Following the Money: Expected and Unexpected Payoff for AI in Medicine by Isaac Z.Kohane | AAAI-20

So will now have our next speaker. Next picture is Professor Jackal Hannah from Harvard Medical School in Oak Grove chair of the Department of biomedical informatics and the professor of biomedical informatics Ed Howard medical school. He develops Enterprise computational techniques to address diseases multiple scales from Whole Health Care Systems as living laboratory is two functional genomics in your development with a focus on autism Zacks I have to be to project is currently deployed internationally 2 / 120 major academic Health

Center is where he drives discovery of methods in research and disease in pharmacovigilance and some of this war contributed to the creation of Black Box warnings that have the airports on drug labels on a more personal note a few years ago when we are just starting Health work with Google Harvard Medical School organized special course for us. It was called executive MD and In 4 weeks. We went to Harvard Vanguard urgent about what doctors are in 20 years of medical education out of this for weeks. I had the privilege of being told by Professor kahaniyan for a couple of hours to about 1%

of my medical education is used to Zach, so, please welcome Professor coconut. So thank you for the program committee for inviting me and it's a real pleasure to be here. Last time. I was at Triple-A. I was in 1993 where we were talking about very old school some techniques for Trend the detection. Let's see. So it's a pleasure to be here and I have two comments about the talk that just to proceed of mine. First of all on the behalf of all male mosquitoes. I want to protest the forcible separation from female mosquitoes second of all, I think you really heard

what's at the Crux of the car problem, which is the payment model. And so I'm going to give you a much more cynical discussion. I'm going to say what big and told you about is exactly what has to be fixed. I'm going to tell you what's going to happen. If it doesn't get fixed which in the near-term seems likely to me and how you personally can be very successful and helping humankind and be impactful in a variety of ways including dollars. And if this is not your cup

of tea, it's still it's my prediction for the overall system. So first of all, something is very is happening in terms of excitement about artificial intelligence. This is from 1951 to the present all the Publications in AI in healthcare. And so there's a big spike happening recently and that's could be because we're deluded about the impact that I have to say when I got my PhD in computer science and the 1980s. We thought we were I saw a I was going to solve

problems and we are roughly wrong and it came in politic to use the word artificial intelligence for about 10 to 15 years after that. I have reason to believe that things are different now. So, why are so many people excited? And where does the hike for short and where can we make a real difference in their term? So you just heard about this paper this paper where we were able to use Google Engineers were able to use a large Corpus of retinal images to achieve Supra

expert performance in recognizing retinopathy. And there were a number of very interesting things to this result including the fact that the only role of doctors in this whole paper was simply serving as labelers of the cases. That was it. They're inflexible contribution. And so when I told this paper, in fact, I wrote an editorial in in the Journal of American Medical Association about this paper when it would appear when I appeared. I called my very annoying cousin in Montreal opthamologist. I said cuz look what we have we can now

automatically recognize retinopathy. What do you feel about that? And his reaction was Zach? That's really interesting. I find looking at these renograms extremely boring. And instead of looking at them. I'm glad to have a GPU School cheap, you run it and for a few cents give me a reading and instead I'll go to the operating room, which I enjoy much more and I get paid a lot more so that it's learning me early on to a completely unintuitive resolved that the use of this technology at least in the developed

work my actual results in higher car because you can actually the surgeon can do the thing that's much more expensive than actually reading the red around and so everybody who said they cost when a cut cut Don't really understand what's going on in medicine. And what I hope to do today is to give you some better ideas at the same time. It is true that the healthcare system is running a little scared about what a I might do to the healthcare system and so are very a gust body the American Academy of medical colleges, which is the

association of old medical schools putting out articles, like will artificial intelligence replace doctors, but they're actually might be a reason why they are worried because with her the song before and when we ignored it, it came to bite us so When our illustrious leader here, I choose ja of Harvard University a public the paper a while ago saying that we needed to have a big investment and electronic health-record systems. I among many cheered this this notion. The problem was now having nothing to do with her. She's

that when it came to actual implementation in 2009. Initiative very happy with the Google at our with the Obama administration's investment in this area led by one of our colleagues David Blumenthal but then when it became clear in 2000. They were going to implant technology that was state-of-the-art. at best for 1980 I believe it's not even a joke and it's funny but it's not a joke the technology that underlies epic. The leading vendor is aligned with cold months. Anybody have an idea what the m and Mom stands for? Massachusetts General Hospital, this was a real

