Subscribe to The Podcast by KevinMD. Catch up on old episodes!

Join Tim O’Connell, a physician executive. In this episode, we explore the fascinating world of medical-grade AI and its transformative impact on health care. Tim provides valuable insights into the evolution of AI in medicine, recent advancements, and practical applications that are reshaping the future of patient care.

Tim O’Connell is a physician executive.

He discusses the KevinMD article, “The clinical AI revolution: 3 things to know.”

Our presenting sponsor is Nuance, a Microsoft company.

Do you spend more time on administrative tasks like clinical documentation than you do with patients? You’re not alone. Clinicians report spending up to two hours on administrative tasks for each hour of care provided. Nuance, a Microsoft company, is committed to helping clinicians restore the balance with Dragon Ambient eXperience – or DAX for short. DAX is an AI-powered, voice-enabled solution that helps physicians cut documentation time in half. DAX Copilot combines proven conversational and ambient AI with the most advanced generative AI in a mobile application that integrates directly with your existing workflows. DAX Copilot can be easily enabled within the workflow of the Dragon Medical application to bring the power of ambient technology to more clinicians faster while leveraging the proven and powerful capabilities used by over 550,000 physicians.

Explore DAX Copilot today. Visit https://nuance.com/daxinaction to see a 12-minute DAX Copilot demo. Discover clinical documentation that writes itself and reclaim your work-life balance.

VISIT SPONSOR → https://nuance.com/daxinaction

SUBSCRIBE TO THE PODCAST → https://kevinmd.com/podcast

RECOMMENDED BY KEVINMD → https://kevinmd.com/recommended

GET CME FOR THIS EPISODE → https://earnc.me/5grFjW

Powered by CMEfy.

Transcript

Kevin Pho: Hi, and welcome to the show. Subscribe at KevinMD.com/podcast, get CME for this episode by clicking on the CME link in the show notes. Today we welcome Tim O’Connell. He’s a physician executive, and today’s KevinMD article is titled “The clinical AI revolution: 3 things to know.” Tim, welcome to the show.

Tim O’Connell: Thank you, Kevin, it’s great to be here.

Kevin Pho: So let’s start by briefly sharing your story and journey.

Tim O’Connell: Sounds great. So I’m a practicing radiologist, I work one to two days a week clinically, the rest of the time I spend working at my medical AI company. And I used to work as a network engineer before going into medicine, and so I had a pretty strong technical background.

And then during residency, because we’ve got an all digital workflow in radiology, I got involved in a lot of informatics projects, did a lot of coding, a lot of workflow improvements, ended up doing an informatics fellowship, got interested in natural language processing. And I realized we could use it to solve a lot of our problems for information management in health care.

And I built an app that just did patient summarization. Right, as a radiologist all I get is a minimum of clinical information on the requisition, and I wanted to know more out of the patient chart. And so the theory was great, the NLP software I was using was open source software, wasn’t great. So I met a wonderful guy who’s now our CTO, and I decided to collaborate with a few other folks, and now we’ve got a company and we’re very busy.

Kevin Pho: All right. So on the date that we’re recording this, yesterday was ChatGPT’s first birthday. From that AI medicine intersection, how has the last year been for you?

Tim O’Connell: It’s been absolutely fantastic, it’s been wonderful. ChatGPT has been in the news so much, it’s so accessible, it’s so easy to use, you just log in and ask it a question about pretty much anything and look at it go.

So it’s created a big tailwind for us, because I think a lot of people have kicked the tires with ChatGPT, and it’s been wonderful in that they’re like, OK, wow, we can do things with AI and language models now. But on the other hand they’re like, oh, there’s accuracy problems and hallucinations and patient quality and safety issues and all sorts of things that we need to sort out first. So then people I think have looked around and been like, oh wow, there’s these people who’ve been doing this for years, maybe we should talk to them too.

Kevin Pho: All right, so let’s talk about your KevinMD article, “The clinical AI revolution: 3 things to know.” So tell us about that article.

Tim O’Connell: Sure. Well, I mean, I think it was really just covering some of the challenges and opportunities in health care for medical AI. I’m happy to sort of answer any specific questions, but I think it was a great opportunity to provide a state of the field discussion or summary about what some of the opportunities and challenges are.

Kevin Pho: So when you say medical AI, what does that mean to you?

