Show notes
Scot Morris, OD, and Rehan Ahmed, MD, on GPT Health in Epic, local vs. cloud AI, and the AI Bill of Rights.
- Industry news: OpenAI’s integration of “GPT Health” into the Epic EHR, letting clinicians summarize records across a network of 325 million patients.
- Local vs. cloud AI: Local (edge) AI keeps data inside the clinic but needs heavy upfront hardware. Cloud AI offers effectively unlimited compute but leaves you stranded if the internet goes down.
- The AI Bill of Rights: Five pillars meant to protect patient privacy and equity as AI enters healthcare: safe and effective systems, protection from algorithmic bias, data privacy, explainable results, and human clinicians in final control.
Transcript
Welcome to Real Talk on AI in Eye Care with your host, Dr. Scot Morris. And I'm co-host, Dr. Rehan Ahmed. In each episode, we discuss current news, add in a little AI education, and debate innovative AI tools and topics that will change the industry. I also want to talk a little bit about news of the day. And so I think one of the most interesting news of this week, and this is being recorded on 9-226, is that recently we just had where OpenAI is wiring chat GPT into Epic EHR. Now, I know for many of us in the eye care industry, we don't necessarily use Epic. If you're in a hospital-based situation, maybe you are. And I know that doesn't sound like a really big deal because most of our EHRs are not wiring any LLMs into the model.
But this is a big deal for medicine as a whole because, you know, they're basically putting OpenAI is going to launch this chat GPT health with a read-only integration that goes directly into Epic EHR. Now, the reason that's such a big deal is Epic serves about 325 million patients. That's a lot. And let's condition summarize notes and labs and medications and all that stuff without actually writing back. And so this is the first really big integration we're seeing about AI into an EHR. Now, I mean, I know we've talked about this many times, is an EHR has a lot of challenges in it in terms of how it's interconnected and how it operates with other technologies.
But I do think this is an interesting first step. So we'll see where this goes over the next couple months, so the next couple, you know, year, as we see what happens with GPT health and its competitors as they start looking how did they get involved in medicine. For today's knowledge bite, we are going to cover a little bit about what's the difference between local AI and network AI in Eye Care. So when you think like this, when choosing AI tools for an optometric or optimal optimal logic practice, whether it be for something like automated funderscreening or ambient charge scribing or even optical inventory forecasting, there's still one foundational architecture decision that kind of dictates everything that we do, which is where does the compute actually happen.
Now, clinicians often assume AI is AI, but the reality is that's not really true. The mechanism and the answer to that question of where compute really happens has some very distinct trade-offs. So let me first of all explain what's local AI or often called edge AI. And this is a model that runs directly on your on-premise hardware, like on an OCT or fundus camera or a dedicated clinic server or maybe in the very near future, your tablet or your phone. Data is ingested, processed and reported without any patient raw data ever leaving, let's call it the four walls of your office versus what we're more used to, which is called network AI or cloud AI.
And that's where data is encrypted and then sent over the internet to some centralized remote data centers like AWS or Microsoft Azure, Google Cloud, and then it's analyzed by a massive foundational models and returned back to your software. Maybe it's an EHR. Now, friction kind of emerges when practices deploy cloud-reliant tools without adequate internet redundancy. In other words, so it's great if you're cloud-based, but what if your internet's slow or your internet's down? Then your network AI is absolutely useless. On the other hand, what happened, the challenge with local edge models or local AI is, well, how do we update those models? And how do we cross-reference those with other models or other examples of that patient chart or that patient exam that's not on that local AI server?
So there's goods and bads of each one, and the reality is that there's one is what I'd call perfect, but nonetheless, I think the future is we're going to combine the two of them. So in local AI, we talked about really all the data resides locally. It's not really internet dependent in any way, shape, or form because you can save it on your device and then upload it whenever internet is existing. It is constrained to some degree by how much horsepower, your local GPU or NPU specifications you have, and many times on a phone or smaller models, that's pretty limiting factor. And the other challenge with that is in many of these models, there's a really relatively high capex, capital expenditure, that's necessary and often to continue to update these.