time operating system that worked in 1966 with a B tree base type hierarchical database and a real-time operating system on a pdp-10. I think that sounds great 1966 but in 2009, it was not the right technology and yet concrete was poured all over the existing Healthcare System to the tune of forty billion dollars of investment by the federal government and many multiples of that by the private sector and the prom was the technology was and is hard to use and there was no real requirement or at least no enforcement of requirement for interoperability. and

we actually wrote an article in 2009 and you can Journal of Medicine saying why can't we actually have why can't we actually have a health it being more like an iPhone where you have a modular architecture rather than the purposefully monolithic spaghetti code that underlies the system and course the guy was cast and so many Hospital Systems including the largest what in the Boston area invested think about it, 1.5 billion dollars. Think about how many startups you could actually doing the spacing? That's just one Healthcare System payment for this Enterprise software.

And this as my software has led, you know, I thought that Vivian was kind and talkin about the contributions to burnout. Because of the payment model documentation is incredibly important to Health Care visit and so doctors who are already squeezed for time, but they 10-15 minutes. 510 minutes are spent just hunting around screens and literally hitting thousands of buttons in order to have the patient counter while they're trying to take care of the patient. So it doesn't feel to the doctor like they're helping the patient. They're not to helping billing

and the patient is looking at a doctor staying away from the encounter and so it's a major contributor to a burnout and so I only race that issue to explain why medicine might be fearful of a faultless implementation of Technology. At the same time we see real Miracles like the fact that artificial intelligence research of keep redefining the boundary of what is artificial intelligence and because of our human centrism, we then dismiss it as no longer being the realm of intelligence but none the fact that

the fact that a reinforcement learning algorithm like alphazero could we learn all the rules of Go and then discard them on the way to World dominating performance suggested. Wow, maybe we could do that in medicine. Maybe we could have a reinforcement learning algorithm on the medicine that could learn all the aspects of medicine and then discard the worthless heuristics and then go on to discover better ones and it is this the future of Medicine. So I think it's very important in that context

to remember a few important points that andrej karpathy now a Tesla has made about the properties of the problem domain the past domain that Alpha zero has taken on so successfully that well, what are those? What are those characteristics the deterministic domain go is deterministic. It's fully observed. There are no hidden States and go to discover. The action space is discrete. It's not continuous. You can have access to a perfect Simulator the game itself to the effects of all actions are known. Each episode or game is relatively short

evaluation is clear and allows a lot of trial-and-error experience. and if shushan sets of human play that is not true of physiology. None of those things are true of of physiology. It's not the street. It's not deterministic. It's mostly unobserved. The game is long. Short evaluation is unclear and they're not huge data sets of human Play because I'm too except they exist and not widely shared. So all the big houses in Madison don't have those properties for

reinforcement learning discovered. So what area of medicine makes come to reinforcement learning anybody take a take a while to guess about that? What? Give those two people the prize they said billing and finance and that is absolutely right because where is the game clear the game clear and reinforcement learning in medicine around the rules this discrete rules off of payment billing and reimbursement. And so you can actually write rules to play that game

and explore all possible games and including who to how to deal with referrals a patient in and out of network all those are not only implementable as reinforcement learning algorithms, but they're the biggest point of money in the whole to one sixth of the economy that is health care, which is that particular task is at the center of all the fights in medicine. How much are bill for something and how much Tanya pay you of that bill for that for that service that's sounds incredibly cynical and it is but that is actually where the big dollars are.

and the more hopeful side of it of this. Well, I'm at will review exploration of a doctor's decision advice and maximizing painting utilities are much less likely to result in sustainable discipline in the near-term, especially because of a payment model that you just heard about So what's up colleagues of mine at the Harvard Law School and one of my grad students in science last year. I believe where we showed that adversarial attacks on the very successful image-based diagnostic algorithms

weather is looking at the skin lesions from melanoma or looking at x-rays for pneumonia or looking at retinas for Retinopathy adding a little bit a little bit of noise. That's so imperceptible that when you look at the image after adding of that noise, I cannot noises purposefully generated knowing what the convolutional neural network is actually doing but adding this noise makes the picture look exactly the same to human being but flips the algorithm saying there's a disease to not a disease and you can

Adjust the noise Otis slight miss or a near-miss are completeness and you could say was Zach other than the bad guys who would ever think of messing around with clinical data. Well, Both of you are not in medicine should know that in the basement of most large hospitals are teams are people whose only job is to look at the billing codes that are being generated the diagnosis. That means everybody a doctor's and saying what do I substitute to actually maximize income to the hospital?