Tim O’Connell: Yeah, well, I think it’s a pretty broad term, Kevin. What a lot of people are talking about these days with AI is large language models, right? But certainly as a radiologist I’m keenly aware there’s a lot of AI going on in computer vision, image recognition, right, to augment or assist the radiologist, that kind of stuff going on.

And there’s all kinds of other AI that can happen, right? Like you could have thousands of models working in a hospital to identify patients who are at risk of things, or process failures, process risks, and all this kind of stuff. So medical AI is a pretty broad umbrella, but what we’re talking about in that article is the use of AI models to deal with medical text.

Kevin Pho: So like you said, it sounds like AI is in the mainstream vernacular for the last year since ChatGPT, but there’s been natural language models even before ChatGPT, right?

Tim O’Connell: Long before, yeah. And natural language processing has been a little bit like fusion energy, it’s been 10 years away for the last 50 years, kind of thing. And the reason why is because it’s an 80/20 problem. It’s really easy to get to 80 percent accuracy, and then really difficult to close the final gap to get to sort of normal language accuracy. It’s never 100 percent, right? Miles Davis said, if you understood everything I said, you’d be me, right? So it’s really hard to get 100 percent accuracy in language understanding.

But we’ve been working really, really hard for the last 50 years in NLP on closing that gap. And there’s been a timeline of advances, right? There was sort of rules based methods in the 80s and 90s, statistical machine learning in the early 2000s, moved on to deep learning in the late 2000s, and then really the advent of transformer models in, I think it was 2016, and now until, like, massive language models. So there’s been this step-by-step technology evolution, and we’re now at the point where very interactive, very capable, very accurate systems are possible.

Kevin Pho: So let’s talk specifically about your expertise, which is the application of AI when it comes to text. Right, as you and I know, sometimes these patient charts are volumes of text and a lot of it simply isn’t relevant to the situation at hand. It’d be wonderful to have some type of model to ask it questions and discern what’s important and what’s not. So where are we now with that technology?

Tim O’Connell: I think we’ve moved a great way along. I think there’s a little bit of confusion going on in the marketplace right now. So you’ve mentioned the problem, and there’s a great quote from Dr. Ashish Jha from a Scientific American blog he wrote a while ago. He said, I see a patient who’s had hundreds of encounters, and so I skim the chart and I hope I don’t miss anything. And we’re kind of all in the same situation. You’re faced with a complex patient and so you try and find the last discharge summary or anesthesia consult, right, those are two high yield documents.

But we need to do better than that. And you can take AI, large language models like generative AI models, and have them do summarization of documents, but that doesn’t necessarily mean it’s not going to omit something that’s really important. And it’s also maybe taking text and putting it in prose form and then providing you more prose form, which may not be the right method of displaying it to you.

My personal preferred method for seeing patient historical information is lists of diagnoses, right? Like, here’s things related to the heart, lungs, mental health, prior procedures, things like that. I think doctors are very good at ingesting kind of list data, because we’re so used to looking at labs and things like that, and that works better than prose data. So there’s issues here with accuracy and inclusiveness and what’s important. What’s important for the radiologist is different than what’s important for the family doctor, right? And so there’s still a lot of work that we have to do in this.

Kevin Pho: So tell me your ideal scenario, where AI, generative AI, transformer models, or what have you, is working to the maximum of its capability. Give us a scenario, tell us this case study or story where the AI is working as it should.

Tim O’Connell: For sure. So I mean, I work mostly in hospital-based care, and I really think that what we need is, there’s a lot of things that humans are really bad at, and broadly they kind of fall into, at least for my role, two capabilities.

One is doing detection tasks, right? So if you have a screen full of white dots, figuring out which white dot is the lung nodule you care about is difficult. But there’s all kinds of ways that humans do those detection tasks, right, not just for image-based things. It could be waveform analysis and ECGs, it could be simply reading chart data. So what we need is, we need AI models to help humans with those detection tasks, in ways that are going to be efficient and accurate and safe.

And then the other task that humans are really bad at is looking at a huge amount of diverse data all at once, right? So integrating something that was mentioned in a chart five years ago with the patient’s current presentation, or integrating a genomic test with the patient’s current presentation. And that’s something that computers in general, big data, but certainly AI models, can also help us out with.

