Precisely Practicing Medicine with a Trillion Points of Data | LSI USA ‘23

This keynote showcases how to navigate the potential of leveraging big data analytics to improve the accuracy and effectiveness of medical diagnoses and treatments.
Atul Butte
Atul Butte
Chief Data Scientist, University of California Health System


Scott Pantel  0:08  

It's my pleasure to introduce Dr. Atul Butte. He's our keynote speaker, Atul as a founder. He's a scientist and engineer and innovator. He's got multiple exits. Here's the Priscilla Chan and Mark Zuckerberg, distinguished professor and Inaugural Director of the baker, computational Health Sciences Institute. He's going to talk about a lot of things today. But he's going to talk about data, he's going to get us inspired. And it's such an honor to have him here, Atul.


Atul Butte  0:36  

Thanks for having me, I get you for the next stuff. 30 or 40 minutes here, it's a real thrill to be up here. First, I'm a medical doctor, and I represent a provider network, the University of California, I'll explain what that is in a moment. But since I'm a medical doctor, I got to start with my conflicts of interest, I have just a few. Actually, I have a bunch. On the top left are the companies I've started I'll talk about those, I consult for almost every major pharma and biotech. So I wouldn't blame me if you didn't believe in other word I said over the next 40 minutes. But I am most proud of the right side of the slide. Those are all the companies started by my students. More than half my grad students now start companies, even if they go into academia, and they do it with the most amazing platform in the world. And it's simply data. It's often big data. It's sometimes open Big Data, I'm going to show you how they did it. I'm going to show you how I did it. And maybe I'll convince you, despite all the bad news, this is still the most amazing time to be in innovation and entrepreneurship. So with that, we're going to be definitely talking about data. You know, you cannot pick up any major magazine now and see a cover like this, okay? Where there's data and it's big data might be small data, it's open data, it's AI, and machine learning. In fact, the economist covers it again, and again, right, you see two covers down there, the data delusion. And then on the bottom right there, data is the world's most valuable resource. Now, there's something interesting about that cover, because they're equating it to oil rigs, right? Do you see that? And that, to me is very concerning. Because we hear that all the time by data is the new oil, right? Maybe you've heard that, I hate that sake. Because when you think about oil, you think about barrels of oil. And either I have that barrel of oil, or you have that barrel of oil, we cannot both have the same barrel of oil, data is completely different. I might take a dataset and make a diagnostic with it, you might take a data set and figure out a therapeutic from it right, we can have the same data and create many different things. And so I think that's the point about data, it's divisible and doesn't lose any intrinsic value. Now, of course, we have data and huge data sets. in biomedicine as well, here are just some of the logos of all the different sources of data. Of course, you see epic, there's an EHR, we'll talk a lot about that. But you see major cohorts the UK Biobank studying half a million individuals, the All of us Research Program, studying a million individuals in the United States, and many, many other datasets. I'm gonna start the talk with very just three simple slides, how I use this data to actually create some value create some companies here. The first of these is a company called New Medi. So what we realized is, so many people write because of NIH mandates, and the Gates Foundation and many other funders, if you create a dataset, as a scientist, now you have to give that away on the internet. And what we realized is so many people are giving out datasets on diseases. And so many people are giving out data sets on drugs, that we could just put these two together, maybe this drug might work for this disease, right? Put these two disparate data sets together. And you know, you turn this crank, you got a lot of ideas, this drug could work and that drug could work. But where are the fun ones where were new uses for old drugs, right drug repositioning, as we turn that crank quite a bit. And I'm really proud of New Medi raise now close to $10 million. Because of course, you cannot keep these in your lab, you got to start these companies, now focusing on idiopathic pulmonary fibrosis, a nasty condition that has two drugs, neither of which are that great for patients. So that's example, and I tell my lab all the time, if you want to change the world, you can't just keep writing papers about it. The science continues in the startup companies. Another example is Carmenta. We wanted to come up with a new diagnostic for an incredibly nasty condition called preeclampsia. When blood pressure goes shooting high in pregnant women, it leads to lots of mortality, morbidity all across the world, every socioeconomic status, every race and ethnicity. But we realized so many people were studying preeclampsia with all these molecular tools. Why don't we just go download all their datasets they give them away and figure out what's in common. I don't care if that researcher saw something that researcher saw something. I'm gonna chase down what every researcher saw. And let's figure out how to make a serum blood test a blood test for preeclampsia, because right now the test that we use is one of the most nonspecific tests, we have urine protein, not even a specific one. It's we turned our crank and we came up with some diagnostics in the middle. Or what did we do in Silicon Valley, we launched a company called Carmenta. Raise $2 million in seed financing. Now, look, I'm not going to brag about $2 million. As you know, in Silicon Valley, that's not a lot of money to raise. But at the bottom right there, you can see what happened next, the company was already acquired. So we went from downloading data, to selling the company in 18 months. Right? Inventors happy investors happy, universities happy. And I'm giving away the secrets here, right? All of these links work, they tell you the story, because every one of you can come up with a new diagnostic for a different disease, we'd never step on each other's toes, we need that many diagnostics. In medicine today. Right? That's the story of Carmenta acquired by Projeninty. The, boy, the next story. The third story is even more amazing to me. When we realized in 2008 was someday a