AI Leadership Insights: The Impact of AI

Mike Vizard: Hello and welcome to the latest edition of the TechStrong.ai video series. I’m your host, Mike Vizard. Today we’re with Paul Brient, who’s executive vice president and chief product officer for Athenahealth, and we’re going to be talking about AI and how it impacts everything we know about health care. Paul, welcome to the show.

Paul Brient: Thanks, Michael. It’s great to be here.

Mike Vizard: There’s a lot of hype out there and everybody has seen everything, and I might argue our imaginations are running away with us a little bit, but let’s just start with some of the simple stuff. What is possible today with AI that we couldn’t do yesterday, and level set some expectations?

Paul Brient: That’s a great question, and there’s been a ton of popular press around AI related to generative AI and the ChatGPT and all that good stuff that’s been coming out this year. The reality is we’ve been doing AI in health care for quite some time, and there are a bunch of things that AI is doing for us and frankly in ways we probably don’t even necessarily know that AI is doing it for us.
So here at Athena we’ve got all kinds of models that are running in the background to help pull work away from physician practices and physicians, everything as simple as taking an insurance card that you might scan and putting all the information into the computer so you don’t have to and picking an insurance package to taking a fax document that comes in and reading the fax and understanding what it is and filing it and putting it in the right place. There are a whole bunch of great uses for AI and natural language processing that just make work that otherwise would be pretty mundane that humans would have to do that they don’t have to do.

Mike Vizard: Yeah, it seems like we’re making everything a little more accessible, but the interesting thing is there’s multiple forms of AI, there are different models of different classes and we’re going to live in this kind of multimodal AI world. Are we prepared for that? I mean, how complex an endeavor is that and how big a specialist in AI do I need to be to succeed?

Paul Brient: Yeah, it is interesting. Certainly I generally view it as AI is becoming very democratized. If you go back 20 years, if you wanted to do some AI, you had to sit down with a bunch of computer code and write a neural network and do all this stuff. And now as a programmer, if you want to build a model, you pretty much go on AWS or Azure or whatever your SaaS offering of choice is and you have all these tools available to do this. And as consumers now, we can log in to ChatGPT and ask it a question and it gives us an answer. Sometimes it’s right, something, it’s wrong, but it’s an answer and it’s pretty cool. I’m in the Google beta, so in my email now, it offers to write responses to emails for me and it sometimes does a good job and it’s helpful, so it’s becoming a lot easier for the layperson to use. You don’t have to be a data scientist as much to start really engaging with some of these cool AI tools.

Mike Vizard: Do you think over time we might start discovering the root cause of things? There are all kinds of issues in the world out there that we have not quite sorted. There are, for example, clusters where there are cancer cases that we don’t know the reasons for. Is all of that stuff that we’re looking for in the data somewhere?

Paul Brient: It probably is in the data. The interesting thing about AI is that AI is really good at replicating things that we’ve already discovered. So if you take a cancer analogy, reading images to see if people have cancer, AI is pretty good at that, and you could argue it’s more reliable perhaps or more repeatable than a human could be. But it learned from humans. So basically we fed AI, we said, “Hey, here are a whole bunch of images, here are the ones that have cancer and the ones that don’t, AI you go learn that,” and it does its thing and then it says, “Okay, good. Now I can repeat that.” But what it can’t do is say, “Hey, here’s a bunch of cancer that you weren’t previously thinking about and a pattern that you didn’t previously have,” without that initial set of data. So it isn’t a tool, at least today, that can discover new relationships, but it is a tool that can take existing relationships or outcomes that we haven’t previously figured out and correlate them.
It’s also particularly bad at telling you why. It can tell you what, but not why. It’s like, “Yes, I found all these images and these ones that have cancer and these don’t,” and a human looks at them and goes, “Yep, you’re right,” but it’s not like a human, if you talk to a radiologist and I say, “Hey, how did you find the cancer in this image?” They can tell you why, because they have a thought process for it. AI can’t really tell you why, which makes it sometimes frustrating to work with, but it’s just part of the deal.

Mike Vizard: It somehow sounds a lot like having an enthusiastic intern.

Paul Brient: There you go.

Mike Vizard: You are the product guy so what kinds of things is Athenahealth going to be able to develop or has developed that are driven by AI that are kind of different and things that people might want to be looking forward to?

Paul Brient: Well, we’re an electronic health record and physicians use electronic health records as part of the care for their patients. And sometimes when you’re using computers, it takes time and it’s frustrating and it takes us out of our workflow, and I think EHRs historically have had those criticisms. We work really hard to make ours not do that, and AI is going to give us a bunch of tools to help continue to take work away from the provider. So the generative AI in particular has got a lot of promise because it writes things and writing things takes time. I always tell everyone that I’m a pretty good editor, I’m not a great author, so send me a draft and I’ll edit it. AI can do the same thing. So imagine a physician getting a text from a patient asking a question, you can imagine generative AI providing a draft response to that patient, considering the patient’s chart because the AI digested it. It’s a thoughtful response perhaps, maybe it got some things wrong, the doctor wants to emphasize something, they make a few edits, send it off.
You can imagine AI listening to a conversation between a patient and a provider and summarizing that conversation and perhaps trying to write a note, perhaps extracting the orders. So the doctor said, “Hey, you need to take aspirin twice a day,” there it is, the order in your notes. So there’s a bunch of stuff that it can do to be kind of a really, really good assistant, or in the medical term we call them scribes. So some doctors actually hire aspiring med students to be at their side and do some of the computer work and they call it a scribe. It’s kind of an expensive thing, but it’s kind of nice for the person because they get to learn a little bit about what it’s like to be a doctor. And you can imagine AI kind of becoming that virtual scribe, that virtual companion that helps the physician understands what’s going on and provides drafts and tees things up so the doctor isn’t having to do stuff from kind of first principles.

