At the second Science on the Hill event, AI, Robotics and Your Health, experts from academia and the private sector talked with Scientific American Editor in Chief Mariette DiChristina about the future of AI and robotics in medicine.
At the second Science on the Hill event, AI, Robotics and Your Health, experts from academia and the private sector talked with Scientific American Editor in Chief Mariette DiChristina about the future of AI and robotics in medicine. Panelists included Pedro Domingos, University of Washington, and author of The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World; Suchi Saria of Johns Hopkins University who is working on diagnostic and treatment planning tools; and Pamela Hepp of Buchanan, Ingersoll & Rooney who is looking at data security, healthcare regulations and digital health records. Hosts were Scientific American, Nature Research (part of Springer Nature) and California Congressman Jerry McNerney.
Steve Mirsky: Welcome to Scientific American Science Talk posted on June 18, 2018. I'm Steve Mirsky. Earlier this year we heard the first in a series of discussions called Science on the Hill – Capitol Hill in Washington, D.C., that is. Scientific American, Springer Nature and Congressman Jerry McNerney of California put on that event which focused on sustainable energy.
On June 14th we all gathered again in the Rayburn Building on the Hill for our second event called AI, Robotics and Your Health. What follows is an edited version of the hour-long conversation moderated by Scientific American's Mariette DiChristina. You'll hear mention of HIPAA – H-I-P-A-A. That's the Health Insurance Portability and Accountability Act, which aims at keeping our medical records secure and private.
Here's Mariette DiChristina.
Mariette DiChristina: Good morning. Thank you so much. Welcome. I'm so glad to see everybody here. My name is Mariette DiChristina. I'm the editor and chief of Scientific American. By the way, if you don't know it Scientific American was founded in 1845. And so it's the oldest continuously published magazine in the United States. We've been working Representative Jerry McNerney who I'm going to invite up in just a minute to start this “Science on the Hill” series, to bring information about the latest innovations to the people who have to work on and decide how best to regulate or to manage those innovations for society's benefit.
Our first topic was on new energy technologies for a sustainable future. But today we're going to turn to the topic of AI, robotics and health care. I mean from R2-D2 to the Roombas that vacuum our houses today, from Evil HAL to things like IBM's Watson helping with diagnoses we've seen robots and AI jumping from science fiction into the real world and into our lives directly. And that's why it feels to me like it's never been a better time to discuss this topic that we're going to look at today.
And now I'd like to introduce our host Representative Jerry McNerney for some opening remarks. Please welcome Jerry McNerney. [Applause]
Jerry McNerney: Well good morning. I just have to say it's always a pleasure to be around scientists and lovers of science. I myself am one of two actual scientists in Congress. I have a Ph.D. in mathematics and differential geometry. I spent a career developing wind energy technology. I wrote the codes that calculated the loads on dynamic wind turbines. And that was a lot of fun. Today we're going to be talking about artificial intelligence and health care.
Now as we know science isn't the answer to everything. But having an informed decision-making process is helpful and would be more helpful if we had more of that around here specifically regarding legislation. I am working other members in developing ideas of how to enhance the artificial intelligence in medical applications. With that I will turn it over to the speakers and I look forward to some good conversation both from the speakers and from the audience. Thank you.
DiChristina: Thank you so much. [Applause] So it's high time I introduce you to our wonderful speakers. I'm going to go from my right to left. Next to me is Suchi Saria, the J. C. Malone assistant professor of computer science statistics and health policy at Johns Hopkins University. She's also a founding research director of the Malone Center for Engineering [in] Healthcare at Johns Hopkins. Welcome Suchi.
Next to Suchi is Pedro Domingos, Professor of Computer Science at the University of Washington and author and co-author of more than 200 technical publications on machine learning, data mining, and other areas. Welcome to Pedro.
And then next to Pedro is Pamela Hepp of the first Buchanan Ingersoll and Rooney in Pennsylvania who concentrates her health care practice on issues affecting large health care institutions including complex transactions, medical staff matters, and regulatory issues. Welcome Pam.
So I think it would be really great just by way of people getting to know you a little bit better if you each wouldn't mind maybe a couple of minutes on a little bit more about your background and the areas that you're really focused on so people know who they speaking to. Suchi do you want to start off?
Suchi Saria: If you are as a patient approaching the clinic, you're visiting with your doctor, and you've had you know a specific set of lab tests and you're wondering do you have a particular – a diagnosis or not? Are you going to respond to a drug poorly or not? Or is this going to have a secondary affect because of some other comorbidity you have? Now what's possible is that you can look at data from millions of patients who came before you and watch what happened to them over time in response to similar therapies that were given or to observe what was done to them and how they reacted to figure out exactly what is the right thing to do for you?