This is for many of us local AI is all of the technology we've used in the past, but now we're adding AI to the picture versus the network AI. We have to worry about business associate agreements and not retaining any data locally and cloud encryption, which brings a whole new challenge. And the reality is that the internet drops, like in a place like me where I'm rural, internet drops were kind of hosed in that case. But the good things about that is you kind of have relatively virtually unlimited compute powers and lots of multimodal LMs and other vision models that you can use. So like I said, there's goods and bads of both. I think the future, just kind of predict what it might be, is we might end up having a spot where much of the processing is happening on local or edge AI, whether it be live internet or temporarily stored on a phone or a tablet or an OCT or whatever, that then communicates with network AI.
And I think they were going to see a lot of diagnostic imaging and slit lamp AI, and it's all going to be local. Whereas a lot of the ambient scribe and how we deal with patient communication, that architecture is going to be network AI. So we're likely to see very much a bridge in how that happens. So kind of pay attention to those terms when you hear edge AI or local AI. Think about the goods that's happening on your phone, versus when you think about what we're used to, which is cloud AI. When you log in to your internet and you're talking to your Gemini or GCLOD or GPT or any of those types of things, that that's your network or cloud-based AI.
So hopefully that helps a little bit in terms of kind of our local or I should say our knowledge bite of the day. And for the fireside, Rehan's taking a little break this week, so I'm kind of covering the whole thing. I want to talk a little bit about an AI Bill of Rights. Now, I know there's something that I'll try to make this as quick as I possibly can or summarize it. And what a Bill of Rights is, much like a Bill of Rights or the U.S., is a set of guiding principles designed to protect civil liberties and privacy and equity as these automated systems and artificial intelligence become integrated into critical sectors of society. In particular, we're talking about healthcare.
Now, originally the White House, the Office of Science and Technology, in its 2022 blueprint for AI Bill of Rights, it kind of gave five core pillars. You're going to hear us talk a lot about those core pillars in the next couple months is, is the system safe and effective? So there are lots of systems out there that are safe but not effective, which means they're useless. There's lots of them that are effective, but they kind of borderline on the safety issue. Then we have to worry, is there any algorithmic discrimination that is involved? Is there bias? And you've heard us all talk about this before. Is there bias everywhere? So I think that's one of the tricky ones is, I think that bias is almost inherent, but is it going to be transparent where we can identify or have the option to identify what the bias is?
Is there data privacy? In other words, who has access to this data comes down to access. How are things explained? In what way are they explained to the general consumer that's out there that has limited technology knowledge? And what's the fallback system? Where do we become augmented instead of our official? In other words, how do humans interact with the system? So let me kind of break this down really quickly. So safe and effective systems. So we think about diagnostic images and risk stratification. Are we going to have the correct information? Or did the system learn from information that maybe wasn't representative of the patient population that it's actually going to do?
And number one and then number two is, are we going to have drift or hallucinations that occur because EHRs are relatively corrupt and have bad data? Are we going to have some way that that affects the safety and efficacy of the system? As we kind of mentioned earlier, algorithmic discrimination, I think that's a big one and same thing. EHRs are real challenge with that. But how do we see is there discrimination present? Will there be another AI that checks the first AI to see if it's discriminatory? Or do we have a place where, as we talked about with humans involved, where we go, wait, that doesn't exactly make sense. Then we get a day of privacy.
And I think we beat this one to the ground in many of our talks in terms of, do we have a very granular consent about exactly who owns, who's the stewards of this information? And who's restricted against using that? And is this data restricted against unauthorized commercial secondary use? And I think that's a really big piece as we get into the day of privacy discussion over the next couple of months and really next six months as we look into who has access to the data and how is that data scraped, aggregated, packaged, sold to third parties without authorization? Once again, transparency becomes really important in that. Fourth is kind of as we said, explainability is, is everybody going to be involved, both the clinicians and the patient?
Do we have a clear, interpretable rationale of how these recommendations were generated? Do we have transparent view of exactly what the brain, the black box, was thinking? Last but not least, as we kind of mentioned in the last three or four pieces is, last but not least is, do we have a human alternative? If AI fails or if AI gives an information that doesn't sound quite right, does the human condition remain in the loop with the authority to override the system? These are all important things that we think about when we get to the Bill of Rights. So as always, thank you very much for joining us. If you have any questions, you can reach me on any of these subjects at ainicare.ai.
Happy to do it. You can also check us out on our website, ainicare.com, where we have tons of resources, podcasts, various different kinds of podcasts, resources, opinions, comments from the audience, all those types of things where you can keep abreast of what's happening in AI this week. Thank you for your time.