SCCM Resource Library-The Future of Artificial Intelligence in Critical Care: A Tsunami of Predictions or a Consensus of Opinions?

Video Transcription

Good morning, I'm Gloria Rodriguez-Vega, and I'm pleased to introduce our first plenary session, the Max Harry Weill Memorial Lecture, being presented by Dr. Gary Clifford.  Dr. Clifford's research focuses on the application of signal processing and machine learning to medicine to classify, track, and predict health and illness.  His focus research areas include critical care, digital psychiatry, global health, mHealth, bioinformatics, and perinatal health.  As a principal research scientist at MIT, he managed the creation of the MIMIC-II database, the largest open-access critical care database in the world.  As an associate professor of biomedical engineering at the University of Oxford, he helped found  the SLEAP-Ensicardian Neuroscience Institute and served as director for the Center for Doctoral Training in Healthcare Innovation.  As chair, Dr. Clifford has established Emory's Department of Biomedical Informatics as a leading center for critical care and mHealth informatics, and as a champion for open-access  data and open-source software in medicine, particularly through his leadership of the Physionet and SynC challenges and contributions to the Physionet resource.  Today, Dr. Clifford will present the future of AI in critical care, a tsunami of predictions or a consensus of opinions.  Ladies and gentlemen, please join me in welcoming Dr. Gary Clifford. Dr. Gary Clifford  Well, Gloria, thank you very much. Thank you very much for everyone coming here today. Also to Craig Cooke-Smith and Tim Buchman for bringing me here today as well.  If it wasn't for Gloria, Tim and Craig, I don't think I'd be standing here. So I'm really delighted to be able to talk to you today about this particular future  that we were all anticipating, AI in critical care. First of all, I want to start off with a straw poll, though. How many people here today are using AI in their hospital right now?  One, two. Two of you. Three of you. Maybe three, four of you. All right. Maybe I'll give you another clue.  If we say AI is a program that can sense, reason, and adapt to the data coming in, how many people think now that they're using a monitor or something like that that can do  this in their hospital? Even less, maybe. No, one, two, three. Well, actually, you'd be surprised.  It depends on your definition of AI, but if you think it's just something that's pulling in data and reasoning about that data, then actually there are many, many monitors already  in the hospital that would fit this definition of AI. I want you to think a little bit more broadly than just this idea of some super intelligent  algorithm that's going to solve all your problems and probably scare you and your children.  So I want to think about the rise of artificial intelligence in life critical applications outside of medicine, first of all.  So actually, you'll see over the last 10, 20 years, we've seen a lot of progress in self-driving cars. We've seen all these sensors on them.  We've seen all this information coming in and being processed. And we're getting to this position where we're starting to think, is the self-driving car  paradigm going to become the ubiquitous way that we travel around? And what will that mean for us in terms of safety as well as transport?  Well, if you think about the mortality rate just due to non-self-driving cars, in other words, humans driving cars, it turns out that this is the single largest killer of humans  between the ages of 5 and 29 years on planet Earth right now. And it's a leading cause of death even in places like the United States.  It's actually double the number of opioid deaths that we have every year. So you can think of us driving our cars as an epidemic, a mortality epidemic.  So let's look at the stats for self-driving cars about whether we think artificial intelligence can provide some benefits, some reduction in mortality.  Well, actually, there's been just five deaths from Tesla vehicles globally, four in the U.S. and one in China since 2014.  So you might ask yourself, well, what does that mean in terms of a per capita or a risk basis? Well, what we do is we normalize it by the number of kilometers driven.  So there's 3.1 per billion kilometers, there's 3.1 deaths, which actually makes self-driving  cars as safe as the safest drivers in the world and twice as safe as the average American driver. So you're probably wondering who are the safest drivers in the world.  Turns out to be Norwegians. And maybe the reason for that is the structure of their roads. I don't know. But it also might be due to population density.  But there are many reasons that are the factors in this. But the key here is that it's safer for AI to be driving your car than you on average, no matter where you are in the world.  So we have to ask ourselves, can AI reduce errors in medicine? Can they make it safer in medicine? That's why we're here today, thinking about what is the future.  So you'll see lots of publications in the last few years talking about medical errors and the fact that it's estimated to be between 200,000 and 400,000 deaths a year.  This turns out to be about 60 to 120 per 100,000 head of population, or about five times the death rate from car accidents.  So this is a significantly high number where we're thinking perhaps AI can do something about it.  AI can potentially provide some kind of oversight or second opinion that can help us reduce these errors.  Well, obviously not all of these are amenable to artificial intelligence helping you out. Some things are completely unavoidable.  