And that's a really, really powerful paradigm shift that's happening in medicine. And our ability to do that – our ability to build infrastructure that allows to safely and reliably use the data to inform decision-making is I think going to be game changing in medicine. That combined with one, another movement, which is the introduction of new sensors. So we're not only starting to digitize tons of data that's already available.
We're also becoming increasingly creative about new data sources that previously we had not thought to collect. So for example there's work done in my lab where we're showing Parkinson's as a disease. We only knew how to measure using these subjective 30-minute tests that people would do. And now it's becoming possible where we've built in new ways by using sensors at home. It's noninvasively watching what the patient is doing – how they're reacting, how they're moving.
And it turns out those actually give you very powerful indicators of symptom fluctuations which can be used to tailor your therapies and detect early if you're deteriorating, promptly timely treatment. So anyway I think a collection of both existing data, new data sources, new sensors, and a science that allows us to exploit all of this to figure out for any given patient at any given time or any given individual – not just patient – because it's all about wellness and disease. For an individual what is the right thing to do?
DiChristina: Thank you. Pedro when you talk a little bit could you also define AI for people? I realize we haven't said what it is.
Pedro Domingos: Sure. So here's what AI is. AI is trying to get computers to do things that used to take – or at least currently take – human intelligence, things like reasoning, planning, decision-making, vision, understanding speech and language. And one very important one is learning. Learning is really the ability that drives all the others. If we had the robot that was as intelligent as humans but didn't learn as well a minute later it would already be falling behind.
So machine learning is the field of area that I in particular focus on. And it's really the field that's been enabling all of these advances that we're seeing today. When I did my Ph.D. 20 years ago machine learning was a very obscure field. One of the main things that I did was apply new learning algorithms to a whole slew of medical datasets where we wanted to diagnose different things based on the symptoms of the patients.
And to my shock at that time we could already do this better than the human doctors. Today the systems are better still but they get not widely deployed. So I think there's a lot that can happen with AI in health care. And hopefully we can all help to bring it about.
DiChristina: Pamela Hepp of the firm Buchanan Ingersoll and Rooney in Pennsylvania?
Pamela Hepp: Hi. I am a health care attorney. I have been my entire career. I received my master's in public health with my law degree and my focus has always been health care. From a regulatory perspective a lot of what I do is in the area of HIPAA and data security including not only HIPAA but the state law implications when we're dealing with super protective types of information such as mental health information, HIV related information, drug and alcohol information – all of which comes into play as we're sharing data.
Another part of what I do is transactions. And that has included creating health information exchanges, which are organizations that take entities that have electronic health records and allow them to share data in a health information exchange typically at a regional level. There's movement to try to do that at the state level. And nationally it really hasn't exploded to the extent that I think it's desired.
But those steps are occurring where health information exchanges are existing to exchange data between providers so that they have the information they need when a patient walks in the door about their medical history and not just the history of their care at their particular facility.
DiChristina: Thank you. So we've heard a bit about processing data in better ways whether it's existing data and new data, adding learning to that. And then a bit about – a little bit about – which I want to come back more on about how do we share that information or regulate it properly so that we get those better outcomes? There are a couple of other technologies that I thought we should at least introduce to the group.
Pedro you mentioned for instance to me care robots and how cellphones can even be used to help improve health care. Could you talk a little bit about that?
Domingos: Yeah so a big application of robotics is elder care. For example Japan is the country that's most advanced there. They have of course a big shortage of manpower. And it's actually becoming quite common for robots to take care of elders. And you'd think that they would maybe be a little put off by that. But on the contrary, they seem to really like it. The robots are actually designed to be attentive to people's needs and to interact with them well.
In fact the robots these days are actually better in some aspects with the emotional interaction with humans than they are picking things up. Picking something up is actually harder than reminding someone to take their medicine in a way that they feel good about it and so on. Robots for elder care actually – And of course as the population ages then you know there are going to be a lot of people with _____ and so on. Robots for elder care are going to be a huge hit soon.
They're going to be very important to the well-being of a lot of people. Another big one is mobile right? Mobility and cellphones are starting already to completely revolutionize health care because they have all these sensors that are on all the time. They have an accelerometer. They have a microphone. They have GPS. They can register your vital signs continuously. This idea that once a year you go to the doctor's office and you get a few readings taken is – That belongs to the last millennium.
We already have someone who said that the cellphone saved their life because they were starting to have a heart attack and the cellphone noticed something unusual and you know. And actually I think before long what's going to happen is your cellphone if you're having a heart attack will call 911 before you do that. So 911 will come to you because the cellphone actually made that happen.
