Subscribe
Share
Share
Embed
The Clinical Excellent Podcast, sponsored by the Bucksbaum Institute for Clinical Excellence is a biweekly podcast hosted by Drs. Adam Cifu and Matthew Sorrentino. The podcast has three formats: discussions between doctors and patients, discussions with authors of research pertinent to improving clinical care and the doctor-patient relationship and discussions with physicians about challenges in the doctor-patient relationship or in the life of a physician.
[00:00:00] Dr. Cifu: On today's episode of The Clinical Excellence Podcast, we have Dr. Sachin Shah talking about the current state of AI in clinical medicine.
[00:00:13] Dr. Shah: You know, this 1.0 version of, you know, of ambient clinical documentation, which is very powerful in and of itself, it's doing my documentation for me. It's giving me a bunch of time back in my day, and it's allowing me to focus more completely on my patients, and the patients very much value that as well and notice that.
[00:00:40] Dr. Cifu: We're back after a bit of a hiatus with another episode of The Clinical Excellence Podcast, sponsored by the Bucksbaum Institute for Clinical Excellence. On this podcast, we discuss, dissect, and promote clinical excellence. We review research pertinent to clinical excellence. We invite experts to discuss topics that often challenge the physician-patient relationship, and we host conversations between patients and doctors.
We took a bit of a spring break, but we are now back with our every-other-week schedule, and we have some great conversations coming. I'm Adam Cifu, and today I'm joined by Dr. Sachin Shah. Dr. Shah is an associate professor of medicine and pediatrics at the University of Chicago and our chief medical information officer.
He is board-certified in internal medicine, pediatrics, and clinical informatics. As CMIO, Dr. Shah helps lead key digital transformation initiatives at UChicago, with expertise in population health and QI, virtual models of care, healthcare delivery innovation, predictive analytics, patient engagement, and applications of generative AI in healthcare.
Dr. Shah is also an associate junior faculty scholar in the Bucksbaum Institute for Clinical Excellence and my son's doctor. Sachin, thank you very much for joining me.
[00:02:00] Dr. Shah: It's really great to be here, Adam. Thanks for having me.
[00:02:02] Dr. Cifu: With that introduction, I can't imagine you actually have time to be sitting here.
So I'm going to start with an easy question. You're an internist and a pediatrician, so tell me a little bit about your history with clinical informatics. Was it something you were interested in during your training or is it kind of an interest and expertise that developed once you became, you know, more of a senior doctor?
[00:02:24] Dr. Shah: Yeah, you know, it was interesting. I didn't have any real background or inclination, I think earlier in training. I think the field is relatively newer, to begin with, but you know when I became a junior sort of faculty member and I joined here in 2010, I inherited a lot of patients, right? When I came, I was pretty busy clinically, you know, eight, nine half days of clinic a week when I started. And so seeing a lot of patients in primary care, you know, as you know, we see a lot of chronic disease, we see a lot of these conditions, and out of training, I was pretty, you know, well-versed in the evidence-based guidelines. And I just noticed that, you know, we all manage our patients pretty differently. And yes, our patients are all different, but there's a lot that's shared and there's a lot of management that should be relatively standardized, you know, to be in line with the guidelines. And I just thought about, okay, how do I improve on that, right? Like I want to make sure my patients and actually patients across the health system and beyond have an opportunity to get the best care in line with the latest guidelines, and informatics became pretty clear, you know, as a way to do that, you know, with things like clinical decision support and embedding some of that knowledge into your daily workflows because it's hard to just run a bunch of education campaigns.
Sure.
[00:03:41] Dr. Cifu: It's interesting to hear you say... Hey, I can't believe you've been here for 14 years, you're like an old man now. I remember some of our early conversations were about the kind of variability in healthcare and technology, you know, maybe being able to standardize that where standardization would be beneficial.
I feel like that was almost a very trendy thing to talk about. I feel like I hear less about that now. Do you think it's because we've kind of accomplished something there? What do you think that's about?
[00:04:14] Dr. Shah: I think to a certain extent, we have. I think, for all the downsides, you know, and the bad press that EHRs get, you know, some of it justifiably so, there are some real benefits to it in terms of standardizing approaches to care delivery and making it easier to show things, you know, at a population level, individual panel level, at an individual patient level. And I think we've leveraged a lot of that in recent years. Not perfectly by any means. And if that was sort of the 1.0 kind of version of it, what we're going to talk about today is sort of, you know, 2.0 and 3.0 version of it.
[00:04:53] Dr. Cifu: I think that's true. It's nice to hear you say that. I mean, I'm sort of weirdly an EMR fan. I feel like I've seen... We all complain about things, but I've seen how much it's improved. And I think you're right that some of the standardization, you know, just kind of comes with everybody working with one of the same tools, which supports us all in a similar way has really helped without feeling like you're being forced to do one thing or another.