What we're really interested in, I think, is misdiagnoses. So whether we can have a second opinion or a third opinion from an AI system to help  us with just this enormous amount of data that's flooding our ICUs and our hospitals in general. So if we think about this and how this is going to manifest, we have to think about  the driving technologies behind this. So what is it that has led to this revolution in artificial intelligence over the last 10, 20 years?  And in fact, more specifically over the last five years, where we've gone from imagining that AI could actually do something to seeing it as rapidly becoming a reality.  Well, many of you are familiar with Moore's Law. How many of you have come across Moore's Law? Good, good. A few at least.  I think many more people actually are familiar with the idea that every 12 to 18 months, the speed of computation in your computer roughly doubles. This is called Moore's Law.  So if you plot it in a log-log plot, or this is a log-linear plot, so it's a curved line. But if you plot it in a log-log plot, you get a roughly straight line.  And so if you look at this plot, you see all these different processes over the years, starting with IBM's earliest machines down in the early 19 to mid-1900s.  And then this graph goes all the way up to about 2010. And then there's some predictions about quantum computing, which are becoming real right on the edge of this curve.  But the real interesting thing that drove artificial intelligence, and particularly self-driving cars, and now more and more all machine learning, was the invention of the GPU.  The GPU is like a CPU, but it's a graphical processing unit. So it's got lots and lots and lots of very fast and very rapid bandwidth processes.  So they can push a lot of data through it. And this can be seen, if you look at these two graphs, the blue one is telling you how  fast computation is increasing in terms of speed in the central plot and memory bandwidth. So that's how much data you can push through these CPUs to actually process the data.  It's not just how fast your computer is, but it's how much data you can push through there as well. And you see these CPUs increasing at a fairly rapid rate between 2008 and 2020.  But if you look at the green line, this is the remarkable part. GPUs have just been increasing at an exponential rate.  We're seeing faster and faster machines that are allowing us to process incredible volumes of data incredibly quickly.  And this is what's been driving the revolution, both in self-driving cars. NVIDIA have been shrinking these devices down.  They've been shrinking the amount of energy it takes to run it so it can fit in the size of a license plate and put it in your car.  And this has been driving the cost of these instruments for processing right down and allowing anybody to just slam them in their computer and get computing.  In fact, one of the other things that drives these GPUs is the gaming industry. So GPUs are very good for graphical processing, in other words, images and videos.  And so this has made online gaming and gaming on your PC much more easy. And this has driven the market as well.  So we've ended up with these very high-end computers that are in the hands of everyone. And it's sort of democratized the power of computing.  Everybody now has a supercomputer in their pocket compared to the machine that I wrote my PhD thesis on, which was programming neural networks from scratch in C. And it worked  about five times slower than your average pocket mobile phone, and it had about a quarter of the memory. And this was my big desktop computer for my PhD.  So things are just changing so rapidly. The democratization of artificial intelligence or machine learning, as we should say, which  is a subset of artificial intelligence, which involves our algorithms that we develop to help artificial intelligence.  This democratization is also occurring in the way that we teach people how to do AI. We're increasingly seeing online courses like Coursera.  Many of you might have heard of Coursera. These are these online courses that you can take a couple of nights a week, and you can teach yourself how to do machine learning.  And you can enroll for free in many of these courses, or it's maybe $50 to do it. And so we're getting this democratization of being able to do, for everybody to be able  to empower themselves to do machine learning. There are all sorts of resources online where you can teach yourself outside of this, and  there are forums where you can dig through the nuances of it. And then, perhaps most importantly in this democratization of AI, was the publication  of some code called TensorFlow, which many of you might be familiar with. This was something developed at Google over many years by Jeff Hinton and his team.  And it was a way of providing auto-discovery of features, and in fact, led the way for something we call deep learning.  So many of you might have heard of deep learning, as well as machine learning. So this idea that we have extremely deep neural networks that can learn a whole variety of things.  And then, more and more, we're seeing online tools that wrap around these things so that we can just plug and play.  So if you go to Amazon Web Services now, AWS, you can open an account, you can get some free credit if you're in academia, and for a very small amount of money, or potentially  for nothing, you can start doing plug-and-play machine learning, uploading your data, putting pipelines together.  So we've got this democratization of artificial intelligence going on, not just in terms of power of computing, but in your ability to take control of it.  