Better Medicine Through Machine Learning: Suchi Saria (Transcript)

Full Text of Machine Learning Expert Suchi Saria’s talk titled “Better Medicine through Machine Learning” at TEDxBoston conference.

Listen to the MP3 Audio here:

TRANSCRIPT:

Suchi Saria – Associate Professor of Machine Learning and Healthcare at Johns Hopkins University

Let me introduce you to Mrs. Manny. She came to the emergency room. She is fifty two years old. She came to the emergency room with a foot sore. Doctors investigated her foot sore and she ended up staying there in the hospital for 22 days.

Here’s what happened.

When she came to the emergency room for a foot sore, they inspected her, they saw no real reason for medical concern but they wanted to monitor in case her foot sore was infected. So, they put her in the general ward.

On day three, she starts developing symptoms of what looks like mild pneumonia. They give her the usual treatment of antibiotics and all’s good but then her condition starts to worsen.

On day six, she develops what’s called ‘tachycardia’ that means in medical speak, her heart rhythm has accelerated dramatically. She then has trouble breathing.

On day seven, she experiences septic shock; that means her body is in crisis. Incidentally, mortality in shock is one in two. Now, it’s only at this point that the doctors get really concerned and they transfer her to the intensive care unit.

ICUs are the units where the most critically ill patients get cared for. Here, they give her every possible treatment to stabilize her but her condition only worsens.

First, her kidneys start to fail. Then her lungs fail and on day 22, she dies.

Mrs. Manny did receive the right set of treatments. The problem is, she received them only too late. What Mrs. Manny experienced was an infection that turned into sepsis.

Let me tell you a little bit about what sepsis is.

Sepsis occurs when infection releases chemicals in your blood to tackle the infection. So, your body releases chemicals to fight the infection. Now, this chemical can trigger a negative inflammatory response. When this inflammation triggers this negative inflammatory response, what it can then do is cause a cascade of changes, leading your organs to fail, leading to death.

Sepsis is the 11th leading cause of death, more than breast cancer and prostate cancer combined. Turns out, sepsis is preventable if treated early, okay?

So, then what’s the catch? Doctors find it very hard to recognize sepsis. In fact, a Harvard study shows with 93 leading academic experts that when they were given several cases of patients with and without sepsis, they couldn’t agree.

Two years ago, my nephew, he was admitted to the best Hospital in India and he died of sepsis. My family was devastated.

I’m a machine learning expert and what I do is study ways in which we can use large messy datasets to enable intelligent decision-making. So, natural question for me was, could machine learning have helped? Could machine learning have helped Mrs. Manny and my nephew?

So, this led to a massive effort with my colleagues at Hopkins to design what we call ‘the targeted real-time early warning system’ or TREWS based on machine learning. I’ll give you a sneak peek into what TREWS is and how we’re using it to tackle sepsis.

Let me take a step back and tell you a little bit about what machine learning is and what’s AI.

Artificial intelligence is a field of study where we teach computers how to learn. Okay? Just like you teach your kids. Machine learning is one way of doing this, by designing code or programs that teach computers stuff over time by interacting with the environment or watching. Okay?

So, I’m going to show you a video of some robots learning how to walk. I find it funny how it shudders. So, you’re probably now thinking this is hopeless.

Well, so the question is, how can we teach robots or machines how to walk. Intuitively, you can think of it as designing a game. The goal of the game is for the computer or the robot to learn how to walk for as long as possible without following. Okay?

So, to do this, first we have to design, write down the goal in a language the computer understands. For this, we’ll use math. Okay, so now you’re wondering, well how do we write the goal of walking without falling as long as possible in math?

Well, that’s often hard for different tasks but you can think of it as writing down a formula and what this formula does is it scores.

[read more]

So, in the case of walking, it’ll score every move the robot makes. If the move it makes helps the robot walk, it gets a high score. If the move that the robot speaking makes the robot unstable, it gets a low score.

And now the robot’s goal is to experiment with the sequence of moves in order to be able to maximize its score. So, how does it know which moves to try? Right?

