, Arkansas who just like fell out of the sky, all dead, no signs of disease, nothing. And I looked at our system and it found a correlation, a very interesting one, that it's not the first time that this is happening. It happened before in Norway and in Texas and always the same time, 6 months before people are looking for oil spills and it was just statistically significant. Now what I was thinking to myself is well if you have an oil spill now I do understand if birds will die okay or animals or fish but why does it take 6 months and why when they found them they found no signs of disease no find of no sign of oil so I thought that we had something that is pretty cool right like big scientific discovery and when I asked my friend she's she was a biologist she told me well this is obvious I was like oh tell me more and she says okay oil spills get the oxygen in the water to a much lower level, oxygen depletion. And then when you have less oxygen, those animals literally suffocate to death and you wouldn't find any sign of disease. They you wouldn't find them to actually starve. It's almost a hidden cause. So this is where it got me thinking, well, what did she have in her head that my system did not have? And what I realized is unfortunately most of you are not searching on Google oxygen depletion. Unfortunately, so I couldn't have enough signal about that. And she had a lot of causal knowledge. In other words, she read it and she understood there's a causality between oil spill and oxygen depletion. And then I started thinking about well, how would I take everything humanity knows about causality and mine it in order to create this causality graph, everything that we know about causality. So I started with a very simple system trying to see if somebody said on the news or then social media and then eventually developed to like Twitter and queries and everything A causes B and then it became something happens after and with time we started building machine learning classifiers to actually predict causality or we built at that time it's considered big a graph of 300 million nodes and billions of edges representing causality that was already verified by treatment control group experiments historically. Based on common knowledge and we started fitting that into correlation machine learning algorithms. So for example, how did we predict the cholera outbreak? The causality graph we already knew that if you have floods that might cause cholera, why cholera is a waterborne disease. So the causality here is clear but we all know that not all floods are going to cause cholera. There's floods all the time. And then the system was started to look for correlations trying to identify places where it does cause as opposed to ones that are not. And it identified the following pattern. If you have a low percentage of water as compared to land, the probability of this one is very big. And an additional one is that you need to have the word drought in the news two years before. This is very weird, right? You have drought 2 years later you have a flood and then after this flood there's a high probability of cholera. And this is what helped us predict numerous cholera outbreaks including one in Cuba it was the first one in 130 years that was a very big thing. What I love about this example and if we talk about causality I think this is my main insight is there's something you can do about it let's say I predict it cholera is very simple to treat if you get clean water in time if you reduce if you get clean water in time you reduce mortality rates from 50% to less than one so all you need to do is get clean water.

So for many years I've been deploying those systems mostly in predicting economic events at the beginning I started a company I sold it to eBay and eBay were doing this mostly predicting commercial events or commerce events I would say on the eBay platform and then I wanted to go back to my roots. I wanted you know I wanted to do something in healthcare like something in biology or chemistry. I was just never good at that. And I thought about kind of what's the biggest problem we see today in healthcare and I started listening to problems in physicians and I think the biggest problem everybody was worried about is that till 2030 we're going to have 4 billion people on earth without access to primary care. What does it mean? It means that if today we're waiting three hours on average inside emergency departments, it's going to become eight and more like it is today for example in China. And what's the best way of solving it? One way is just to automate. You know, it's like we're going to automate as much as possible from the primary care physician war, but how much can you automate? Today, you have, you know, like one primary care physician per 2,000 patients. In certain areas, it's one per 5,000. It just doesn't scale. And I think most people would understand that it's about proactive care is if we can get less people sick which is a lot of the work of the primary care physician then we can actually control the system. And I was lucky enough to get introduced to Elevance and through a partnership they gave us access to more than 60 billion claims. We received additional access to almost 100 million medical records through different other partners and we started building AI algorithms attempting to predict which patients will deteriorate in the next 6 to 12 months. So that was a very cool experiment, right? And we're really good in predicting it. And now we're coming back to the biggest question. So what, right? Yeah, they're going to deteriorate. So this started me through a journey. First of all, how do you even define what's preventable? Luckily enough for me, somebody else answered that question. In 2007, there was a very nice paper published by Beth Israel. They had 60 physicians manually look at codes in inpatient visits and manually categorize them into preventable or not preventable. Next question is okay what do you want me to do about this? Tell me the secret. So this got me into the field of causal inference. Causal inference was a very hot field of study approximately 10 years ago. Somebody even got the Turing award based on this. This is Judea Pearl's work. And what we're trying to do is take specialists and ask them to model in a graph the knowledge they have in their head in a causal manner and based on this information try to do a lot of simulations. If you talk to clinicians the way they talk about this they say we don't believe anything that did not happen in a control trial but I cannot do infinite amount of control trials. But if I had to summarize this very high level, the way we do it is for each patient we find their twin. Somebody who looks exactly like them like same age, same social determinants of health, lives in the same area, same diseases and they just have different physicians. One physician did A another physician did B and then we historically can see what happened. And if you do it with enough data, that's a very good way of mimicking the control trial. And we then we started seeing based on our historical data, what's the best clinical next step to do in order to proactively prevent deterioration of patients. I'll give you an example. Some of the examples are very simple. If the patient stops taking medication, that's not good, right? You need to help them with that. The caution is usually patients stop taking more than one medication. You have 15 minutes with your physician, which one do I fight for? Okay, they don't take 10 of their medications. Well, what do I do? So, the system even ranks the medication. And I'll give you a very nice example. It turns out that we had a patient that had hypertension, diabetes, CKD, a lot of diseases. And he stopped taking his depression medication. And very strangely the system ranked the depression medications higher than all of the other hypertensive drugs which is very counterintuitive to clinicians. Clinicians are still thinking about it very clinically and are not looking the mental behavior of the patient. And the explanation the algorithm even gave was well if he's depressed he will stop taking all of his other medications as well. So if you have to spend your 15 minutes spend your 15 minutes here as opposed to others.

I think one of our main discoveries here, just as an example, we noticed that patients who have some hypertension problems, they're taking PPIs. Think about this almost a drug for upset stomach. You see their blood pressure drops. And if you compare to the control group, there's no explanation. It's literally because of those PPI. Again, it's not a full control trial. it's mimicked but still you see a very big difference. So we showed this to the physician and he's like ah so I don't believe in that like until you explain me why I'm not going to look at that. I was like I don't know I see a correlation it's very statistically significant. So what that physician did spent almost two weeks looking at papers till she managed to convince herself that she's correct. So the explanation was she says oh okay patients who are taking this drugs they probably sleep better if they sleep better there's other studies to show their blood pressure drops. But she wouldn't even go into a control trial until she felt comfortable that the literature supports this and a lot of the explainability of machine learning models today is like explaining it from the data that you have very rarely we're trying to explain from data like from the literature or like external data. So what we then did we built a system that reads this medical literature every time we identify correlation it tries to explain it. So one of the main discoveries we made we were studying hypertension because it's a very important disease like 40% of the population under the age of 50 is being treated with some hypertension drugs. And we identified that patients who are taking beta blockers. Beta blockers is a family of medications as compared to other drugs for hypertension like ACE inhibitors and others 15 years later have an increased risk for Parkinson's and increased risk by almost 40%. Now there's like oh my god this is super interesting but once you give them all of this explanation physicians like oh interesting and if you could get the biological explanation that's even more amazing right. So my biggest breakthrough was first of all explaining only from your data is not enough you have to explain from external data especially when you work with healthcare people they wouldn't like really understand the science a lot of them are scientists themselves and then connecting it to an action using the causal inference.