So when I think about how AI could be helping us live our best life as clinicians and helping our patients to the best way possible, it would be sort of in one of those two domains. To prevent things from slipping through the cracks with humans doing detection tasks, and also for integrating massive amounts of data to provide useful insights, for either preventing disease, preventing disease progression, catching problems before they happen in health care. I think that is one of the most important ways.

Kevin Pho: So you outlined your ideal scenario. Tell us, as of today, what are the obstacles that are preventing us from getting there?

Tim O’Connell: Oh my gosh, there’s many, there’s so many. I mean, I think we have all the normal health care system obstacles. For one thing, number one, who’s going to pay for it? That’s always an issue in health care, particularly in Canada. And number two, regulatory stuff, like whatever you do, it can’t hurt patients, it has to be safe. Number three, we just have adoption speed, right? Like every different institution is going to be somewhere different on the adoption curve. Number four, we have major workflow design challenges. I was literally just in a hospital meeting about workflow stuff. So integrating things into our workflow in a really efficient and safe way is also going to be important.

Kevin Pho: So I have some doctors who said, I don’t want to wait, why can’t I just simply cut and paste discharge summaries into ChatGPT and simply ask it questions? And now with ChatGPT you can create custom GPTs and have it specialized. So why can’t doctors just do that, rather than kind of waiting for all these regulatory hurdles to fall?

Tim O’Connell: Sure. Well, first of all, you may be violating data sovereignty if you do that. You’d be sending Canadian health care data to the United States, which there are different provincial laws regarding that. Number two, you may be violating provincial privacy laws, would be another thing that you’d be doing. Number three, you may be violating patient consent at your organization. And number four, you would be using an unregulated device as a medical device, so Health Canada would probably have something to say about that. So that’s the first four things that come to mind. So there’s all kinds of reasons why we can’t do that just right now.

Kevin Pho: I know that you’re in Canada. Are there any hospitals that you know of that are currently using any form of AI, and if they are, what’s the current state right now?

Tim O’Connell: Yeah, absolutely there are. There’s hospitals all over the world who are using AI right now. There’s certainly a lot of image analysis AI being used.

One thing I’ve seen at one of our local hospitals is, there’s a thing in stroke triage called the ASPECTS score, and so it will take a head CT scan and automatically calculate an ASPECTS score for it. It’s not always right, but sometimes it saves you time. And so that’s certainly a micro case of where AI is being used. But there’s lots of things like that being used in radiology departments all over, lung nodule detection, things like that. And we’ve had these CAD tools for years, like literally 20 years, but they’ve just gotten better with new deep learning models, really, so it’s not too transformative there.

There’s certainly other hospitals, for example, where my company’s software is in use, for being able to process documents, to do cohort identification, to do follow-up tracking. Really useful use cases like that that save people time and hopefully increase patient safety or assist in research.

Kevin Pho: So the technology that ChatGPT popularizes, the large language models, which is essentially just pattern recognition, is that the technology that we have to look forward to, that future AI advances is going to be built off of?

Tim O’Connell: Hopefully much more than that, Kevin. Yeah, so those large language models, it’s not quite pattern recognition, they’re actually more like sentence completion models that have layers and layers built on top of them. And so that is just one tiny area where we could be using AI. It just happens to be something that’s really interactive and neat, and for the first time allows humans to interact with computers naturally, and that’s why it’s providing a lot of the press.

But that whole class of models is based on something called the transformer architecture, which came out in 2016 I think, it was out of Google Labs. And that transformer model is sort of like a new type of neural network, and it’s being used for all kinds of different things now. But five years from now someone may have a newer or better type of foundational model, which continues to further improve on performance and enable new applications. So we’re just at the beginning here.

Kevin Pho: We’re talking to Tim O’Connell. He’s a physician executive. Today’s KevinMD article is titled “The clinical AI revolution: 3 things to know.” So I want to ask you a couple of forward thinking questions, but the first one is specifically with radiology. And you mentioned that already AI tools are being used to help radiologists. So I always like to ask radiologists, are you concerned about AI replacing radiologists in the future?

Tim O’Connell: Oh God no. I want to be the Maytag repairman, right? I want to sit around with nothing to do, or just be seeing patients or providing consults or doing procedures on patients.

Like I said, there’s a lot of work that we do that is literally like humans doing a detection task, and human beings are never going to be amazingly good at that, like not as good as potentially an incredible AI model could be. But there’s a real integrative part to our job that we have to know a tremendous amount to be able to do, and is based not just on textbook knowledge but on years and years of experience.

