Welcome to NVIDIA's AI podcast, where we explore the expanding world of artificial intelligence with the people who are developing and applying AI to the problems and opportunities that cut across our lives.
Today we have Mark Michalski, who's the director of the Massachusetts General Hospital Center for Clinical Data Science.
He's a physician. He's a tech nerd. We like that.
And he's taking IA on to healthcare, and that's what we want to talk about today.
So, Mark, welcome. Thanks very much. So let's talk about the Center for Clinical Data Science.
You guys are applying technology and all kinds of technology to the problems of healthcare.
But let's talk about how AI can even help the problems that we all face as patients and also, you know, for the physicians listening out there as practitioners.
Yeah, sure. Artificial intelligence is one of those technologies where if we began the list of applications, we'd be sitting here for the entire podcast.
There's not a place within medicine where artificial intelligence doesn't have a potential to make some impact.
So you say it has potential across all kinds of things in medicine.
How so? Because medicine, you know, it's been biology and chemistry And AI leans on data.
And so are we at a place now where there's enough data or we're gathering enough data to apply the tools of AI to medicine?
Yeah, I think we definitely are. I'm a radiologist by training, and my field is really all about data.
We look at images throughout the day, essentially interpreting that data and making meaning out of that data.
We take pixels on a screen, And then we translate them into meaningful information about how to treat that patient.
So we look at the screen, we see a tumor.
And we see that it's grown and it's strong.
That gives us input on the way that we change our therapies.
Radiology has been a data-driven specialty ever since effectively the digitization of these images, which has sort of proceeded over the past couple of decades.
In fact, nowadays, radiology is very data-driven.
Everything we do here at Mass General is essentially on the computer.
It's all digital. We look at the images on the screen.
We report out in text into a text reporting environment.
We're increasingly trying to make that report structured data.
So in a very real way, radiology has now been sort of the data sciences for a while.
But it's not the only specialty in medicine.
Increasingly, This is happening in pathology and many of the clinical subspecialties as well.
So what you describe, and when we're people looking at images, we don't call it image recognition, but that's what it is.
So in radiology, what we're talking about is image recognition of tumors or irregularities in an image period.
How is it that an AI system is able to do this?
I mean, we can teach systems to recognize cats.
Is it the sort of same thing where We're teaching a deep neural net to recognize either irregularities or specific characteristics in an image and in a tumor.
Yeah, so there's a couple things. I think the kinds of problems that are being solved today are in radiology are kind of analogous to what you're seeing in the computer vision community earlier on.
Things like segmenting or classifying things in an image. you draw a line around a tumor or you say that this is an image of a CT of the chest, a cat exam of the chest.
But I think that's what we're mostly involved with now in the ecosystem.
But I think that's just the beginning. once you're able to do some of these things you can do quite a bit more sophisticated things and things that start to permeate throughout the healthcare system.
So you can use machine learning to use those inputs, say the way that the tumor is growing or shrinking, along with clinical inputs like the kinds of medications that the patient is taking or the kinds of activities that they're performing at home.
You can incorporate genomic data, you can incorporate clinical laboratory values, the really interesting thing starts to become when you incorporate some of these broader, more heterogeneous inputs.
It sounds to me like what you're describing is the ability to pull together signals from a variety of different sources.
So it could be genomics, it could be clinical. response to a treatment, and it could also be electronic health records.
That leads me to think that what we're going toward or heading toward that AI helps us with, I should say, is this notion of personalized medicine, which is something we've been talking about for a long, long time.
So are we getting there quicker or finally, should I say?
Yeah, I think it's a great question. You know, personalized medicine was largely predicated on two things, the genomics revolution and the incorporation of electronic medical records in a broad way throughout the healthcare system.
And that sort of gave us this foundation of data in which to suss out differences between patients We could use that data to find things.
We could torture the data really to find things that we wouldn't otherwise be able to find. that were important for patient care.
I think one of the things that's important, that's different about what's happening now is it's not just about big data analytics and And it's not just about genomics.
It's about the application of machine learning to suss up very valuable signals in places where you wouldn't otherwise find it.