And so they attack Of the hospital on the reimbursement algorithm. And the pears happens every day and every hospital is investing literally millions of dollars to play that game. So if you don't think that ml is going to be deployed very broadly as part of that game weather from taking Ml algorithm for left a melanoma. And how big is the legion is or the degree of retinopathy? That's a biomarker. It's a by Mark on the same sense of smell protein level in the blood can serve as an outcome in a trial. So if I get my algorithm just

slightly shifts the the outcome measure by 5% for my trial is Mike Trout from a billion dollar drug is a slam dunk if you think that's not going to happen. I don't think you're right. But there are plenty of Greenfield and that's where I'm going to live it less cynical where medical doctors would rather not tread. And those are the field where I think we can have real impact today. on patience and as an example, I give you this one child who was fine until Age 3 was walking was talking and then

over a six-month. How does a significant decline no longer walk to normal. Nowhere talk was referred to melt multiple top Healthcare Systems Top Doctors. No diagnosis. He was then hello. whoops He was then referred to the undiagnosed disease Network, which I have the privilege of being the principal investigator of the coordinating Center of this network and a pedestrian to you. But we wait for it when we proposed this in 2012. It was a kind of classic. I said 12 academic Health Centers Coast to

Coast will take all the genomic data of the patient and the clinical data both identified and put it into a hipaa-compliant cloud in this case doctors wherever this patient goes. So we created the network. And so if patients were accepted into the network to go to the Harvard portal they ask they apply with a spray simplification if their accepted as about 50% acceptance rate. Gino get sequence clinical get there to get some tape and this patient goes to the right doctor in the right institution whose expert this area and

sulfur example in this case of this child. Turns out that the child had a mutation in a gene cause gpp cycling hydrolyzed one, which you haven't heard of but I haven't heard of either before we get to it. Let me explain you the artificial intelligence or machine learning challenge. Here. We are we have millions of bearings in our genomes each one of us. Not only that but we have each one of us doesn't very upset. If you look to the medical textbook today is associated with disease. We're clearly not if you want to look up a word that I created them free product.

the incidental on the ohm of August Dental findings, that's part of the challenge we have but when you when you come through that list of disease possible disease generating mutations sorting that the list is a machine learning task where you have to use all other sources of knowledge to and to correctly prior type which are the few mutations your investigate and think hard about In order to be able to diagnose a patient this one this was a result of such a

pipeline working in conjunction with doctors who understand the disease and turns out this Gene codes for a protein which is important and making your transmitters in your brain and because of this child had a mutation is making less and less of these and so by giving him a cocktail of neurotransmitters right after that blockage right after that loss of function within a few months. This child was Walkin and talkin again. And this is remember after having been seen by lots of doctors and this is about a dollar issue caused from mine from United Healthcare. Tell me the cost of an average

undiagnosed patient is $18,000 per month cuz it keeps on going to the Health Care system and keep getting work the full workups. And this is out of date but we've seen well over fifteen hundred patients with me diagnosis for well over four hundred patients novel diagnosis and eight year. We had a new moon Journal paper where we summarized our success over the first two years and I want to call out to you a few things one is third of the African store Network already had a genome RX on some merely having the

measurement doesn't do the trick. They were still undiagnosed after their XO more genome was sequenced. The trick was the proper use of human intelligence add ml to focus correctly on the right. That makes a difference between able to make Defender diagnose and treat someone versus saying that they have a mutation or they don't that is irrelevant and not helping the patient and just so you can ask in case you're asking why do you know that you were more correct than other people? In addition to pack. The patient's responded when we thought they should respond

every single one of the mutations that we that was suspicious or questionable. We would use crispr Gene editing technique to knock in that a mutation into a fruit fly or a zebrafish to see if the cause of the significant phenotypic change. The other area where the one area there are many areas like that where doctors the average doctor's going to have confidence. So areas like genomics is a great application domain where doctors don't think they have the expertise and it's I think Greenfield for us to use our techniques. The other area is

where patient utilities that is what they're trying to maximize and dollars are lying here is a great example. This is one patient at all. The things that are being built for drugs and visits and procedures with inflammatory bowel disease a Bad Thing some of you may have it and that requires a lot of treatments how many treatments with monoclonal antibodies for example of cost easily couple hundred thousand dollars a year. In fact the leading drug. She would probably know this number better than I do but the leading drug Humira, I think it's several billion dollars a year just

for that one. That's being used. So we asked ourselves could we use look at the Healthcare System who use a VA Healthcare System at all the data that's being generated for each of those patients and just predict those patients who will not respond to the oldest drugs and will ultimately have to have their colon removed. Because that's time that their suffering that's time that they are not at work or go to school and it's all the time that the healthcare system is paying $100,000 for care and to make a very long story short. We seem to have good results