Well, there are two strategies for doing it. First, it learns by interacting with the environment, okay? So, here the robot will just make a guess. It guesses, it makes a move. If the move gets a high score, that’s positive feedback and the robot builds on it. Okay?

The second strategy is by watching other robots. In other words, the robot finds data from past robots that are similar to this robot. It watches what moves that robot did when it was in very similar positions and now it emulates or replicates those moves. Okay? So, those are the two strategies.

So, now I’m going to show you a video of robot learning how to walk using the strategy I just described. Okay, so in the beginning it’s going to look hopeless but I promise you it gets better. And just to be clear, this the skeleton of the robot and so this is not a human animator going there and just moving or animating this video. This is really the robot, the algorithm choosing which moves to make by moving the joints of the skeleton that you’re seeing and you can see it’s already getting better.

Now, suddenly the robot’s able to walk and run for a lot longer than it was doing, right? So, essentially the basic principle is as follows.

You figure out a game that the computer can play, you write it down using a language it understands and then we train it to optimize the score, right?

This is how we teach cars how to drive, computers how to play the game of go, an Alexa to understand say your preference of coconut water.

So, let’s go back in our case, the problem of sepsis. So, the goal here is to identify sepsis as quickly as possible, right? And for this, TREWS learns by watching.

In other words, using data from past patients. This avoids the need for TREWS to have to experiment on new patients, right? So, to do that, what are the pieces TREWS needs to do?

So, one big change that has happened in medicine that’s interesting to note is, in the past five years, the introduction of electronic health records in EHRs, every single measurement, every single lab test that is ever done when you walk into the clinic or you’re in the hospital gets collected.

TREWS analyzes this data from thousands of patients to identify subtle signs and symptoms that appear in patients with sepsis than those without, okay? But that’s not alone. What TREWS also needs to do is to figure out how to think about every signal in the context of every other signal.

Let me give you an example.

Let’s look at the example of creatinine. So, creatinine is a waste molecule, okay? And your kidneys filter it out, but here’s the catch.

So, when your body is septic, it affects your kidneys, it deteriorates your kidneys’ ability to filter out creatinine so creatinine level rises, but there are many other things that can affect your kidney’s ability to filter out creatinine, for example, if you have chronic kidney disease, you’re very likely to have high creatinine levels.

So, now what truth has to do is to figure out, is your creatinine high because of sepsis or because of chronic kidney disease or the numerous other factors that relate to high creatinine levels, but that’s not enough. It needs to do this for every single signal that exists in the electronic health record and TREWS is thinking about every signal in the context of every other signal to identify signs and symptoms that occur more often in patients with sepsis than those without.

Let’s return to Mrs. Manny. Research by Kumar and colleagues have shown that for every hour treatment is delayed, mortality goes up by seven to eight percent. So, timing is critical.

We went and took Mrs. Manny’s data and we ran TREWS on it and here’s what we found. TREWS would have detected Mrs. Manny’s sepsis 12 hours before doctors currently did. As my clinical colleagues would say, that is the difference between life and death.

Last year, we showed using data from 16,000 patients that TREWS on average would have detected on most patients, on average, more than 24 hours prior to the shock onset. That’s nuts, 24 hours, in two-thirds of these patients, their sepsis was detected prior to any organ dysfunction, whatsoever.

And to put this result in context, that’s 60% increase in performance over state of the art. So, what TREWS is really doing is doctors a much longer window to come in and intervene in order to prevent organ dysfunction and mortality.

This year, the independently validated TREWS in data from Howard County General Hospital in Maryland and now we’re working to do real-time integration in order to make something like TREWS available to every doctor at Hopkins.

I’m also really excited because after we’ve published our papers, several other health systems are now already implementing the published version of TREWS in order to be able to develop it in their own environment.

So, I’m going to highlight like a few, perhaps three salient characteristics that I think makes a strategy like TREWS very powerful, okay?