There's this competition to try to design you know more censors to put on the cellphone to record the things that need to be recorded. So we're looking at the world where your health is being continuously monitored and the interventions can help as soon as they're necessary. Of course the quality of understanding of what's going on with your health will be much greater. There will be a lot of AI machine learning on top of that to figure out what's going on and what needs to happen.
And all of this is still in its early stages. But once it's done I think it will be a huge revolution.
DiChristina: Thank you. I'd like to dive in just a little bit more about some of the benefits of AI itself in research areas. Suchi maybe you could speak a little bit about what AI and other technologies are going to help research, to develop ideas and so on. And Pedro feel free to add to that as well.
Saria: So I want to elaborate a little bit more on Pedro's point of in order to talk about this in terms of what is AI and what is machine learning? I think there is sort of this sci-fi version of AI machine learning as some magical thing. So the mission is very clear right? The idea is to build machines that can think and act and sense and perceive like humans do. And when we use the words "like humans do" or goal isn't to exactly build it the way humans do it.
It's just to draw inspiration from the way humans do it. And really the bigger mission here is more that well are there ways in which we can write programs that can effectively --? You know if we could in a room see that there's a red apple in the back couldn't we write programs for a computer to use cameras to similarly sense the room and identify that there's an apple in the background? There's a very simple, curious mission here which is one of building these capabilities that allow you to sense, act, perceive, and reason.
So reasoning is probably the most interesting one. And so reasoning is of the kind where these very, very simple machines, or very, very simple algorithms do this version of associations which is saying when I see pixels that look – Pixels are sort of in a camera image you have colors that you are sensing. And if the color is the value – The numerical values look like this and that. If this and that, then maybe it's red. And maybe if there's a collection of red then maybe it's an apple.
That's effectively something very simple. You're learning these rules. But as you push the boundaries further the kinds of things we do are things like well we take into account alternative reasons, alternative causes. For example oh well it's in a room and it's red. There is some yellow next to it and some green next to it. So much more likely that it is a fruit basket because those things go together and make sense.
And in the health care setting this is needed. This is needed in a very powerful way because just because you see a measurement as high doesn't mean that measurements are just it's this one disease. Very often it's the case that – Like an example would be creatinine is a measurement of kidney function. If creatinine is high it means maybe your kidneys are not doing so well. But the question is why? It turns out maybe it has nothing to do with one particular disease.
There could be five other diseases, or that you could've gotten a specific medication that temporarily increases your creatinine level. So the ability to take multiple things into account, reason about how they affect each other and influence each other and reason is effectively I think one of the more powerful areas where I think AI will make a lot of progress. And the ability to do this in a way that you're doing it from data. So it's one thing where you're building reasoning engines because somebody else is telling you how to reason.
So you learn from your parent when you say well okay the reason I believe it's not the red apple was instead it's like a red toy is because of X, Y, and Z. And you learn it by rote. But instead what machine learning allows is that we take tons and tons of data and we learn. We build learning algorithms that learn reason from data to say well the last time you saw something red what else was next to it? And therefore what else could be happening that can explain away this?
So I think in terms of building machines that can reason is a very powerful construct. And once you have something like that you can imagine making all sorts of discoveries. It has nothing to do with just health. It can be discoveries about how the world works and why it works the way it works. And the more data we have the more experience we collect. That experience allows you to discover. So that's effectively I think what's exciting about it.
DiChristina: If you have a machine that can reason and something that has infinite patience – You've talked to me about things like developing new drugs and other ways that these sorts of machine learning can advance things. Could you speak to that a bit?
Domingos: Yes. So one very exciting example of what can be one with AI today is to actually design your drugs in vitro right? So there is to be in vivo – in the living organism – sorry and in vitro. But now we can design an in silico. In vitro was in the lab. The way drug companies used to design new drugs was to be trying out new things in the lab and seeing which ones work. But the problem is that that's mostly run its course. The probability of finding something really important in that way has dwindled. Luckily there's a new way which is we actually use machine learning to figure out from the structure of a compound what its effect is.
And then we can actually try billions or trillions of different variations to see oh we want something that binds to the AIDS virus in this way. We try to design a drug that does that. It is actually quite conceivable that in the not very distant future when a patient has a tumor we can try all the different drugs that exist for that tumor. But the problem with cancer is that every tumor is different. But it's actually quite possible in the near future that there will be a drug that is designed specifically of that patient, for that tumor to kill the tumor cells without harming the health ones.
