The Current Applications of and Opportunities for AI in Healthcare

Simon D. Murray, MD: Medicine is a perfect, fertile field for AI. There's so many tasks and so many things that it can do. I didn't realize I was using AI until I started thinking about it. I use Dragon, and I can use my cell phone now to dictate into Dragon, I mean it's an amazing thing the newest versions of the Dragon medical 99% pick it up right out of the box. That’s Kurzweil’s stuff right?

Eric Daimler, PhD, MS: As I recall there’s a different researcher that actually developed on that, though Kurzweil is obviously a brilliant man. The discussion around that technology, about voice recognition technology, also is a great way to talk about the difference with AI, what's changed. How is AI different than the last three four or five generations of digital technologies? Voice is one of those, so we've been working on voice recognition for as long as computers have been around but it's not really helpful when it's at 70% efficacy, 80% efficacy. Even 90% efficacy can be a little troublesome. We might think 90% accuracy in lot of things is good, in voice recognition not so much. But when it works it scales infinitely, it's suddenly available for everything and at 99% accuracy suddenly we just realize “wow we can use voice.”

That's why it's it appears as though those 24 by 7 microphones, what we call speakers in our home, like Alexa and Siri and so forth, those microphone speakers, that's why they occur as having suddenly appeared. It's because the technology suddenly reached a threshold, it's useful, and boom everybody can use it in seemingly infinite number of applications.

SM: Voice recognition really improves life, I know in radiology they get their work done much quicker. That's one application of AI in medicine. What are some of the other applications that that you can name off the top of your head?

ED: I really put them in four buckets: drug discovery, I have patient records that's in the nomenclature all over this, I have logistics, you know just the moving of supplies around a hospital, and then you know we have medical robotics about which I've written.

And so in those four I'd say that the one that's exciting for me is drug discovery. We have a kind of increasing sensitivity we might say about human trials on drugs. Yet we have this treasure trove of past drug trials. It can't be tapped right now or it's very difficult to tap it because the data is dirty, it's clustered in a particular way where we don't often know what data was collected and then it was collected in different ways. So we might look at an old test where they ask the patient did you smoke yes no, another one might say how much did you smoke, and another might say how long ago did you smoke. They're all trying to get at the same thing, three different studies, over three different times maybe over three different sponsors and three different companies.

How do you merge them, if those were three different drugs, into something as a collective? You can't normalize them or average them, you want to keep that fidelity, so the innovation that's coming in data integration, this is very boring low-level IT technology, is going to allow for that to come together for analysis. This is hugely exciting.

SM: What other applications do you see for robots and for AI?

ED: You know logistics is an obvious one. So I was in Pittsburgh where we saw this r2d2 looking thing shuttling supplies around a hallway. This is a great way for anybody concerned about jobs in AI to think about the jobs that are going to get taken away. Jobs will get taken away. Here's one shuttling: supplies around a hospital. But you know, no job gets automated that humans want to do right? The way we've thought about this traditionally is dirty dangerous jobs those are the ones that get automated but even the ones that are just boring we find benefit automating these sort of jobs. Nobody trains themselves to shuttle supplies around a hospital, it's not a human task in some sense, it doesn't use our skills. So logistics, the supply chain of hospitals also. It's dreadfully inefficient. I was talking to a large Hospital in Manhattan who doesn't really even pay attention to the inefficiencies because they get measured on standard of care not the efficiency of their business, but I think the increase of some of these tools will help bring efficiency and lower the cost of delivery.

SM: There's a lot of ways, they don't even keep track. A nurse opens this sterile gown it gets wasted, throw it away, I'm not even sure it gets tracked. I think they just built into the cost so high to account for it. Which is not very efficient and we can't keep doing that, we have to track these things better, track supplies track who’s using what.

ED: I can say another example is patient records. So AI is going to make a difference in patient records. It has to some degree. I mean I look at the fundamental IT infrastructure of it, the guts of the technology, and the guts of the technology say that we need to integrate heterogeneous data, data that's not alike, and heterogeneous data doesn't like to be integrated.

This is the reason that healthcare.gov had such problems, it really was brought to its knees for a time, it's because heterogenous data doesn't like to be merged data. The British National Health Service is trying to do a similar sort of data integration of patient records, Italy has a 30 billion dollar effort trying to integrate health records, and I hate to say that they're likely to have unhappy outcomes because the technology is really just emerging for a new way of integrating heterogeneous data.

SM: They made the same mistakes we did by allowing 90 vendors to come in and do 90 different EMRs?

ED: They may be making different mistakes but similarly at the IT infrastructure level the same mistakes. I virtually guarantee will be made.

SM: That was a tragic mistake actually. In my office now I can get an EKG or an x-ray report sent to me through the EMR it's perfect it comes right in, but I have to print it and fax it to somebody else. I don't have the ability to send it directly from my computer to your computer. And faxing is the least secure of all. It's so inefficient, we're still printing things out and faxing them, I mean it shouldn't happen.

ED: Well this touches on other parts of our discussion where we really need to look at incentives. You know what are we incentivized to do and if we're incentivized just to continue meeting a practice of standard of care then faxes versus email really have nothing to do with it and we'll just keep doing what we're doing.

You know another way to think about it is how we will introduce new information thinking that's going to solve some sort of problem. It often doesn't it. And we experience this in our day to day lives, when we will in the United States go to find a mortgage application and the resulting mortgage paperwork has become thicker and thicker, no one's incentivized to make that smaller but that deluge of information on a home mortgage is ultimately giving you all the information you need but it doesn't help in your understanding.

SM: No it certainly doesn't, and when a person goes home from the hospital with a stack of papers that thick and there's instructions it's overwhelming actually to them. Every prescription has a four-page blurb with it and no one's reading them anyway.

ED: This is actually the perfect place, that physicians, clinicians, researchers, administrators, everybody in the healthcare industry already experiences many of the issues that that we are confronting in AI because you already in the healthcare profession you already are interacting with patients on issues of explain-ability and understanding even when somebody is sitting right across from you they're alert they’re nodding their head they still may not understand what you're talking about and you have a very difficult time, but you're completely used to the issue of your context being different than theirs and the level of understanding being different and this is very similar to what we confront with learning algorithms or any number of other parts of the totality of the AI system.

Transcript edited for clarity.