Mike Vizard: There is, of course, always a lot of concern about privacy when it comes to anything to do with healthcare. So as we kind of take all this data and put it into some sort of AI repository, how do we make sure that the data doesn’t wind up in the wrong place?

Paul Brient: That’s a very, very important concern, and certainly you do not want to put any of your health care information in any of the public ChatGPTs or things of that sort. I would highly recommend not doing that. However, the private versions of those, so at Athena, we have our own version that conforms to the HIPAA requirements, which is the privacy protection requirements. So we can put a PHI, protected health information, into that ChatGPT-like thing, and it stays within our four walls and we protect it in all the ways it needs to get protected. So there’s nothing inherently wrong with AI in terms of protection, it’s just that the public ones are, well, they’re public and they learn from the prompts that you put into them. So you don’t want to put any confidential information of any kind into any of the public stuff because it will not be private anymore.

Mike Vizard: If I look at the health care system today, it is generally run by well-intentioned people trying to navigate a massive number of data silos, and lo and behold, bottlenecks emerge. Do you think AI will help us identify those bottlenecks and streamline a lot of the patient experience because so many folks are dealing with multiple hospitals and health care systems that it just becomes a major challenge and some of them are even hiring human assistants to navigate it, and maybe that could be all become less friction filled as it is?

Paul Brient: Absolutely. I think anyone who’s tried to use the US health care system has probably experienced some frustration along probably two dimensions. One is data is spread out all over the place and it’s really hard to get a coherent view of your data. And I’m not sure that AI in particular is going to be a major unlock for that. However, there’s a bunch of work going on, there’s a piece of legislation called 1st Century Cures, which essentially says, “Hey, if you have health care data, you have to make it available in some meaningful readable electronic way using some standards.” And that’s really unlocking the ability now to pull a patient’s record together. So whoever you are, if you’re a doctor and you’re using modern tech, that modern tech can go out and pull all the data together. Our term for it is experiential interoperability so we create basically a virtual chart of all the charts that we can go find on your behalf, not necessarily using AI.
But where AI comes in, which is really cool, is okay, we do that, now there’s a whole bunch of information for the poor doctor to have to review. As a patient, you’d sort of like them to review that, but they don’t have a lot of time to do that. The visits, doctor visits are quick and you’d like the doctor to be talking to you, not sitting in front of a computer. So what AI is really good at doing is saying, “Hey, summarize this information for me. I’m a cardiologist.” Once we’ve gone and gotten all that information, now have AI go, “Hey, what’s relevant for me as a cardiologist about this particular patient?” So the AI kind of comes up and says, “Hey, here’s the three or four things you, doctor, need to know so I can now have a conversation with you as a patient and be informed so I don’t have to be speedwalking through the whole chart.”
So it’s kind this neat one-two punch, the standards and the tech and just the interconnectivity of the health care system is getting better so we can get the data, that presents a problem in itself and then AI helps solve that problem by allowing the physician to more easily access all that data.

Mike Vizard: Do you think that will ultimately help level the quality of care that everybody experiences? Because a lot of folks today would go to say, I don’t know, I’ll pick MassGen in Boston, they have a lot of resources, but the folks who are living out in a remote spot in Oklahoma, maybe not so much. So do you think as AI kind of progresses, we might be able to even out the quality of the health care that everybody experiences regardless of where they’re physically located or what organization they’re signed up with?

Paul Brient: Yeah, that is certainly a tough one. Just like in any field or any profession, there are better and worse members of that profession, and health care is no different. I do think that the technology in general has the opportunity to do a couple of things. One is it can help physicians make sure they’re aware of and don’t forget things. So there’s a protocol for a disease, the software can remind the physician of, “Hey, here’s the protocol,” or, “Hey, read this article because it summarizes nicely from an expert this particular condition.” So if you have someone in a rural area that doesn’t see weird diseases or unusual circumstances, they find a patient with that, the software in the technology can really help that physician become more of an expert on that.
The other thing you can do is provide access. When you go to a Mass General or a major research institution, you get access to therapies that aren’t yet approved in clinical trials and things of that sort. And one of the things that is particularly cool is if you’ve got technology now that can identify a patient and the trials and say, “Hey, there’s a match.” You’ve got a patient, you’re in some remote town or not in a major academic institution, you’re not doing a lot of research and you get a patient that’s got a condition and there’s a clinical trial going on, you’re kind of out of therapy, you can say, “Hey, is this patient eligible for any trials?” The computer is can do the matches and say, “Yeah, hey, there’s a trial going on at NYU,” or wherever it is, and you might want to enroll your patient in that. So that provides more equity around clinical trials and cutting-edge treatments as well.