So this is kind of like one person – One of the big themes of AI in medicine is personalized medicine. It doesn't have to be the same recipe for everybody anymore. It can be tailored solutions. And one good example of that I think is literally designing drugs that are for a particular person for a particular problem. Or even more broadly these days a drug to get approved has to basically work for a lot of people and not harm anybody we're actually moving to a setting where the drug is useful if maybe it only helps five people.
And maybe it harms ten people but as you know which ones – If you know which ones it is, if you can predict, then you can just use it for those five percent of the people where it's helpful. But of course we also need regulations to allow that to happen. We need – For example something that machine learning enables that is very important is not this idea of we just do one drug trial with a few thousand people and then the drug is deployed and we never hear about it again.
It's like where things are being continuously rolled out and more patients are trying them and we are continuously getting the feedback from those patients and refining our models of what that drug is good for and what it isn't good for. So these are just a couple of examples.
DiChristina: We've heard how elder care can be done with robots even already today. And people – some of them at least – seem to like it. We've heard about cellphones for health care. We've heard about going form in vivo to in silico and personalized medicine where you can treat a specific tumor. Or I love the idea also of identifying whether a drug could be useful for five percent of the people and not used in the others, which could save just so much. And I'm thinking Pam you work on all of these issues.
Hepp: Let me first start from a policy perspective because what Pedro just mentioned with respect to research and development of drugs there's a constant struggle to contain costs in health care. And the reimbursement system continues to change. The Affordable Care Act was the most disruptive but over time the reimbursement system has continued to evolve in ways to try to contain costs. And one reason that costs are expensive is research and development related to drugs for example.
And so the research applications can actually help to reduce costs. The reimbursement system changes with respect to value-based medicine can also help to achieve cost savings to the extent – only to the extent that it is improving outcomes, that it is keeping people out of the hospital. It's keeping them healthier. It's keeping them in the home. And to do that it requires a robust amount of data, not only about that specific patient, but to predict how that patient is going to react to a specific treatment or to remain compliant with a specific regimen requires predictive analysis based upon the population.
It's a population health analysis dealing with all of this data. But to share that data has HIPAA implications. And it also has state law implications. But to the extent that some may say this really is research that you're using to develop a treatment protocol for me in particular research requires specific consent by the patient. So I think there is going to be some need to determine whether or not consent is needed from the patient to be able to share broadly this information.
But even more particularly when we're dealing with the super protected types of information, which often for predictive analysis is really the type of information that's most vital. And that is the mental health of substance abuse. That type of information is going to be vital to this analysis. And so we need to be thinking about how we structure the sharing of that data and what consents may be required to do so.
And again while those are issues that we need to think about having said that one very powerful use of artificial intelligence is to prevent cyber-attacks because artificial intelligence can be continually gathering data to see where there are changes in activity in a health information environment and react. So it has the power to prevent these issues from a cybersecurity perspective. From a policy perspective there are a lot of issues to take into consideration.
DiChristina: We're talking about data as if it's an unalloyed good. How do we know that it is? I mean how do you know – Is it just the volume of data gives you confidence that you can rely on it as you're processing it to do other things? Or what are some of the ways that researchers look at that?
Domingos: Actually that is not an unalloyed good. It depends entirely on what you do with it right? There is information that allows it to make better decisions depending on what your goals are. If your goals are bad you'll do bad things better. So there are a couple of definitions. One is like the quality of the data. If the quality of the data is not very good it doesn't matter that it's very large. Another one is the volume. The volume is actually a double-edged sword.
In a way having more data is always better. So for example having more patient records allows it to build a more accurate model. Having more information about each patient you'd think would be better. But actually it can also – It's actually a balancing act because too much of the wrong kind of information can actually confuse the system. It's actually one of the things that machine learning systems have to do all the time is figure out what data to ignore.
And also a lot of that is very specific details that don't generalize. So learning is actually – I think this is true of human learning as well. Learning is as much about forgetting what is not important as it is about remembering what is important. So it's great to have a lot of data. It's like having a big river. But if the river floods your house then actually it wasn't for the better. So we have to figure out how to make the best use of it. And that's actually what a lot of AI is about these days.
DiChristina: I think I'd love to let you all ask some questions.
Gonzalez: Hi, I'm Graciela Gonzalez from the University of Pennsylvania. I'm at the [Perelman] School of Medicine and do research on social media for health. A lot of people are tweeting about their medications, what it causes. And there is also some policy that is going in that direction as to how can we use this information that's made publically available voluntarily by people that are concerned about their health?
I think this participation can only be enhanced and facilitated if researchers can work with it. And HIPAA has its place and it's important. But it has to be rethought.