So this has led to a tsunami of machine learning in medical research. It's not just engineers and computer scientists producing this work, it's anybody who wants  to get their hands on it. And that's both a double-edged sword, a good and a bad thing. If we just look at a search term for the words deep, as in deep neural network, neural, as  in neural network, and EMR, electronic medical record, on Google Scholar, you'll get 9,230 results. And these are mostly from the last five years.  In fact, actually, if you just search in the last year, you'll see that you get 1,690 results. People doing deep learning on electronic medical records in Google Scholar.  Just the publications. This isn't the whole web, just scientific publications. That's a lot of reviews that I have to do each week.  So what does this mean for medical diagnosis and prediction? Well, we're seeing more and more the excitement around the idea that AI is going to help us with doctor shortages.  It's going to fill in for places where we don't have the capacity, we don't have the human power to deal with the throughput.  And probably in low- to middle-income countries, this is going to be true, that we're going to see more and more the use of AI to fill in for resource-poor locations.  But there's also the idea that instead of replacing the human, it's going to be an assistive technology.  It's going to give you a second oversight, a second pair of eyes to look at what you're doing and flag when you're having a bad day.  You know that everybody has a bad day and that the failure rate for most people in any application is somewhere between 10% to 30%.  So 10% of the time, at least, we'll be making a mistake. I'm sure as I'm talking here, I've made a mistake already and told you something that isn't true.  So whenever you're listening to somebody, you should think, I wonder what mistakes they're making right now.  So having that second pair of eyes behind you that's providing some kind of fact-checking,  providing some flags to tell you when something's abnormal and the diagnosis that you've made is out of the ordinary for what you're seeing in that actual image is a very good idea.  And so there's more and more excitement around this instead of replacing but being assistive technology.  So we're seeing this incredible promise of deep learning in medicine in particular. We've talked about AI and more specifically machine learning.  And then more specifically than that, this sub-branch of machine learning for deep learning,  which is these incredibly complex, deep-layered neural networks that have lots and lots of parameters, thousands of parameters that can learn very complex relationships between your  data and the output. And what we're told is that the promise of deep learning goes something like this.  This is standard machine learning, that as you increase your data set size, this is performance metric on the vertical axis, data set size on the horizontal axis.  And this is actually real data that I made up in PowerPoint. So you can see that we're told that with conventional machine learning what happens is it plateaus reasonably quickly.  Once you've got to 10,000 examples, it doesn't matter if you add another 20,000 or 30,000 in there.  Conventional machine learning isn't going to give you anything new because you've made up your features. You've decided what the right covariates are for the outcome.  And you may have put a few extra things in there that you weren't really sure would actually be correlated with the outcome. But that plateaus pretty fast.  What we're told with deep learning is that it's going to auto-discover the features in our data, things that we could have never thought of, associations that we could have  never thought of. We can't even comprehend ourselves because there's these thousands of neurons that are connecting very, very complex and intricate patterns within it.  And what's going to happen is that no matter how much data we put in there, it's going  to increase our ability to do pattern recognition and classification, and we're going to discover more and more and more features. And that's never going to end.  And unfortunately, what you find is that, although this seems like a very, theoretically it seems like a great idea, it turns out to be a bit of a myth.  Because what happens is, as you increase your dataset size, you become less and less able to curate the quality of it.  So if I take 10,000 ECGs and I manage to get 20 people to overread those 10,000 ECGs because I've got millions of dollars to pay people to do that, then I get fairly high-quality  ECGs. People will disagree anywhere between 20% to 40% of the time, but I can find a consensus decision and get some really high-quality labels on them.  But now if I go to a million ECGs or I go to 50 million ECGs, which is the current problem I have, what do I do with the labels? I can't employ 20 people to label 50 million ECGs.  It's too much of a lip. So you start to cut some corners. You start to use just a single label on each one, or you use an algorithm, you use the  monitor's generated algorithm as your label. So your quality of your labels are going down as you increase the size of your data.  So what you see is a drop-off in performance instead of an increase in performance. There's actually an optimal data size if you have a labeling problem.  If you have a quality of labeling problem, then there has to be a point where you're  adding in more and more noise, and it's not helping out if you take a conventional machine learning approach.  If you take vanilla deep learning, so in other words, just the standard algorithms that are out there right now, state of the art, then you see something even more frightening.  It's going to plummet through the floor, and why is that? Well, if it's an auto-discovery mechanism, it's going to start learning the noise in  the data more and more, and it's going to build in more and more spurious outcomes with your input.  