Chest x-rays, for example, we get hundreds of those a day.
And we're typically looking for, say, pneumonia or we're looking for a pneumothorax.
You can actually use those chest X-rays to determine things like how calcified the bones are.
And if you look at this from a population perspective, you can use machine learning to describe the radiolucency of those bones, it actually can give you a sense about Things that you wouldn't otherwise ask about.
Things like, is this patient osteoporotic?
This is something that we should we should investigate and use that kind of data in a way that at a population level that you never would have before.
And you never would have before because it's just too hard.
Again, I feel like. data, and not by design, but these things lived in silos, and to bring them together, is it just too hard, and was it just too many combining pieces of paper in a folder and kind of synchronizing things.
I think it's sort of like silos within silos.
The challenge, let's take that chest x-ray example.
Machine learning allows you to pull out of that chest x-ray bunch of data that you wouldn't have otherwise seen so you'll have the pixel data the data lives in that chest x-ray but Machine learning allows you to add some value to that by pulling out this metric.
This is an osteoporotic patient or... The BMI of the patient.
You can find those sorts of things. The size of the heart.
You can find those sorts of things with machine learning that you couldn't otherwise find.
As a patient, it sounds to me like I get more precise treatment for myself. but then when you expand that across millions of people, does that change how we think about treating all kinds of people who are sort of me in profile or genetically or otherwise, does it help push healthcare generally forward into a better place?
I think the answer to that is yes. For example, imagine that over the course of practicing, a clinician can see tens, hundreds of thousands of patients.
It's impossible to remember all those patients, but imagine if you had a tool that allowed you to tell you that the patient that you're seeing today has a hard diagnostic question that you have to solve.
Looks exactly like the patient that you saw 12 years ago, five years ago.
And this is what you did then. And this was the result.
And Not only is that patient back then, five years ago, resemble the patient you see today, but here's a whole host of patients that look very similar.
And what were the outcomes for those patients with various therapies or various treatments?
That's a really powerful tool for a diagnostician or a clinician.
As a diagnostician or clinician, how does it then change what your job is, or not even not what your job is, but how you do it?
Like, what does it free you up to focus on and do what physicians are best at?
I think it's one of the things that increasingly radiology, pathology, the diagnostic-facing specialties will ask themselves.
Stepping back, radiologists, their jobs were not always about interpretation of the images all the time.
And when I was a resident, what I did with a lot of my time was scroll through CTs or MRIs and try to pick out the abnormalities, describe them.
And that's a big part of what radiologists do.
But before the digitization of medical images, Radiologists were kind of known as doctor's doctors.
They looked at their radiographs and they met with other doctors and sort of described what this radiograph really meant.
Well, it turns out humans are not really good at some of the tasks that we do right now in radiology.
And I think where we are going is a place that looks a lot like our past.
We take the data that comes out of these images The images are, say, pre-interpreted or we get the interpretation with machine learning on the back end.
That gives us structured data. And making meaning from that data, describing the meaning from that data, is, I think, where we're going. we talk a lot these days about radiologists and pathologists, not just as people that are looking at computer screens or under microscopes, respectively, but people that can weave all the different threads of data together to paint a meaningful picture of a patient and describe what the best way to treat that patient is.
You are a physician and you just use terms like structured data and big data and meaningful data.
So you're clearly bilingual, let's put it that way.
What took you to this path as a physician Why did you think this was the place to go?
And then how unusual are you and how unusual is Massachusetts General and the Center for Clinical Data Science.
Are we going to see this kind of approach more broadly and fairly quickly?
I'll tell you, this year at our big conference in radiology at the RSNA conference, There was a tremendous amount of interest in machine learning from just about everyone.
It was the theme at a conference. I think there's a broad recognition that This is coming.
I think within radiology and pathology, tech nerds like myself are less rare maybe than they are another.
I didn't call you a tech nerd. You called yourself a tech nerd.
Let's be clear. I appreciate it. I wear the Namacurl with pride.
But I think that for us to do our jobs well as radiologists, to serve our patients well, We have to start thinking in these terms.