in predicting six months. If you have 6 months of a six-month window prior to that we can predict with a decent accuracy with a uses of round 98.9 those individuals who will not respond to therapy and will ultimately have a colectomy within 6 months. In closing I want you on ask you. What is this graph of anybody have an idea? What flu but thanks for being raped for answering and stackable for a pulse when I know what this is is weight. And this was a response to my mother is my mother is a situation where

four and a half years ago. She had been admitted twice to Brigham Women's Hospital for heart failure where the only thing she was in the hospital is because they need to get fluid out of her body had to have an intravenous director director called a Lasix and I then did something that all them from athletic literature has shown to be ineffective but I was desperate probably because and mostly because I care about my mom and because she was suffering and another admission to the hospital because Woman's Hospital is 100 Mi away from my office

and when she was in the hospital she expected me to be by her bedside. And so then my building to work wasn't getting on attack. So I did I did something happen shown not working randomized controlled studies, which is to give the patient a scale and then to modify medications based on the scale, but I have No options cuz I already had purchased for her a concierge doctor who has 400 Pace instead of 4,000 and he was not keeping her out of the hospital and she has big fat legs full of fluid and it was not looking good. Our scale and and

I'm praying as a pediatric endocrinologist adult International adult in his friends were horrified when I told him that I applied this very fancy machine learning algorithm that goes like this if weight is greater than weight Yesterday by 1 lb add one additional pill of Lasix. That's it. And what happened was? It worked. So she was going again and get the hospital until I started this and her. Feet with her legs Getting Thinner and thinner free of water. And she had many years after that completely free of the symptoms of

heart failure. And so I wrote a tongue-in-cheek article for a local national public radio station called what my ninety-year-old mother taught me about the future of AI in healthcare, but here are really the lesson. One is who completes the loop who is trusted my mother? Who when she wants to use foul language or lapses into Russian would not accept initially my recommendation to even get on the scale. So I had to get on the scale get on the scale. I would look on this Fitbit scale it cuz

it was connected Wi-Fi whether she got it or not. Then Mom to please take a pill and no you didn't you didn't take that Lasix Lasix pill and I got to do it at to the point that within a few months. It was autopilot and with her everyday, who is the best with the recipe I was pulling up her call her doctor to make sure that he was tracking what was going on. I was text aporte if the the scale of the lost its Wi-Fi I was doing that. and who does the expert

catch what I mean by that so I thought you heard my algorithm, which I should publishing AAA I but then I saw this. Her weight was going down. I see you're even better than you thought, you know, it's going down. and then I did I realized I had done when all doctors do don't listen to a patient while not she had told me that she was Getting up at night to pee and sometimes was having accidents, but we can cut it and I know what I said. Yeah, I was thinking and

then I realized that something else is going on when I saw this guy went to the CVS Pharmacy and got what's called a test strip to dip in your urine and sure enough there was sugar in it and she's getting diabetes and she was peeing out the calories and to have something to her urine. That's what I do. I just gave her a pill of Metformin and f&f in stabilizers. Stop peeing at night and I wake stabilized but think about all these things the AI algorithm

is just a very small part of it. All these other things have to happen in a real-world for real patients to be helped and so many tasks that can be done by the algorithm for the same time. The the I'm not about the human implementation question. It's absolutely essential to the success here at with that. I want to thank you and give you a free advertisement to a conference that were holding in Bermuda. This April where we have major players from the Big Data

companies from the New England Journal of Medicine talking about what it looks like when a I try looks like and where we have practitioners who actually have Programs that have actually made it into the clinic will be Talkin at without. Thank you very much for your attention. one question Gaufrette I saw you say about revenue maximization is quite horrifying. What I heard utilization maximization Revenue maximization, right parts is better for me. Thank you very much. Well, it goes back to

to Vivian's right now. The revenue attack on the system is true of all health care. It's done by the drug companies. It's from some of the hospital everybody. We're all trying to maximize our pull out a dollar. The way to get it to get there that those utilities on a system aligned with a patient is to go to value-based care, which is what a business that is extremely hard and Ashish has shown how simple short-term measures that people can bend game don't really contribute to improving the outcome or really improving

the finances or you have to have as much more holistic approach to measuring how your populations doing and although we're making some stuff there. We're quite far from it. And so that explains why some very smart people when they've heard me speak exactly was popping but a different audience have come to me and think I'm a VC. I think this area about gaming the system is where I want him Best Buy dollars. Alright, thank you. Thank you.