So when you test this, even though the training performance will be very high, when you test  it on data you've never seen before, the actual performance is going to plummet through the  floor because it's created all these spurious associations with your input variables compared to your output variables. So a good example of that, I think, is looking at sleep data.  So sleep polysomnograms and sleep studies have been shown to be pretty consistent. You get a cap of anywhere between 0.6 and 0.8, so you'll get a very high agreement  between different professionals if they're using a polysomnogram, a full montage of electrodes, EEGs, EOGs, EMGs, to record people's nocturnal activity, neural and physiological.  But as we try and scale that and say, well, what's the sleep health of the population? The direction we've gone is saying, well, it's kind of difficult to measure these stick  electrodes over everybody's head and measure all this. So what we're going to do is we're going to just put wrist devices on people and measure  how much their arms move around at night. And that's extremely limited. It only gives you a very, very small slice of your behavior during the night.  It doesn't pick up things like insomnia or if you have restless leg syndrome and your  devices on your arm, it's not going to pick up that particular problem. So actually, as we scale and start to look at sleep on a larger and larger basis using  these wearables, the quality of the data is plummeting through the floor. And it's exactly what I'm predicting you'll see from this kind of machine learning.  So what can we do about this?  Well, I think the international competitions that we're running, like the Datathon that  happened over the last couple of days, are the solution. Not because they create one single winner, but because of the whole industry that springs around it.  If you look at the Ansari XPRIZE, how many of you remember the XPRIZE for suborbital space flight? If you didn't, then you clearly don't engage in any media whatsoever.  It's hard to miss the fact that somebody put up a $10 million prize purse, this is about  five or ten years ago, for the winner of the first private industry that could reach suborbital space.  So you had to be a certain distance from the Earth for a certain period of time to get the exact criteria.  26 teams invested over $100 million in aggregate for research and development.  And the breakthroughs led to a private space industry worth over $2 billion today.  So Virgin Galactic licensed the technology that came out of this and will be the first private company to do space flight.  And this came from just a $10 million prize, which is amazing. Amazing to create that kind of industry out of it.  And it's more than just a single winner that's important. It's this competitive industry that springs up around it.  So what do we do about this? And in particular, how does this apply to medicine?  So we run a series of competitions every year through Physionet. This is a resource that Ari Goldberger and Roger Mark started at Harvard and MIT.  And if you're not familiar with this resource, it became the standard reference database for all arrhythmia detectors that the FDA required you to test your algorithms against.  So every one of you who uses a bedside arrhythmia alarm has to thank Roger Mark and Ari Goldberger for setting up Physionet.  And also George Moody, who was instrumental in the design of all the databases and all of the mechanisms behind this.  And for the last 20 years, we've been running a series of challenges that range from detection of sleep apnea from the ECG to predicting acute hypertensive episodes to looking at  T-wave alternans. And then most recently, last year, we ran a competition on early prediction of sepsis from EMR data.  This is a 20th annual competition sponsored by the Gordon and Betty Moore Foundation, Google Cloud, and MathWorks.  They provided cloud credits and funding and prize money to help us build this competition. And what we did was we posted a whole series of data where people had either—it was a  generalized ICU population where a subset of people had got about a 20% incident rate of sepsis.  And just to recap on sepsis for any of those in the audience that might not be familiar with the history of sepsis, sepsis 1 criteria based on SERS was delineated in 1991.  This had a high sensitivity but a low specificity. In 2001, this was redefined as sepsis 2 as suspected or documented infection and other symptoms.  So this is more difficult to quickly diagnose. So the sensitivity specificity tradeoff shifted.  And then in 2016, it was redefined again to be suspected or documented infection and an acute increase of more than 2 in the SOFA score.  So this is low sensitivity but high specificity. And we used this for the competition because we wanted something that clinicians would react to.  What we asked people to do was say, can you predict sepsis before the clinician suspects that this patient has sepsis?  Because if you don't do that, then you're creating an algorithm that's of no use. You have to guess before the clinician is guessing.  So it's a harder task than predicting something that will happen. You're predicting that somebody will guess that something will happen.  So having that human in the loop is an extremely difficult problem. So we want to take somebody who appears initially healthy. They have a run of healthy data.  And then at some point, though we're not absolutely sure, the house starts setting on fire. We're not really sure when that is. We don't notice it immediately.  The alarm doesn't go off.  But at some point, the blood pressure starts dropping and some of the signs start to indicate it. And we want to get it before we get to this point.  