We are purveyors of data, and we have to know how to hold that data, how to interpret that data.
Increasingly, I think what I'm seeing around me is we're using the terminology that came from statistics and computer science disciplines in sort of my everyday practice of radiology.
As patients, how might we experience this shift?
I'm sure I'm not going to go into my doctor's office and say, okay, the machine learning results are in.
Let's look at them. I mean, how do you think we'll experience this as patients?
What I hope is that you'll first start to see your doctors look a little less stressed.
That's what I mostly hope for my colleagues.
So right now, what's happening is the amount of data coming into the healthcare practitioner's brain is increasing. substantially.
So we're getting more and more data that we have to interpret.
And human brains just don't scale that way.
So we're having this deluge of data, which is making it in some ways harder for us to figure out what's important.
What we need is a way of getting that data into a place where it's easier for us to interpret and convey.
We had this problem in genomics, right? It turned out that in genomics, we created these machines that all of a sudden spat out just all this data And we didn't have the mechanism on the back end to make meaning out of that data.
And that's part of why the value of genomics is only starting to hit really in an impactful way now because we have all this data and we're only now figuring out how to make it manageable.
You guys at Massachusetts General and your center are building the technology to are we going to start to face a sort of platform war?
I mean, it seems to me like if Massachusetts General has its own data and my hospital has its own data, That's not kind of where we want to go.
We want this data to be shared and, in some sense, accessible.
So how does that play out kind of as a platform for, In the world of medicine, in the world of hospitals for that matter.
Yeah, well, I think the platform point is a really good one.
This is one of the grand challenges that's facing us as we move forward.
How do we take these models? and distribute them in a way that doesn't land us in something of a standards war.
We've seen that before in healthcare, and how do we avoid it?
I think that we at the center are really thinking about this as a global effort.
I pretty strongly believe that big data has to be big.
And if we create a system at Mass General that helps us with diagnosis here, It may be useful for our patients, but what we really want is for these systems to be useful for all patients.
Geographical variation in patient populations is really important.
And the different healthcare systems that take those models and sort of use them for patient care, it really requires us to think ahead of time today.
I think at Mass General, we have a great patient population and a huge one.
But for these systems to be really useful broadly, not just here, but throughout the world. we have to think about big data being big.
And that means more than just mass general data.
In fact, we don't want to get into a platform war or a standards war.
I think that that would not help anyone.
These sorts of centers will start to rise up in other places.
And I think the important thing is that we all sort of are rowing in the same direction.
I hope we are all rowing in the same direction.
It sounds like you guys are at the center of making sure that happens.
You mentioned radiology and pathology. Where do you think that machine learning can go and goes next?
What are the opportunities when you step back and say, ah, gosh, it'll be great.
Let's knock these things off first, and then we'll go after these other areas of medicine.
Yeah, you know what I'm really excited about is when we get all this heterogeneous data together.
And we're able to use the entire picture of the patient to give us a sense about what to do with our patients.
My belief is that it really will take all the data that we collect in the health care system and beyond to really optimize our care.
I don't want to find out that I'm sick because I start showing symptoms.
I want all the data that the world collects about me to proactively tell me, you better get checked out because you're going to start feeling sick pretty soon. you better go get a CT scan or you better go get an x-ray because you got to get something checked out.
You got to see your doctor. That's what I hope for because I think We do collect a lot of data on ourselves.
And right now, I don't know that we're getting all the value that we could out of I really do think that that vision, which is proactively knowing what you're at risk for and what you should do about it.
I think that's the real promise here. You described something, too, where we as patients have both a responsibility and a way to participate in a different experience. manner than we have thus far in our own healthcare.
And I think that that's exactly what this offers us as patients too.
I think so. And I think what I've found over time is that know all the details of what's going on in that CT scan, that MRI.
They really want to know what that means for their lives.
And for many patients, that means sort of controlling the data, actually understanding The things that data scientists understand, which is what you're at risk for and what you're not.
I think this is an incredible time for people to really understand what it means when you have an abnormality on a laboratory exam or an imaging exam.