patient's going to show up in our medical system with a genome, right one of these DNA sequences, and we were so not ready for that. And this one really started with a high school kid in San Jose. But we say you want there's no master list of all the spots in the DNA and all the diseases they've been associated with, let's just start making a list. And a really smart San Jose, San Jose high school kid could read 50 papers in a summer, and curate them right read them and forgot this snip. And this spot in the DNA meant this disease and this odds ratio. We said if a smart kid can do this 50 papers in a summer, let's just go read every single paper in genetics. And we hired on a massive team in India to just read every single paper in genetics. And we made the master list of just every single spot of the DNA, every single disease and odds ratio p value, turn that crank, what do you do in Silicon Valley, you launch a company. Now this became Personalis. And in 2019, the IP owed we IP owed at the peak, of course, everything's down from the peak, we reached a $2 billion market cap. And this started with a San Jose high school kid who was reading papers to build a curation right just to build that database. Nothing I just showed you, these three companies needs magic computers. In fact, we don't even have the computers anymore. We rent them from Amazon, right? Amazon cloud, right? The data is out there, the papers are out there. It's up to you to realize this is coming. What can I do about it now to address that need in three to four to five years. And that's no different today. No different today. Now let's think about those. But then let's focus on to me the next big data, electronic health records. EHRs are so old, they're new again. Okay, because everyone has them. Here are three press releases illustrating how much money we spend on electronic health records. Sider is a health system in the United in California, our neighbors here, up and down California spent $1 billion on their EHR system, you see that in the back there, partners, which is MassGeneral, and Brigham and Women's Hospital spent $1.2 billion on their EHR system. And Kaiser Permanente spent $4 billion on their EHR system. Now, amazingly, they all bought the same system, epic, okay. And the price tag goes between 1 billion and $4 billion. Now, it's not just for the software, it's for training and training the trainers and all these trips to Verona, Wisconsin, and all the badges and certificates, you get and redesigning workflow, the boy, the narrative I want to communicate here is this is now the most expensive data in America. We're paying doctors to type all of this in. And really the take home point is it will be a national tragedy, given how many billions we spent here, if we don't use this data to improve the practice of medicine. Of course, we have to do this safely, responsibly, respectfully, but boy, what a tragedy will be if we don't use this to improve the practice of medicine. And that's what we're talking about here. And with that, let me just reintroduce where I come from. I represent the University of California, right, you're surrounded by UC campuses here, you might have passed UC Irvine, on your way from the airport. University, California is massive. We have 10 campuses, three national labs, including Lawrence Livermore, which I think as the number two or number three supercomputer in the country, if not the world. We have 200,000 employees who are one of the largest employers of the United States, a quarter million students per year. And we have six medical schools and thus six academic health centers. That's UCSF, UCLA, Irvine, Davis, San Diego and Riverside. We have 14 other health professional schools, nursing, pharmacy, dental, veterinary public health, we train half the medical students and residents in California. We got about 10% of the NIH extramural budget. And coincidentally, we write 10% of all scientific papers in the United States all scientific productivity. Well, it's ecology to the economy. 10% of all papers written by the University of California, it's about 13 to 14 billion in clinical operating revenue. So I think Places is now number eight in the country for in terms of revenue or for health systems. Number eight size 5000 doctors get a paycheck from us every month, but 100,000 Doctors write orders on our patients every year. It's all the residents, all the fellows, and UCSF and UCLA and the US News top 10. But that one's an interesting one, because everyone else is leaving that list. So we're moving up no matter what we do, until we decide to leave that list we're still in there. So Olson, we're getting better, no fault of our own. And we have five NCI comprehensive cancer centers and five NIH Clinical Translational Science Awards to take from discoveries to products. And then we got the secret superpowers that are so secret, even we forget we have them. IRB reliance, right? If one of our medical schools approves an IRB, the other five automagically approve them too. Okay. Now what that really means, instead of weeks to months, it might take days, two weeks, right. But once one IRB approves that, then we can rely on that one, and centralized contracting, which some of you might know if the University of California Office of the President signs a contract, the other campuses will agree no renegotiation needed, what an amazing partner to help you scale, right, because we can just scale like crazy. We have friends on every campus here. And what we now have is this umbrella across a health enterprise called the University of California Health, UC Health. That's what I'll be explaining. Now, UC Health really was started more than 10 years ago, but really came out, in my view stealth mode in 2016, because we announced this declaration that we want to build a single Accountable Care Organization for the University of California. Now, this may happen someday, five to 10, to 15 years ago, and we partner with United Healthcare to learn how to take on risk, right? If you don't learn how to do this, you lose money in a hurry. But the idea is we're going to have to build we want to build a single Accountable Care Organization for the University of California. Now, the first thing is when you make a decision like that, you got to ask yourself, well, UCSF does it this way, and UCLA does it that way, which is the right way, we're going to take care of liver cancer, transplants, even pneumonia, right or diabetes. And all sudden you realize we got to figure out the right UC way to take care of everything. Boy, it's a lot easier to figure that out, when you have a central database of all the clinical