We want to get it before we get to septic shock.  Chris must be laughing now because this is his slide.  And then, of course, we want to get it certainly before we get to the end. But it's that house on fire part that we want to avoid.  So how do we assess this? This is non-trivial. You can't just do area under the receiver-operator curve or accuracy or sensitivity or specificity  because you're constantly making a prediction.  Most of these machine learning algorithms and most of those papers that I showed you earlier had this assumption that usually what people do is they take a bunch of septic  people and a bunch of non-septic people and they say, look, I can differentiate these two groups with an area under the receiver-operator curve of 0.8. Isn't that fantastic?  Well, it turns out that that isn't a very good prediction mechanism because you have to repeatedly try and make these predictions.  So what we did was we came up with a new penalization method. We gave it a continuous output score that gave if you were six hours ahead of time, you got your maximum marks.  You got zero marks if you were late.  And if you miss a prediction, you got highly negative marks. And if you were too early, we gave you slightly negative marks as well because telling a  clinician that somebody is going to become septic before they're going to believe they're going to become septic is probably a false alarm, even if you're correct.  So there's a lot of psychology behind this actual function. Of course, you know, the actual temporal dilation of this and the magnitude changes based upon your different ICUs.  But this quantifies the benefits of early sepsis prediction and harm from false alarms and overtreatment. The data was sourced from multiple hospitals.  There were 2.5 million data points, 40 vital signs, labs, demographics, over 63,000 patient records from three distinct hospitals.  And we posted half of the data publicly and half of it privately for two of the hospitals. And then a third hospital, we kept it completely hidden.  We didn't show anybody the state and we never let them touch it. And we waited right until the end of the competition. And they had 10 shots at the public and the private data.  And then right at the end of the competition, we took the algorithms that performed the best and we ran them once and only once on this hidden hospital.  And it's interesting to see the results. When you sequester some data and never let people see it to have a look at what the performance is, it's very, very interesting.  So how did it go? Over 122 days, 101 teams, 840 submissions, about half of them were successful.  And most of them managed to complete in less than three hours their computation on 60,000 patients. So all of this can be done in real time.  We did a total of 107 days or more of computation in the cloud on this, running all these different algorithms.  And we analyzed the performance over these three sets. So if you look at these three columns, this is their performance in hospital A, B, and then C.  And what you notice is this drop in performance on hospital C because they never got to see this data.  And there's a substantial drop in performance there in almost every single algorithm.  And that indicates that there wasn't enough generalization outside of the training data. So if they did well on hospital A and B, they did poorly on C because this test set C has  some slightly different noise and error characteristics, which is always going to be seen in external hospitals.  So the question is, can we overcome this by learning which algorithms to trust and when?  And the example I want you to think about is creating a hurricane model for this. And what do I mean by that?  Well, actually, when we do hurricane prediction, we tend to run a lot of models in parallel.  And when they converge, we say, this is a good prediction. So if you look at this plot, this is Hurricane Sandy.  And the intuition is you look at it and you say, OK, well, when they start to agree, when they all start to have a central tendency, I'm going to start to believe this.  And if any of you remember Hurricane Sandy, if you were in the States at the time, there was a lot of concern that it was going to hit the East Coast, even though it looked  like it was heading out into the Atlantic. And the reason for that was most of the predictors, after it passed over Jamaica, most of the  predictors collapsed and started to agree that it was going to slam into the New Jersey shore. And so when they're all together, we trust them.  And when they diverge, we don't trust them. That's the analogy I want you to keep in your head. So when you've got a lot of models agreeing, that's the time to respond to it.  So we said, can we create a voting algorithm?  We used all 88 models initially, trained it on the training data, didn't use any of the test data, and said, which of these models can we put together, find the context in which  a model does good and a context in which the model does bad, and say, how do we predict?  What's the prediction on our hidden data as well as our training data?  And once you create this voting algorithm, you find, you look at the top of this plot, that's what the voting algorithm does.  It guarantees this superior performance across all of them, not much better than the best algorithm. But you notice the best algorithm is different from database to database.  And we created this algorithm that's always better, or at least as good as any of the best predictors, without knowing which one is actually the best predictor.  So it's a security blanket, if you like. It's always going to create your best prediction.  