Or like as you said, what I can maybe do to fend things off or to not get that far down the line because I've got some early signals that, kind of relate to all kinds of other signals out there in the world that these kinds of systems will bring to us.
Well sure. We've talked about the implications for treatment and diagnosis in the United States in particular, but also throughout the world, the cost of medical treatment keeps rising.
So can machine learning help us with that?
Can it change the economics of healthcare?
I really think so. Right now, it's very, very difficult, very, very costly to interpret some of these examinations.
So just, again, in my field, radiology, We have to go through thousands of these images.
And that's part of the on a typical MRI scan of, say, the head.
So that's a real cost adder. The potential here is that by using machine learning, you can reduce the energy required to do the interpretation. and thereby reduce costs.
So I think that that's a real important feature to recognize.
But beyond that, If you reduce costs, then it actually allows you to get those examinations more frequently, say, if a machine learning algorithm tells you that you need to be screened more frequently.
Oh, right, right. So say you're at risk.
Now, if an MRI is no longer $1,000, but it's $100.
Maybe you have a higher resolution temporally, you know, a higher temporal resolution on your scan.
Maybe now, What you're doing is you're getting scanned on a regular basis for screening because you're at risk.
And maybe you know that now because a machine learning algorithm is taken into account.
All your clinical information, your genomic data, even your social feeds, and knows this about you.
It's an interesting idea. And you catch it earlier and hopefully it's easier to treat and or less costly too.
You're applying machine learning to radiology and pathology, and there is this belief, and I'm with you here, that it's going to translate to better outcomes Is there evidence already that it's changing how we treat people and that in some sense, we're getting better outcomes?
I think that's the hope that these systems will impact outcomes.
I think what we are seeing now is some of the things that we do on a day-to-day basis in healthcare, that these machine learning algorithms can help us do them quickly and potentially better.
So I can give you some examples. We have already seen machine learning algorithms that do the job of picking up pulmonary nodules in a CT scan. screening for lung cancer. it performs that task pretty well.
We've seen some other examples of systems that look for cancer cells under the microscope. on pathology slides. and do that pretty well on par with humans.
If you look at the ImageNet competition over the past couple of years, you'll see that These days, many of the systems are beating human performance, and we're also starting to see that in translation.
Well, I think there are some tasks in healthcare that are analogous. if we're working at this as long as people in the ImageNet competition have been working at the ImageNet competition. and there's a good chance that we're going to start to see systems that do just as well as us or better.
The hope there is that that translates to better outcomes for patients.
And this doesn't mean that you're taking people out of the equation entirely.
It sounds to me like, look, you can screen way more images they get flagged and they get sort of run up to the next level and somebody else takes a look at it and decides how to move forward.
Is that fair to say? Yeah, that's absolutely true.
I think for the foreseeable future, there are some tasks that we do as radiologists that machines can really help us with or as clinicians.
And there are some things that for the foreseeable future, it will be very hard to automate. automate the things that you can, especially if they do as well or better than people. that should translate to better outcomes for everyone.
And frankly, it might make us as practitioners happier.
We have this deluge of data and patients.
And systems that help us to deal with that deluge are real critical.
We need them now. So I'm not worried about... doctors being replaced by robots anytime soon.
As a person on the other side of the doctor-patient relationship as a patient, I look forward to a time when a doctor can actually slow down and take some time because maybe they're not cranking through all those images and trying to synchronize all that data.
But that's being done for them in the background, as it were.
And they can face people face to face and do what doctors are best at.
Yeah, I think ironically, one of the things that patients are noticing about their doctors is they are talking more with their computers than they are talking with the patients.
And they're sitting and looking at a computer screen. ironically, these systems, which are digital systems, will allow your doctor to talk with you more about what this data actually means.
Well, I hope we get there. And it sounds like with your help, we will.
Mark Michalski, thank you so much for joining the AI podcast.
Really my pleasure. Postscript here. If you are listening to this podcast on iTunes, Google Play Music, or whatever your favorite podcast player is, do us a favor and leave us a review.
Thanks for listening.