data in one place. And that's what we ended up building. And that's what I'll be talking about here. So I wouldn't be an IT guy, if I didn't know how to make boxes, point the boxes on PowerPoint. So here's the slide that does that. And here you can see the logos at the bottom. So UCSF and UCLA are the two kind of big gorillas there. Irvine and Davis are smaller. San Diego's in the middle, and Riversides are baby siblings. All right, they're a brand new medical school, they don't have a hospital system yet. They have some clinics here, but I'm going to show them just like the rest. So each of these six has their own data warehouse, same common data model. And every month, every two to four weeks, all that data moves Central, you know, we were happy with every two to four weeks until COVID hit. And during COVID. We moved all that data nightly, every single night was going into the central database here, right central database. So what does that mean? Okay, so this is a little bit of a busy slide here. But these are the latest numbers look like. We're just coming up on 9 million patients now that we've treated over the last 11 years. And this is literally every dose of every drug, every vital sign every pain score, every respiratory rate for 9 million patients, you can see some of the finer details there. 378 million encounters 1 billion procedures 1.3 billion medications we've ordered or prescribed, relevant to this audience 44 million medical devices we've used. I'll give you some more details in a moment. 1.4 million providers over 11 years, 1.1 billion diagnosis codes, and 5.2 billion lab tests are vital signs. Another way I like to explain this, we have everything from Tylenol, to car T cells. Tylenol is probably the cheapest thing we can offer to a patient right tablet of Tylenol will still charge quite a bit for that. But that's why the cheapest thing we can do to car T cells, well, we'll take out your white blood cells, train them to fight cancer and put them back in right to fight that cancer for you. That's why the most expensive thing we do and we have everything in between, right, so we're not just primary care. We're not just tertiary quaternary care about 800,000 or more of these patients with primary care that we follow. It's merged with state regulatory data pathology, radiology text, we can barely run the death index for state of California. So we know everyone who dies in the state of California merged with our database. And all of this is safe, respected, regulated, responsible use of clinical data. Now, let me just give you some examples. If you just did the potassium measurement, you know, we get serum or plasma potassium measurements, here's 24 million of them, okay. 24 million serum potassium measurements. Luckily, most around four or five or six, there's some here's zero and some are at 10. Those are not so good for human life. Hopefully we did something about it and we can tell if we did something about it right because we have all the medical records there. That's the kind of easier one a hard What is this one? almost 1.4 million Birad scores. Right now this comes from the the mammogram that we do that generates a text report that we parse out every buyer adds every ejection fraction from every echo every mammogram across the entire University of California, right? So there's 1.4 million bired scores, looking, these are for breast cancer screening, for example, right? When we need to get to the text, we can parse out all of those numbers from the text still. So I'm going to argue probably there is nowhere else in United States where six academic health centers, bulk data share, like we do constantly, like we do in University Counseling and make it relevant for you medical device, folks, I looked in the hotel room, we have the Edward Sapiens Val, for example, we've used 1544 times the Zimmer personalized knee, which is super expensive, I will tell you, I'm not sure why I'm not a surgeon, particularly a pediatrician, we use 56 times. But here's where it gets fun. The Zimmer 75 millimeter by 3.2 millimeter tracer Tip Drill pack, we use 480 times, we track every screw of every kit for every device, because you have to still order it in epic. And boy, are we great a billing for those things. And all of that data is sitting there sitting there waiting for some interesting scientific questions. The Medtronic Bovie pad, I think that's for grounding right? 130,615 times the Ethercon proximate stapler, we've used 59,629 times we track 119,000 individual device elements, okay. In epic across the whole UC Health System. Probably your devices are in our system, or want to be in our system, right? But if you've already gotten them approved, chances are we're using it somewhere. We actually know if it's working or not. And all that data is sitting there. What can we do with that together. So this is what it looks like. Of course, it starts as an identifiable database, right, we have every zip code every address, right? We're using this for quality improvement. So you're about to 7.6 million patients in California, northern California is UCSF and UC Davis, Southern California is UC Irvine, UCLA, UC San Diego, and there's a tiny little drop dot for UC, Riverside. You see age, race, gender, ethnicity, of course, that's common primary care, not primary care. But because we can geocode every single patient, we know the characteristics of where they live, the tip of the top of social determinants of health, the social vulnerability index, for example, it's it's a rural area without a lot of resources, an urban area with a lot of resources is an urban area with few resources, right, we have all of those kinds of elements that we can track that I'm not saying we've solved social determinants of health. But it's actually the first step that we can do all of that they're the first thing before I go into detail what we do with all this data, I want to point out, we give all of this data to patients, okay, every good health system should do this. Now, if you really want to learn how to do this, if you go see a doctor regularly, if you have an iPhone, you should just click on that little heart icon, you thought it was tracking your steps and sleep, it actually has a little tab there for medical records, you can put in your user ID and password. And that's a great way to actually get all your medical records on your phone. If you got care of five different places, you got one timeline, with all of those elements in there, Apple doesn't have it, they just gives it to you on the phone. What we realized is not everyone can afford an iPhone. So we're working with CommonHealth, to do this