Hepp: Yeah and that actually brings up a good point because a lot of the data that is coming in will come from fitness trackers for example or personal information that is not protected by HIPAA. And those personal types of records may be to the extent that it gets into decision-making – If it's a tracker that is decision-making the FDA may regulate it but otherwise those software platforms are not regulated. That information is not protected.
I'm not suggesting that it should be but not all of the information that we're talking about is necessarily subject to whether it's a federal or a state law protection unless it makes its way into the medical record. Once it's in the medical record then it is protected. But you're absolutely right. Not all of this information that you're looking at is necessarily regulated.
Question: There's a tension that I detect between big data and individual medicine. You have all this data and it gives you trends and the ability to understand. But you want to develop medicines for an individual as Pedro said. So I'd like somebody to address that. Is that an actual tension or is there something there? And the other question – Pedro mentioned robots for senior care in Japan but what about robots for surgery?
Are there limitations to how much can be learned? Are actually robots someday going to be able to actually perform surgery on their own? Thank you.
Domingos: So on the first question what big data creates? I wouldn't say so much there's a tension is there's an opportunity. If you have 3,000 data points you can't do anything personalized. You can't do anything personalized. You can only do things at the population level. When you have big data what that means is that you have a reasonable bulk amount of it about individuals. So when you have a lot of data then you can start to model individual people.
Right now we have a model of the population as a whole. But what we want to have tomorrow with big data that's possible is a model of you specifically, a model of her. And then we know what good medicine is for her. We know what good medicine is for me. We know a good intervention for different people. Having said that there are a lot of issues on how the data gets used. And then I think there is for example a tension between the organizations that have a lot of data and the ones that don’t.
Knowledge is power and data is knowledge. So data is power. I think if you look at how things are evolving you'd think to get a lot of big data concentrate it – The question is always who is going to do what with this data? And some of us maybe have less data than we should have. And that's maybe part of what we can try to do. On the subject of robots for surgery this is a very exciting frontier. For example the Da Vinci robot is an extremely accurate, very effective robot for surgery.
For the most part the robots are still controlled by human surgeons. But I think the time is not far where we will actually have robots that do the surgery themselves. And they have many advantages because for example they're not limited by the visual system that we have. They can have a continuous integrated sensing of what's happening at the exact point of surgery. So I think this is another area where we're going to see a lot of progress.
DiChristina: That's really interesting. And they don't get tired.
Domingos: Yeah exactly, they don't get tired.
DiChristina: I think we have time for one question more maybe? There's one in the back.
Question: So just considering the problem with really, really high prescription drug prices currently in America I guess I have a two-part question. The first is how significant of a factor in these high prices is the money spent on R&D? And secondly how much more investment both in terms of time and money is required to get AI to a point where it can at least do R&D on these drugs as well as humans?
Saria: I think there are actually very legitimate efforts that are starting to happen in doing R&D both - like Pedro mentioned earlier – inventing new drugs, finding new indications for existing drugs. And this is using very detailed data off existing patients, seeing how people have reacted, all the way to new data that's being collected in the lab in dishes to make measurements about what possible new drugs or candidates are.
So I think that's definitely getting accelerated and a place where we will see a lot of activity I think in the next ten years.
Hepp: What's commonly understood is that R&D is a big factor for behind drug prices but I don't have a corner.
Domingos: Let me give one element to the answer to that. In the drug development pipeline where is the huge cost? The huge cost is actually not in coming up with new drugs. It's when you have drugs that look promising enough to be used and they're going to human trials. And human trials are extraordinarily expensive. They cost billions of dollars. They take many years. And then the drugs often fail.
And this is really the single thing that most contributes to the cost of drugs. Now fortunately this is actually a place where machine learning can really help and is starting to help which is what you do is when a new candidate drug is produced you try to predict whether this drug is going to have harmful side effects or not. And then you can weed out a lot of candidates that way.
In a way it may be ironic where the biggest role in machine learning is not necessarily in coming up with the new drugs but in weeding out the ones that are going to fail at a huge cost before that cost is incurred. So I think there are good signs of hope on the horizon. And this is already starting to happen.
DiChristina: I wanted to ask you all to join me in thanking the speakers for a really fascinating discussion. [Applause] And thank you for coming. Thank you.
Mirsky: That's it for this episode. Get your science news at our Web site: www.scientificamerican.com. Or you can also find the Sustainable Energy “Science on the Hill” session in our podcast archive. And follow us on Twitter where you'll get a tweet whenever a new item hits the Web site. Our Twitter name is @sciam. For Scientific American Science Talk, I'm Steve Mirsky. Thanks for clicking on us.