First, TREWS runs 24/7. What it does is it gives doctors a second pair of reliable eyes, right? Two, it’s hard to scale-up doctors. It’s easier, I think much easier to scale up computers and what TREWS is really doing is allowing us to get expertise from the best doctors everywhere. Here’s the third one, which I think is very interesting. In many cases, like we see in sepsis, we might not need new measurements. The signs and symptoms were already in your data and what TREWS is really doing is discovering these signs and symptoms to learn something that we couldn’t see by eye.

Finally, there’s been a lot of buzz about big data and I want to make a little subtle point about a technical problem that I think TREWS is solving that is very interesting. TREWS would be able to learn much faster if it had a lot of data on you or it could get more data by experimenting on you but we don’t want that, right?

So, what TREWS really has to do is leverage your limited data to figure out what’s right for you, right? So, in other words, what TREWS really has to solve is a challenging small data problem.

In other words, it has limited data on you and has to figure out what is the right treatment for you and for that, it leverages vast amounts of data from other patients and figures out what information to borrow in order to make these assessments reliably and precisely.

So, I also want to tell you a little bit about how the strategy is not unique to sepsis. So, very broadly, if you think about it, in many diseases essentially, where you have profile of symptoms and the response to treatments varies a great deal across individuals, you can use the strategy like TREWS in order to target treatment.

So, you’re wondering like, for example, if you consider cancer, diabetes, multiple sclerosis, Parkinson’s, lupus, so there are many such diseases on which a strategy like TREWS is amenable.

In fact, in our own lab, with experts in rheumatic diseases or immune diseases in particular, we’re looking at how in scleroderma, for instance, we can use strategies similar to tools to avoid giving strong immunosuppressant’s to patients who don’t need them.

Other colleagues, this is William Pelham, Susan Murphy and their team, they’re studying kids with ADHD and looking at how using similar data-driven strategies, they can identify when kids can be benefit from behavioral therapy and we can avoid the need for giving them psychostimulants altogether. So, this strategy is very powerful.

So, I was speaking about sepsis, so let’s go back to sepsis again. So, I said it was sepsis Awareness Month and the CDC has declared sepsis to be a medical emergency, rightfully so.

Remember, 750,000 people annually are affected by sepsis. A patient’s family recently asked me, what will it take to bring this to a hospital near us? I think that can be done.

In fact, it can even be done within a year, but we don’t want to stop there. We want it to be possible to bring strategy like TREWS to hospitals everywhere. And so the question is, to do that, what will it take? Right?

So, I think the three key things we need your help for. One, we need super smart engineers to be working in healthcare. We need your help in building and scaling up such technologies, don’t go to Wall Street, healthcare needs you, right?

We need policymakers to create incentives to open up electronic medical records. As an expert at a leading health institution, it’s taken me more than a year because the EMR is so closed, in order to be able to figure out how to implement rules against the EMR. It really should be easier than this.

Three, we need a healthcare system that’s based on quality. Our current healthcare system is incentivized to optimize volume rather than quality.

Right now, you can choose which restaurants to go to, based on the quality of food. Should you be able to choose the hospitals you go to based on quality of care? Part of the problem is that quality data at the moment is not very visible to consumers and we really need to make a bigger effort to make this quality a visible, so that you can choose based on quality.

So, to summarize, sepsis is one preventable killer. In many pressing medical problems, like we saw in sepsis, the answers for knowing whom to treat, when to treat and what to treat with, might only be in your data.

Sometimes, I wonder if we had done this work two years earlier, if I could have prevented my nephew’s death. I can’t wait for this to be the way medicine is practiced. Thank you.

Resources for Further Reading:

Think Like Eagles: TD Jakes (Full Transcript)

Matt Beane: How Do We Learn to Work with Intelligent Machines? (Transcript)

Martin Luther King Jr. on Why Jesus Called a Man a Fool Speech (Transcript)

Chronic Stress, Anxiety? You Are Your Best Doctor: Dr. Bal Pawa (Transcript)

[/read]

Better Medicine Through Machine Learning: Suchi Saria (Transcript)

Resources for Further Reading:

Related Posts