with them for Android as well. So I want to make sure you all get that we have to enable patients with this data. It's not just a cool thing to do. It's federal law at this point. But it's still amazing how many health systems you might go to that aren't even on this list. Not sure what that says about them. But I'm telling you, we're going to share this data with patients, including notes now, right, all those op notes, right, all those surgical notes, all those discharge notes are all available to patients matching what they should or shouldn't be doing with them. Alright, so a lot of operational teams have been benefiting from this data, saving millions of dollars. I'm going to give you a couple of examples here in a moment. But because we have central tools improve the quality of care, we can decrease some unnecessary drugs, centralised population health management. I'll go into a couple of details here. This slide is deliberately blurred out. These are the top 10 drugs we use on each of the campuses. So for example, up in the gray, there is UCSF, and the top left little corner there is a particular drug for breast cancer. I'm not going to say the name and I'm not gonna say Well, I'll tell you about that little square there is about $200 million. This is just a charges from 2022 charges. This entire slides about $4 billion. So I'm showing you $4 billion on one slide. Now top 10 drugs as you'd guess most of these are biologics right? No surprise here. They're pretty expensive. But then why aren't they the same biologics? Why does for example for anti TNF alpha what is UCLA use a lot of this one and UCSF use a lot of that one. which is the right one to use. And all of a sudden, now it's in our interest to do the comparative effective. So we don't need a grant from a quarry to do this, we want to do this. And then we can do something super magical. Let's buy them together. Right? Something we call leveraging scale for value. Something I think Medicare is just learning how to do. We can put in one purchase order and get a better price. Because we can count and start to standardize on some of this right? It's no secret some miracle here. The minute you start to look at this, you start to realize what you could do with the data together. We'll be doing a lot more of these. And other examples My favorite example, IV Tylenol not actually found, well, that's a brand name IV acetaminophen. Yeah, everyone knows acetaminophen right now, it's pennies for a pill. But the IV dose hopefully this company is not here. There's one caveat makes him United States. IV Tylenol IV Acetaminophen is $50 a dose you crack the vial? You bought the whole vial? Okay. 50 5050 5050. Now, there are appropriate uses for IV acetaminophen, right? Maybe it's pre op or post op, right? You can take anything by mouth. Maybe it's a kid, a child pediatrics, maybe it's an opioid, you're trying to avoid opioids here, right? But what we realized when we look our database, and boy, so many patients are getting IV acetaminophen, at the same time as other oral pills. Why couldn't they get the acetaminophen by pill? Because it's 50 50 50 $50 each one, right? We're not paying any differently for using this drug, P pre negotiated DRG. This is an absolute waste of money for us, right? All I can do is show you the curve. It's the Hawthorne effect, you start to look at something it gets better, right, you should write down. Of course working with the chief pharmacy officers who came up with this project, the pharmacist who make it harder to order this in epic and elsewhere. This example, we're just this one drug Saves us millions. Because we don't need these IV form of this drug. Right, we can use the oral form. Expect to see more of this happening at University of California. All right, here's one dashboard for all of our patients. Now with type two diabetes. Just to give you example, how in one dashboard, we can manage entire courts, it's nearly 50,000 patients on this dashboard 48,700 or so this data as of last month, we update these things monthly now, you can see most of followed by a primary care doc from us, but some see a specialist and endocrinologist. We have the latest hemoglobin A1C and we have the agency changes on the right it hasn't been increasing or decreasing over the holidays, a lot of A1Cs go up when that's the blood test. We used to track type two diabetes. And all on the bottom you can see are they getting optimal diabetes care? What is their eye health, kidney health, foot health? Are we even looking at these, we track all of this for diabetes. Now for hypertension, we're working on lipids. So on one dashboard, you have five academic medical centers here, right? Nobody else it's never going to happen in Boston. It's never gonna happen in New York City, because they compete with each other. And we're far enough apart that we don't have to compete with each other. So I love being at the University of California. Now our cancer centers for those who have cancer devices, and cancer therapeutics are all working together. All five of our NCI cancer centers have now one big cancer consortia. And here I can just show you 105,000 Cancer Center patients every year, our best guess that's tripled or quadrupled the number one in the country, that's MD Anderson, we know we're stronger together when we work together. And even here we have advanced cancer patients that are getting immuno oncology in car T cells, about 31,000 of them, where do they lived the age, gender, all of that here. By the way, we also have their cancer genomes, because we do so many DNA sequences now on cancers, the same database as all the cancer genomes, I think Roche paid $4 billion for for flat iron and Foundation Medicine to get those two together. We definitely have that kind of database now here too. And we use this a lot in partners looking for new targeted therapeutics with pharma and biotechs. Now, I'm gonna just really bring this to the next step here. And now I look I'm a researcher, I write papers and grants, I still gotta write those papers, publish or perish. So I used to get a lot of my inspiration from NIH, but boy, I'm getting a lot of my inspiration from the FDA. Look at these covers here, right look at the covers these, these guidances AI and machine learning software as a medical device, real world evidence, digital health innovation Action Plan, the FDA, I would have been thrilled to see one of these. We got all three of these in the past five years. Digital Health Innovation means we can now write a prescription for more than 20 apps. I don't know if you know that 20 apps are prescribable. Real world elements we'll spend a lot