So let's go back to the Hurricane Sandy example, because you have to think in your mind, when do I make this prediction?  Because if you had made the prediction very early on, before it crossed over Jamaica, so if you look on the right-hand side, this is the storm track much, much earlier on,  about four or five days before landfall.  And it hadn't crossed Jamaica at this point. And so all the models were disagreeing. Some had them going out to the Atlantic. Some had them hitting as far north as Greenland.  And no one was really sure where to make the call. And the reason for that is that there's too much uncertainty with a hurricane passing across the land.  So once you get across Jamaica, there's less perturbance in the signal as it moves through just across open ocean. And so at that point, you can start to make a strong decision.  The lines have collapsed, and we're confident about it.  This is another example.  This is Hurricane Dorian, which is more recent.  And these are the confidence intervals for all of these predictors put together. This is a 95% confidence interval.  And I want you to look where the confidence intervals span as they go up the East Coast. It's very, very difficult to try and work out where this is going to make landfall.  And if you remember Hurricane Dorian, it was very difficult. Everybody thought it was going to hit Florida.  And then it moved further north, and it just sort of skimmed the coast in the end. But it was very difficult to predict.  And that's because all of these confidence intervals sat right here on this East Coast of Florida.  And you can see that these confidence intervals went out through Georgia. And you can see that it just sort of intersects with the bottom right corner of Alabama and  Georgia as well, which is important. Because you have to make a call about where would you make this prediction.  So here it is. This is the real-time or the retrospective analysis now. But this is the yellow is the hurricane itself and its extent.  The white cone is the 95% confidence interval of where it's going to go. And then the dotted area is the less certain area where you're not really sure it's going to be.  But you can make some kind of guess about it. But it's really not something that you would want to react to.  So if you were thinking about this as a patient, the white area is the area that you want to react to.  If your patient looks like it's going to go into that white area, then you want to respond to that area.  All these algorithms are saying this is pretty close to certainty that something in this white cone is going to happen. And if you're caught in that white cone, it's bad.  And so you should react to it. You should get away from the coast.  Or if it's a septic patient, you should treat with antibiotics. So it's a way to utilize those resources. But if you're in that dotted area, then you can't be absolutely sure.  So you also want to ask yourself, what if somebody provides misleading information? That's why I said look at the...  I'm not making any particular statement here except for the fact that sometimes you can be misled about the information that you're receiving.  Turns out that you can learn if someone is almost always wrong because they disagree with the known facts or the consensus too often.  And what you can do is flip their vote.  You can say, well, because they are making a prediction of X, I know the answer is really Y.  And you can build that into your machine learning model. So this is a serious point.  You can learn that anybody who's giving you a piece of information in this consensus model adds utility to your score. Doesn't matter whether they're right or wrong.  You can work out what contexts are right, what contexts are wrong. And even a consistently wrong algorithm is giving you some information.  The only thing that doesn't give you any information is a white noise generator.  But to be fair, this was actually the set of storm tracks that I think Mr. Trump was making his prediction from. And you can see that there are all these lines twisting in and out.  And some of them go as far west as Louisiana.  So if you're not familiar with how to interpret these kind of plots, it would be not unreasonable  to say, OK, it looks like there's a chance that this could go all the way through Alabama or maybe even Louisiana.  But of course, hopefully now you've realized that these are very highly improbable tracks and that the cone of activity that goes through the center is where the hurricane is really  likely to go. The other thing is, look at when these predictions were taken. This is when the hurricane was very small, before it had passed over any significant  land masses in the Caribbean.  And so your ability to predict at this point is extremely diminished. So this is the point about not trying to treat too early.  You want to watch and wait, which is a familiar thing for us in medicine. And you want to watch the track and then watch till these lines converge.  So eventually they converge and then you can make a prediction now. And you make that treatment at that point.  So that's the main point here. We should listen to each other. We should work with each other. And we should be working as a group. Medicine is always all about working in a team.  So I just wanted to thank the audience, of course, the Gordon and Betty Moore Foundation, Google Cloud and MathWorks, the NIH and the NSF for supporting this, Roger Mark and Eric  Goldberger for the years of dedication for Physionet and supporting this. And for this particular work, I really wanted to call out Matt Reimer, Shichimim Nomadi and  Chris Joseph for their contributions to these slides and all their hard work on this, because they're really the engine behind this. Thank you very much.

