Chapters Transcript Video Artificial Intelligence in Cardiovascular Medicine I'd like to talk to you for a little about artificial intelligence and cardiovascular medicine. This has been a topic of interest to me for a while. And uh, there's stuff that I can share and there's stuff we're working on that, that is investigational still. And uh, and I have to keep my comments limited on uh, have the button. Okay. And this session everyone knows it's the green uh, forward. So if you think about computers long ago that moore's law was formulated and the idea was, if you look at the ancient computers from decades ago, the pace of progress is such that we're doubling our ability to compute speed, memory and everything like that about every two years. And at that pace we've gone from the original mainframe, ancient computers that couldn't do much more passion projects for people now to where all of us are carrying around iphones that are capable, for example, of beating any grandmaster in chess out there nowadays. So the computing technology at our disposal is tremendous in health care. That has been both helped us. You look at other things available, but also because of some of the restrictions in health care. It's been complicated in time for Halloween. We know there's a lot of negative projections out there in the world about ai from the terminator movies to comments by Elon musk and Stephen Hawking about the real dangers that exist. We do have to be careful in anything we do in this realm Ray kurtz. Well, who's a thought leader in this area has taken a more positive view of thinking. Eh, I won't really displace us humans, it's going to enhance us. At the end of the day, we've seen examples of that already. It's interesting in this original projection of his, it was thought that with the moore's law type improvement in computing performance that were in that era now where we'll get close to the ability for a computer to have the number of connections and complexity of something like an animal or human brain. So the future is getting close to us. Now, when you think about applications of AI, I'm gonna start by talking about games, that's where the most progress has been made. And this is a nice little display. There are easy games like tic tac toe, tic tac toe has been completely solved. This is the best algorithm for playing tic Tac toe, you'll never lose, you'll always win or tie depending on what goes on. Very easy. That's easy to program. Chess was much more difficult. Um, in 1997 IBM built deep blue and deep blue built Gary beat Garry Kasparov, the chess master, that was the highest ranked at that point in time. There was a lot of skepticism, there's great movies about what was going on and everything else. But nowadays it's clear that even uh, iphone application has the computing power to be able to really beat a grandmaster in chess and that's created its own set of problems, those of you following the nerd chess news out there, know that uh, magnus Carlsen, who's the world champion was recently beaten by an american younger player. And there's been allegations of cheating and lots of craziness around that. There's, yesterday's news is there's $100 million lawsuit related to this now. So, so in games, computers have really shown supremacy. Now there are more complicated games than chess. There's a game called go that for a long time was considered to be the most complicated game out there. And for those of you that really are finding this interesting. A company called Alpha Go produced the first computer algorithm that could beat these folks who in asIA that have spent their whole lives becoming go masters. And that was felt to be very difficult because the complexity here chess, there are 10 to the 120th combinations and in go It's 10 to the 700th number of combinations. So you can't use brute force to solve go, you have to develop a program that's going to have some level of intuition, it's just too many moves. There's more moves in the game of go than there are atoms in the universe. So you can't just out compute what alpha go did. That was amazing is it made a huge splash. It's beat all the major players in the, in the world, frankly. And ultimately they were acquired by Google. And what google did is there were publications about this artificial intelligence system. It's deep convolutional neural nets and it's amazing the way that they work and it's beat the best players in the world. It went on to get eight hours to train itself to learn chess and then was able to beat the highest ranked computer chess algorithm out there as well. So that technique of neural nets is now kind of the leading force in image recognition in game theory. And just to show you where that's gone, Google purchased, it created a deepmind division and now is turn that algorithm over to working on health problems. They aren't sharing much of what's going on. It's gonna be fascinating to see what they do from that. And I can spend hours talking about some of the deep specifics. It's more time than we have, but it is fascinating that from game theory to medicine, we're seeing this evolution of Ai as something that's gonna become a part of all of our lives. The problem when you take these algorithms which have been insanely successful in games and try to apply them in medicine is the issue of kind world versus wicked world problems and that's where we're stuck right now. Kind world situations are like games, there are repeating patterns, you get immediate feedback, you know, immediately what was right and what was wrong in a chess game, One person, one person lost. You can look back and there's always immediate adjudication as to did you make the right decision or not, medicine is a wicked world application. The patterns are complex, huge data streams. I mean big data. This puts to shame, feedback takes a long time and in fact, we're not sure of a lot of the feedback we've seen talks today where we've discussed all the gray areas that exist in medicine. How