of time talking about in a moment. That's our clinical data, and software and AI as a medical device AI will definitely end with today's just the future here. But one big use of the clinical data we have is assessing the care we deliver right otherwise known as real world data, or real world evidence here. Now real world data quickly overcomes and overpowers pivotal study data. I'm not going to go through this detail here. I'll just give you an example of just what this means for any of you. Humira is a well known drug from ABI right. Those four bullets, they're the first four studies that they use to get approval from the FDA from humera. Those are no secrets are on their website. The first one, they had 542 treated patients in Humira, the second one 690 patients that study 299 854. And then what happens after a drug is approved, boom, we use it more than 11,000 patients now that we have data on treated with humera. Right. At a certain point, we can walk into Abbey and say we might know more about this drug now than you do. Because we use it like crazy take any drug after it's approved and the device. After it's proved we use it like crazy, we might know more about these drugs than you do. Now, this really makes sense and the FDA is embrace this many examples now of how clinical data can be used to gain new indications or even approvals for drugs. The best example is Pfizer brands for male breast cancer. male breast cancer is certainly not a public health menace will take forever to get a trial. But the molecule is being used off label, it's already approved for female breast cancer, why don't use off label for male breast cancer enough clinical data existed in EHRs. That fiber is able, Pfizer is able to gain approval without running a clinical study. Right? That is happening now with the sapien three artificial heart valve gaining a novel indication there. We've got synthetic control arms, that's not synthetic data, fake data. This is just the control arm could be in one place that tested arm could be in another place. So much of that happening now. We've started now a center for real world evidence our we partner at UCSF and UC wide that people pharma biotechs. I think we have 15, pharma, biotech and device partners now that come to us with questions that we can help answer. And here's our best example in the device world working in 2021, partnering with ATTREX cure this a cryo nerve block, they were trying to gain an indication for adolescents purely using off label clinical data, they gained that indication from UCSF data, right. So we helped partner with that this is what we can do with existing data, you might be able to get an approval might be able to get a new indication here. More importantly, can we study where these things are working and not working? In fact, I think real world data is just the tip of the iceberg. 21 uses out there that we can at least count, whether it's updating side effects, better patient selection for clinical trials, long term post trial outcomes, just because you ended your trial, we still treat these patients, we know what's happening to them after your trial is done. You might want to know some of that too. And on and on and on so many uses for that kind of data. What I really want to end with is big uses of another big use of clinical data is using it to train computers to learn what we are doing in medicine or artificial intelligence. I don't think it's any surprise to any of you that AI and machine learning are just taking off. Now I think we will talk about Chat GPT in a moment here, but it's certainly a new world. For all of us in this area. That's just level set, make sure everyone gets what these terms mean. Artificial Intelligence is the broadest term, computers mimicking aspects of human intelligence, cars driving by themselves language is translating by computer that's AI, machine learning is one kind of AI, where we're using data to train that computer. Now you might ask, you're using data to train the computer, how else would you do it? Well, in the 70s, we didn't have a lot of data. So we used to use experts and rules, right? We'd ask a doctor, right? What do you do if you see this? And what do you do, if you see that, that really didn't get very far, it's much better to see what a doctor does, for example, in the EHR data, then ask them what they're doing. Okay. And so that's why the data revolution we have so much data is taken off. And deep learning is one kind of machine learning, where the models that we're building crudely resemble the human brain with one thing pointing to another one model pointing to another model pointing to another model, and you get the ability to learn nonlinear things like images and voice. And you see many examples of that out there in terms of using turning that in reverse and generating pictures and generating texts. Now, AI machine learning sounds scary. There is no wizard behind the curtain. You can literally buy dummies books on all of these $17 each. All these are our software tools we download to put in our software, and many of them are free now, right? You don't even have to pay money for these tools. A lot of them are developed by Google or Facebook or even Berkeley and a lot of tools and communities out there. AI in medicine goes way back to the 70s but one of the first examples was helping docs pick the right antibiotic right you know some Thompson hospitals have resistance bugs that are resistant to particular antibiotics. They have these anti microbial, anti microbial grams that they give out to residents and how staff. Well, can you put that in the computer? Well, in 1973, it certainly showed that you could look on the right there, literally from the New England Journal of Medicine, medicine and the computer, the promise and problems of change. We got to have that article today, because we're still stuck with this challenge of what to do with AI in medicine here. Why is AI back? We have an incredible hardware from the video gamers companies like Nvidia, yes, they make video game boards. But those same boards can be used for deep learning. If we buy so many of them, sometimes it's hard for gamers to buy them. We have great software libraries that have free and great communities. Many high school kids know how to program but the stuff now it's unbelievable. And in medicine, we have two amazing big datasets to work with. We got the whole molecular world, which is taken off, we got the EHR world, we just take it off. And we still have these hard unsolved questions to help answer. Now, this is not just theoretical, okay, since the FDA put out their guidance 500 AI and machine learning