Video Summary

Dr. Gary Clifford presented on the future of AI in critical care during the Max Harry Weil Memorial Lecture. He discussed the application of deep learning and machine learning in healthcare and the potential for AI to assist in medical diagnosis and prediction. Dr. Clifford emphasized the importance of using AI as an assistive technology rather than a replacement for human doctors. He highlighted the advancements in computing power and the democratization of AI, making it accessible to everyone. He shared examples of how AI has been used in other industries, such as self-driving cars, and how it has led to improved safety and outcomes. Dr. Clifford also mentioned the challenges of using deep learning in medicine, including the need for high-quality data and the potential for the algorithm to learn noise in the data. He discussed the importance of international competitions, such as the one run by Physionet, in advancing AI in healthcare. Dr. Clifford shared the results of a competition on early sepsis prediction and highlighted the need for algorithms to be robust across different datasets. He concluded by emphasizing the importance of collaboration and teamwork in using AI in critical care.

Asset Subtitle

Administration, Procedures, 2020

Asset Caption

The Society of Critical Care Medicine's Critical Care Congress features internationally renowned faculty and content sessions highlighting the most up-to-date, evidence-based developments in critical care medicine. This is a presentation from the 2020 Critical Care Congress held February 16-19, 2020, in Orlando, Florida, USA.

The Future of Artificial Intelligence in Critical Care: A Tsunami of Predictions or a Consensus of Opinions?

Gari D. Clifford

Learning Objectives

Distinguish the main types of artificial intelligence and their impact in medicine
Identify opportunities for machine learning prediction in critically ill patients
Review scientific strategies to demonstrate the clinical impact of machine learning models in critically ill patients

Meta Tag

Content Type Presentation

Knowledge Area Administration

Knowledge Area Procedures

Knowledge Level Intermediate

Knowledge Level Advanced

Membership Level Select

Membership Level Professional

Membership Level Associate

Tag Informatics

Tag Monitoring

Year 2020

Keywords

AI in critical care

deep learning in healthcare

medical diagnosis

assistive technology

advancements in computing power

international competitions in AI

early sepsis prediction

	Society of Critical Care Medicine 500 Midway Drive Mount Prospect, IL 60056 USA Phone: +1 847 827-6888 Fax: +1 847 439-7226 Email: support@sccm.org	Contact Us About SCCM Newsroom Advertising & Sponsorship DONATE	MySCCM LearnICU Patients & Families Surviving Sepsis Campaign Critical Care Societies Collaborative	GET OUR NEWSLETTER
© Society of Critical Care Medicine. All rights reserved. \| Privacy Statement \| Terms & Conditions The Society of Critical Care Medicine, SCCM, and Critical Care Congress are registered trademarks of the Society of Critical Care Medicine.