do you know if the intervention we made at any one point was the right one. Sometimes it takes 5, 10 years or a lifetime to know what was the outcome? Was that the right decision or not? And there are so many factors. It's much more difficult to train an Ai platform in health care than it is in. For example, a game where you have open information, everything is available. You can see it and codify it quickly. But that's what people are trying to do right now is find a way to make this work. The one health application that has been successful is that Deepmind group turned alpha go into another iteration, which is alpha fold. One of the hardest things to do is to take a sequence genomic sequence and predict what the protein that it produces will look like. And there's an annual competition to try to see who can do that the best. And if you look, it's obviously from from 2006 to 2016, the winners have been humans but alpha fold has outperformed any large organization or human workgroup in predicting folding. They actually among the different proteins. They ask people to solve one was a covid viral envelope and you can see alpha fold is by a huge amount. Outperforming the winners from the decade before. So these applications are clearly coming and starting to make an impact. Um erIC Topol recently wrote a book called Deep Medicine and again the idea is there's humongous data streams available to us, there's room for a i to make our lives better. It's the growth in Ai and healthcare is huge and we've been involved in some of that. One of the interesting things is you think about a lot of our algorithms we use one dimensional or sometimes two dimensional analysis, one dimensional analysis, EFEF 35% is our cut off for defibrillator or not. Right. But really that shouldn't be a yes or no, there's probably more continuous nous to that variable. But right now we don't have easy ways to turn that into simple algorithms that we can follow as clinicians, two dimensional analysis, calculating a B. M. I. We're good at that. What we're not good at is when you have many dimensions together and really complex patterns of inter reactions and inter relations between these and that's where computers can be very useful to us. This is from a Jack article that was published about two years ago talking about what's coming in the world of machine learning and with data streams from apple watches, fitbits, cardamom devices, implantable loop recorders. There's huge streams of data that is clinicians. It's hard for us to have the time. Um, and sometimes the grounding to to take a look at. And so that's where I think there can be a real impact. Some of the projects Sentara has been involved in over the years. We were working with IBM Watson, that project is over now and we made some progress. Watson again did very well with games. It beat the jeopardy champs. But in health care it still hasn't really made a lot of progress. We did a several year project that was a multimillion dollar project with a company called JV in looking if we could predict sepsis heart failure readmissions in other areas. And it's interesting. Again, the data streams are so complex that ultimately when we finished those projects we found the additive value was limited compared to what we were doing with just sort of clinical support to physicians and apps. They thought they could predict sepsis. So we could prevent it. Part of what made it hard to see a benefit from is we don't have data on what you if you think the patients can develop sepsis in the next week. It's hard to know what interventions make a difference. More handwashing, antibiotics. Not really proven. And so those have limited some of our projects to date. We're about to embark on one on EKGs. There's so many EKGs done in the system. There are so many data streams of tele metric data we've seen now. This is a publication in nature that there are algorithms out there that can provide cardiologists level arrhythmia detection. Uh this slide is just pointing out that looking for a fib a flood or other arrhythmias like that. Some of these algorithms and we're getting ready to try to launch into this here. Could look at large data streams and make it easier and safer for us to pick up patients that need anti coagulation or other therapies. And what's coming next is even crazier. This was published Alexa. Those devices that you can hook up at home standing in front of a monitor. There's actually pretty impressive data that it can detect the regularity of your heart rhythm in a quiet room just from the contact and the sound waves that come off of us from our heart rhythms. Obviously we've got other devices like cardio devices, Apple watches and fitbits as well that are going to create huge amounts of data that could potentially be utilized to get us on track. So my quick predictions here are as the amount of patient data increases artificial intelligence and big data will become an essential part of the daily practice of cardiology like vision has in radiology and pathology. They're ahead of us in this patient center. Disease focused real time biometrics will allow teams of providers with the right ai guidance to focus resources on the right patient at the right time. A lot of what we do that's very tedious E HR related interactions billing maybe I could help us in some of that. I think we do need to embrace the change at each stage of history. Humans have been worried that aI will displace us and that's never been the case as we've gotten more sophisticated tools at our disposal. What's happened is we've extended what we've tried to do and it's actually usually created more jobs and more opportunity. Thank you guys for your attention. Published December 7, 2022 Created by Related Presenters Deepak Talreja, M.D., FACC Sentara Cardiology Specialists Dr. Talreja provides top-quality, patient-focused care for the prevention, diagnosis, and treatments of disease and conditions of the heart and blood vessels. View full profile