algorithms have been approved, the master list is on the link there. By the way, I'm tweeting all of this stuff out. So you can find this on my Twitter. So you don't have to write all this stuff down 500 It's updated weekly 150 of them just in radiology, and American college, college, radiology runs their own list. And so this is happening every single week. It turns out, a lot of these folks can work from home during COVID. So this took off during COVID, you don't have to go to a lab to do this, this is not going to stop this is just going to continue to grow, 500 Plus most of these by small companies, some by big companies, many of the companies that you know have already filing for these. Now, we're not too behind at UCSF, I think the fun once we've been partnered with a lot of different companies. I'll give you an example how this works. So at UCSF, we had so many pictures from chest X rays, that we figured you know, you could partner with a company like GE that makes an x ray device, right? And so if you're in the intensive care unit, for example, your patients not going to go down to radiology, right, they have a breathing tube in there might be on a ventilator. So you bring the X ray to them, right, the portable chest X ray? Well, the GE portable chest X ray unit now that you bring that you buy, when it takes a picture, it automatically will detect two things for you at the bedside. Where is that breathing tube ET tube? You'll kind of want it up, you don't want it in one of the lungs or the other? And is there a pneumothorax a little bit of a collapsed lung. Both of those now are auto detected by machine learning, trained with UCSF images, right? So you went from images that are already existing? And of course, you do have to do this safely responsibly taking out what patients want. But they've consented to all the way down to now the device has the algorithms built in. That's just one of many of these kinds of examples. All right, I'm gonna skip through a couple here to show you how this is going to work here in the future. Because all of you make devices, some of you make drugs, some you make drugs on devices. But all of these go together in a trajectory, like how do we take care of patients? Here's my favorite example, looking at a complicated problem called type two diabetes, perhaps you've heard of diabetes, of course, this is what the American Diabetes Association says, a doctor like me and endocrinologist should do for patient diabetes. Well, at the top there, it says get them to eat healthier lose weight exercise. When that fails, there's a big Metformin at the top right. And then when that fails, you got the six boxes in the middle choose one of these others to add to the Metformin. And then when that fails, add another category. And when everything fails, go down to metformin and insulin at the bottom right, you see that? Right? So I was curious to wonder where exactly the patients go. And which of those boxes do we choose? Because in the middle there, some of those boxes cost 200 times more than other boxes, but which ones do our patients go in? And so that's a question for us. So the analogy of this right, we have a guideline, it's great. We have a guideline. Look, it's pastel colors, it's like made for California, right? But there are a lot of diseases where we don't even have a nice diagram like that. You know that? Right? But I think of this, the analogy I like to use is it's a pachinko machine. In Japan, they have these really odd pinball machines that are vertical, and the ball starts at the top and kind of drops down. I think people place bets on these things. And the patients are like that, to which box they go to next. Right? And so that's what we want to ask at UCSF. So here, we could don't generate these kinds of figures. We now we call we used to call these diabetes doughnuts. And then we realize donuts are not so good for diabetes. We call these Lifesaver plots now. Okay, you're at 26,820 patients. We started on a drug for type two diabetes, okay. And the yellow means Metformin, that's great. A third of our patients started Metformin. That's what it said to do. A quarter of them at the bottom start on insulin boy, okay, that's interesting. Maybe they're sicker than they're they're supposed to be. And all the other colors are other drug. Some other combinations of drugs. And the gray at the top means there are so many slices there. I can't even show you them all. Right. So what happens in medicine, we take care of this patient, we say, Okay, try this drug come back in 90 days. Let's see how you're doing. And that's the next rake. The next thing is the next thing we did to this patient. yellow to yellow means they're still on that Metforman, we changed the dose. Yellow by itself means they're happy on those that dose we've never changed the dose. Any other color change means we added a drug subtracted a drug change the drug. Okay, go home, come back in 90 days. Let's see how you're doing. That's the third rake. Okay, now what will go home come back in 90 days. Let's see how you're doing. And that's the fourth rake. And then we realize we have 1500 different ways to start a patient on type two diabetes with at UCSF 1500 slots there. Okay, probably too many. Can we get it down to 1000? Maybe 100? Maybe 10? Right? We have docs who are like, Well, I was trained him with everything and then peel it back. Where does it say that in American Diabetes Association, right? Right, we're gonna do exactly what's that in the diagram, right? That's, and we hit one button. And boom, we have now the entire University California, one figure 159,000 type 2 diabetes patients. But we have more than 5000 ways to do this at university, California. Probably too many, we got to get down to one UC way to do this. And that to me is going to be the future here. Right? I'm not saying we're gonna get rid of evidence based medicine. But a lot of evidence based medicine has been expert driven medicine. Right. The experts convened experts look at the trials. I think the future of evidence based medicine is data driven medicine here, right. Now, why am I sing all this? Because think about how we treat patients, right? Let's try this. Let's see what the disease does. Let's try this. Let's write morning orders on morning rounds. Let's see other patients doing let's write some more orders. medicine is practiced synchronously, we write orders, we wait to see what happens, we write more orders. The same way how games like chess, checkers and go are played. Boy computers are really great at that. And it's only a matter of time before computers can look at all the HR data, figure out this