So it’s one thing for an AI model to find a lung nodule, it’s another thing for a human to go, yeah, I’ve looked at a prior chest x-ray from 10 years ago, which is the only prior in this case, and that lung nodule is stable, and so we don’t need to do anything about it, right? So there’s a lot of higher order functions that our AI models are not going to be capable of doing, at least not in the near term, not in five years. Who knows in 10 years.

But there’s a lot of things where it’s going to go, oh, this looks like some terrible disease, and we’re going to go, no, this is just a patient with a congenital deformity or a congenital thing, and it’s nothing, right? So we are a long ways away from any kind of general purpose image recognition thing. We’re still going to be looking at pictures for a really long time.

And then there’s all the procedural work we do, which is just increasing in volume with image-based procedures. A busy day, if I’m covering procedures, it’ll be renal biopsies and putting in chest tubes, abdominal drains, thyroid biopsies, things like that. And you still need a pair of hands to do that stuff.

Kevin Pho: So that example that you gave about the lung nodule, so if you fed an AI model all the patient x-rays, that model wouldn’t be able to discern whether a lung nodule has changed over the years?

Tim O’Connell: It certainly may, but you’d have to have a model specifically built to do that case. And comparing against modalities is also very different. So if I’m looking at a lung nodule in a CT scan, my only prior might be this 10-year-old chest x-ray, and doing that cross-correlation between different modalities is exceptionally difficult. I don’t think anyone’s really even trying to tackle it so much right now in AI.

It’s one thing to do registration across one CT scan to a prior CT scan, or one x-ray to another x-ray, that has its own challenges. But then doing cross modality correlation, like ultrasound and, like, no, it’s super difficult. For whatever reason the human brain seems really good at that kind of stuff, and AI models are particularly lousy right now.

Kevin Pho: So so much has happened in the past year, but looking into your crystal ball in the immediate future, say the next 6 months or so, what do we as physicians have to look forward to in regards to AI?

Tim O’Connell: What we’ve got is, for years I’ve been working, anyone I work with says, well, what do you do in informatics? And a really great mentor of mine once told me the right way to answer that question: it’s improving quality, safety, and efficiency of care, right? And so we’re not installing Windows for people, that’s information systems, information technology. So informatics in health care, I think AI is just going to be used by informaticians to help improve the quality, safety, and efficiency of care.

And it’s going to have all sorts of wonderful knock-on effects. It’s going to have the potential to decrease things like physician burnout, because you’re spending less time scanning the chart, you’re spending less time creating documentation or sending emails to patients and this sort of stuff. It’s going to help catch diseases at earlier stages through doing inference and predictive capabilities. It’s going to reduce repetitive tasks.

And at the end of the day, I think we’re having a bit of a sustainability crisis in health care, right? In the 1960s we didn’t have MRI machines and drugs that cost half a million dollars a year. And so the question is, how are we going to pay for everything? And I think AI is one of the few new technologies which is coming out which has potential for efficiency improvement, which will hopefully yield sustainability, so that expensive imaging tests and other diagnostics and expensive treatments can actually be afforded by the health care system and patients, and help those patients in the ways that they need to.

Kevin Pho: And my final question, tell us some of your take-home messages that you would like to leave with the KevinMD audience.

Tim O’Connell: Sure. I think one of the most important take-homes is, AI is coming. It may seem really slow where you are today. Don’t be impatient, don’t rush it, do not take on any personal risk. Please do not cut and paste into ChatGPT patient documents. Your organizations are working at every level. I was just at a chief medical officer conference in the US and it was almost turning into an AI conference, right? So the AI revolution is certainly coming to health care.

Be patient and participate, right? The one thing that we need, there’s a lot of very well-intentioned vendors out there who have very few people on their staff with clinical experience, and they need people to help them understand what medical workflows are and what medical safety is, and why they might be making a product that’s dangerous. So participate any way you can, because at the end of the day, if you’re helping improve sustainability of the health care system, if you’re helping improve quality, efficiency, safety of care, you’re helping patients.

Kevin Pho: Tim, thank you so much for sharing your perspective and insight, and thanks again for coming on the show.

Tim O’Connell: It was my pleasure, Kevin. Thank you very much.