worked, this doesn't work. Now, in fact, we can get so good. Now we can predict if we start this Med, that's a prediction, we can guess where the hemoglobin A1Cs going to be in 90 days. Right? That's the observed on the vertical there, right? We're pretty good now thinking that we start this drug on this patient. That's where the A1C is gonna go in 90 days, because we just have so much data. In fact, you don't even need deep learning to figure all this out. Here's a simple decision tree learn from the data. Can you remember the yellow write the patients who did well with metformin, and the folks who had to change, can we just predict who does well with metformin, and on the right there, you see the red, if your hemoglobin A1C was ever more than 8.8, or your fasting blood sugars ever more than 206, you're in the Red Square. And that means we know the Metformin is not going to work. I know what the American Diabetes Association says, but we've treated over 100,000 patients like this, it's not gonna work whenever the next move you're gonna make, make it now to get this patient under better therapy sooner. And that's what I mean by data driven medicine to help figure all that out. Doctor needs to make a choice, let's make a little prediction which one's going to work doctor makes the right choice personalized for each patient here, we can already start to do this and other diseases like rheumatoid arthritis, we can do this across hospital systems. In fact, I think the future is not just a learning healthcare system where we learn with our data, I think the computer is going to teach us automatically how to improve our care. Because I think that why are we waiting for experts to tell because they'll look at something, I think the computer is going to constantly look and say, by the way, I've now seen 10 patients with kidney transplants who are not doing so well on this drug. Maybe you should try this other thing as well. That that seems so far fetched. It's really not. You're just three examples we just curated just in the last week or two with Chad GPT. Now one thing is we have we have all of our notes de identified so we can legally put them into chat TPT here's a complicated Note that you can barely read from a cancer patient at UCSF a progress note and we just simply asked at the bottom give me all the cancer biomarkers well ChatGPT says well here er, estrogen receptor that's not in the note but it's explaining it to us. PR progesterone receptor hurt to write off the shelf ChatGPT knows what a cancer biomarker is and explains it to you. But at the bottom there's MP MammaPrint. Our own oncologist miss that the MP in this context is another biomarker happens to be a private company. It makes a product called MammaPrint. It's tiny MP high risk, but off the shelf ChatGPT got it our own oncologist didn't. GPT can do drug discovery. Given a currently available drug find compounds and similar property Let's modify them to make sure they're not patented, put in the purchase orders, synthesize them for me. Geordi now says I'm testing all of my diligence questions on every company that's pitching to me. Looking at trial design, clinic hypotheses, endpoints target validation. The days of the $1,000 an hour KOLs is over. He says, chatty, Billy can write the IND itself. Okay, given enough publicly available data, rewrite this section, and literally it's writing an IND off the shelf. How much do we pay consultants for that? I wonder, right. But it's a complicated environment with biases and lack of transparency, biases in terms of the data, we're using biases in the way the algorithms are being built. HHS is saying you Thou shall not discriminate by algorithms. Our own attorney general's asking us how do we curate and store these algorithms? FDA says sepsis algorithm shall be regulated further, epic itself says we're getting rid of our epic algorithm now. And then also, we have an AI Bill of Rights. Every screenshot I'm showing you here is to six months old or newer. At this point, it's that active of a field. And I think it's amazing time, what else can we do with clinical data? We're going to empower patients, we're going to power communities, we're going to empower all sorts of researchers, entrepreneurs, I'm going to end with this, what am I planning to do with this data? I'm trying to build maps of death and disease in California. Yes, that sounds morbid. But I want to know how our patients are going to die. And so here you see an example of a map, learn from the data. On the top left, we see patients showing up with alcoholism, each arrow is a year or a year later, they have liver disease, or they go to the left to go to liver abscess. And the squares mean yet they died of those disease, you can't die of alcoholism that easily in our system, but you die of the complications. Here's a harder one patients showing up on the left with heart attacks, heart attacks, you die right away, or a year later, you're in heart failure, and heart failure. Heart as a pump is too weak to pump that blood starts to build up in the lung, you got lung disease and orange. And on the right, there's septicemia. Now that was a surprise to me, because, of course, I trained as a pediatrician, I didn't take care of a bunch of patients with heart attacks, but I always thought it was the heart that killed you in the end. In fact, it's the infection that kills you in the end. Because if you don't take the northern route on the map, you can knock out your kidney and take the southern route on the map. And that's the point of having maps to figure out what's going next. But I don't want to just build maps, I want to show where all our of our patients are on the map. This is the real prototype with real Californians moving. This is how as they get older, the colors get brighter, they move from disease, to disease, to disease to death, a whole bunch of them are gonna get sepsis in a moment, they're in purple, and they're gonna die. There they go. It's okay, you can chuckle we can be morbid here. And now the future is predicting what's gonna happen next 90 days? What's going to happen next year? And what are we going to do about it? And that, to me, it's gonna be the new definition of an accountable care organization, one that knows how to account for the care of all 9 million of its patients. That's what I'm really proud of building with a lot of friends and colleagues in the University of California. I thank a lot of collaborators, a lot of folks make the database happen, a lot of funding, a lot of organizations, I'd love to thank my family for letting me go all over the place to give talks like this. Thank you very much. 


Audience Question  43:16  

Thanks for that talk. I feel like you blew my mind right off the bat. I'm from Texas, and I see a lot of parallels with the University of Texas Health System. And I'm wondering, do you work with other state health systems to do similar work? 


Atul Butte  43:28  

Great question. In fact, I believe the only other state that can pull this off is Texas. Okay. We have been trying to help them for years now at least eight years. I think they're a little bit behind where we are. There is definitely an effort in Austin UT Austin, so their office of their president to make things happen. I really believe UT can do this. The harder issue in Texas though, is they they're not all at the same peer level, meaning they don't all believe they're all at the same peer level. Right? I'm gonna show Riverside even though their baby, their baby medical school, but they're one of ours, right? Like MD Anderson and UTMB are going to be at they don't perceive themselves necessarily at the same level, let's say right. So I think you have to respect each other first. But we're trying to help them technically don't make that happen. Go ahead. 


Audience Question 2  44:16  

Chief Digital Officer in Austin, and I'd love to talk with you.


Atul Butte  44:20  

 I totally know Zane and Zane Zane is totally there. I'm aware we have exchanging emails constantly. I think that's a struggle he has there for sure.


Audience Question 3  44:29  

Yeah, I'm Jack Mormon from the US, Japan healthcare society. A lot of hospitals use up to Mayo Clinic has its own how do these compare or compete with the kind of work you're doing?


Atul Butte  44:44  

Look, I think there's gonna be a new role for partnerships here. Right? Look, this is partnering for progress leads. Oh my god, I think so. I think we got to get a look. It's a complicated healthcare system, right. We are all competitors in some ways. And for some reason, we love a competitive healthcare system. United States. I'm not saying it's right or whatever. So but it what it means. So as neighbors are gonna have a hard time sharing data with each other, I know Principal Investigators share data with other principal investigators PIs to do cohort studies and things. But in general, competing health systems are not going to bulk data share with each other, right? Because we put billboards next to Stanford, they put billboards next to us. I love them still, right. But that's how it is billions in revenue. What I do think is going to happen is we're going to have a little bit of a hopscotch, right, I'm gonna share with a neighbor, by my chair with the neighbors neighbor, right, where we're not competing. And essentially, that's what we have University California, right. We're far enough apart that we can actually share because we don't compete now. There are, there are, there's actually one avenue in Orange County where UCLA and UC Irvine a bot. Okay. And yeah, maybe in some discovery stare, I don't know. But in general, we're big enough that we don't have to compete with each other, we can see the business value of working together. So I think that's, it's a business model, right? It's not a technical problem. I'm gonna be the IT person who stands up, he said, This is not a technical, technologically hard problem. It's about the business model to want to work with someone else, or many others, and see those advantages, enough to put some money in to actually make this happen. And yeah, me it was working with Duke, Mayo working with many folks, Mayos working with Mayo, Mayo has a lot of different sites. So there are a lot of these kind of hopscotch is happening now. Little by little trevenna is another big one, which is a data store with about 20, different health systems, Providence and many others. Go ahead.


Audience Question 4  46:28  

Thank you very much. I really appreciate the energy you bring to this when when you look down the road 10 20 50 years, depending on how long it takes to get to that standardization. How do you think that will change the practice of medicine in terms of new treatments? So for example, ICU right now, there's not many randomized clinical trials from ethical considerations, but we can compare it this hospital does it this way, this hospital is that way. And you kind of get to Well, here's the best practice. What does that look like then when every system has this is the UC way of doing it? Does that actually at some point lead to now less opportunities for new insights from that data?


Atul Butte  47:08  

That's a good question. I think, look, I think we have a long way to go before any of us can tell a doctor what to do. Okay. And I'm also guilty of making everything seem easy on stage here, right? Of course, no one has to listen to me, I have 1000 dash lines, right on my org chart here. Right. So but I think it's about working with teams that want to start to standardize improve, reduce cost of care, right? We started this morning talking about a depressing environment, right? 45% of health systems are now in the red this coming year. So no one's gonna argue about reducing the cost of care, right? So you pick some kind of business planks that are unassailable in some ways, get buy in and go that way now. But in general, to answer your question, yes, I do believe in the future, physicians are going to have fewer choices, not more choices, because a lot more will be prescribed to them. But you can also argue it's kind of already like that, from payers. And what I'm gonna say is, let's use our clinical data to drive that. Not just some payer guidance and prior bought, right, which you could argue kind of made up, right, let's use our own data to figure out what's working and not working. Now, let me be super provocative in the last few last minute here, right? I am just asking simple, stupid questions. You all do trials for your drugs and devices, you all get approval, I just want to make sure my patients benefit, the same way yours did in the trials. Because you know, what, if they don't if they're not benefiting, maybe we shouldn't pay as much for this device or drug anymore? Right? How would we think about it that way? I kind of want you to make sure every one of my patients is benefiting now, this same way they did in your trial. That's all I can ask. I'm lucky if you got 60% of trial, I want to see 60% of my patients. But why is it we're not even looking at this right now? Right? I'm going to argue it's not just my responsibility. It should be my duty to study every damn thing we do in our health system. Make sure every one of my patients benefits from it. Right. That's a simple, stupid question. I should be asking more. Every health issue. I don't need a pakora. Grant. I should want to do this. Right. That's the future here that the data is there. We spent billions collecting this data by time we use it. We don't have time for another one, or should we just end on that? I don't want to be depressing. There's one more hand here. Let's quick one, just to say it and I'll repeat up Lisa, please. Yeah, so amazing. So a question about outside the United States. So I think the world is going to watch what we're doing here. In some some weird way we thought we were behind with electronic health records. And all sudden we're head because we have federal standards now, right? Again, we're all kicking and screaming to enter this world. China now is starting to get to some standards. Japan has those, GDPR makes things harder. We've had HIPAA for 25 years and so we kind of know what the rules are. That's a benefit for us. India's trying to get some standardization of EHRs. I think they have to cross that step that we crossed 10 15 years ago in terms of meaningful use of EHRs. But it can happen, right, it can happen. So I think the future is all of us showing what we can do with data. And really just teaching the world this is the right interest. This is the right direction to go in. We've spent a lot of resources doing all of this collecting this data. Let's make sure it's all working for all of us here, right across the world, across the state across the county, and certainly across the United States. All right, I think I should probably end here. Thank you very much.


LSI Europe ‘24 is filling fast. Secure your spot today to join Medtech and Healthtech leaders.

September 16-20, 2024 The Ritz-Carlton - Sintra, Portugal Register arrow