[00:05:23] Dr. Shah: Yeah. You know, the way that I think about it is, you know, it's GPS, you know, it's not self-driving, you know, but it's GPS. You know, every time I drive home from work, I know the way, but I still put it into the GPS because it helps me get home more efficiently and bypass this or, you know, streamline that but I'm still driving, but it gives me a little guidance and additional support that makes me better, more efficient.
[00:05:47] Dr. Cifu: Do you worry that it'll delay your diagnosis of dementia? Getting assistance...
[00:05:51] Dr. Shah: Yeah. Possibly. .
[00:05:54] Dr. Cifu: So, you sort of gave me a nice segue into it. I feel like we've heard a lot of people, many of whom I don't think actually know a whole lot, talk about how AI is going to change medicine.
And I got to say, I've been a little bit disappointed with people predicting the future because I feel like they almost lack the creativity to understand how enormous it's going to be. And so what I was really interested in talking to you about is kind of the current state of affairs, sort of like where we're already seeing AI, how you think AI is already helping us before we get to some predictions.
[00:06:33] Dr. Shah: You know, I like that topic because I think there's plenty to talk about in the current state. And it's evolving so rapidly. I think we all sort of recognize that there's certainly a lot of buzz but there's some, you know, actual reality to it now too. And, you know, the timing of this conversation is good because we're just two weeks into, you know, deploying here. A pretty wide deployment of, I think, what's the most mature sort of use for generative AI right now in healthcare, at least on, you know, the healthcare provider side, and that's ambient clinical documentation. And I think, you know, a lot of us have probably heard some version of this, but you know, the short version is that you know, as a clinician, I walk into the room, you know, on an app on my mobile device. I start it. It records the conversation, of course, with my patient's consent. And at the end of that conversation, I press stop and I click generate note and it'll do my documentation for me. It'll write like a very coherent, pretty nuanced note for me.
And this is something that, you know, you and I both know can take a long time for us and certainly for a lot of our colleagues. They're spending, you know, hours on end after work, you know, at night, that's time, you know, away from the things that they might otherwise want to be doing. And that's pretty fascinating.
It also, you know, in the room, in the moment allows me to sort of turn my chair and face my patient, you know, instead of... You know, I'm someone that... You know, a lot of us can sort of maintain eye contact with our patient and type but we're still splitting our cognitive bandwidth while we're doing that. We're missing things and patients notice that, when you're sort of dropping a bunch of, you know, a long 'umm', you're like buffering, right? Because you're thinking about what was just said and trying to document that so you don't forget it and you miss what's being said. And the conversation becomes much more... With the focus on the patient, with Full eye contact, with full attention, it just becomes a much richer conversation and it feels like it's bringing me back to, I think what we all want to be doing, which is focusing on that real sort of clinical encounter and that interaction.
[00:08:43] Dr. Cifu: I think maybe we'll try to put in the notes a link to some of the videos that show how this works because without having experienced it, it really seems like magic.
And what's been interesting is hearing some of our colleagues talk about things that they've been surprised that the note writer has kind of picked up where they sort of mention off-handedly, "Oh, you have a little bit of a rash here, maybe you could pick up some, you know, hydrocortisone," and then you know at the end of the assessment and plan, it will say, you know, "Chest rash, recommend over the counter hydrocortisone." You know, like, "Huh! I would have not even documented that if I was documenting it myself!"
[00:09:23] Dr. Shah: Right. And that's, you know, like a good segue to some of the real power of it. You know, there's 1.0 version of ambient clinical documentation, which is very powerful in and of itself. It's doing my documentation for me. It's giving me a bunch of time back in my day, and it's allowing me to focus more completely on my patients, and the patients very much value that as well, and notice that, but what it also will do over time and these iterations are coming in a matter of like months, not years, but it'll take all this unstructured information that historically lives in our notes and help me capture it, right? Like structure it and say that, look, if I said that, yeah, why don't we go ahead and get a chest X-ray and an EKG today? And then why don't we get this set of labs and you can, you know, call out a couple of the labs, and it'll just tee that up for me, right? It'll say... If my patient, you know, a lot of the patients we care for here on the South side of Chicago, they have, you know, complex social and, you know, medical needs. And that absolutely, you know, has a big impact on the level of care that they need and the ways that we can, you know, support them better in their healthcare. And if they say off-hand, you know, "Gosh, you know, like I'm sorry, you know, that I missed that appointment, like just my transportation didn't come," or "I had this childcare issue," or you know, and we're not in the habit of systematically documenting that because it's effort. It's additional effort. And if this is note number, you know, 12 that you're writing at, you know, 11:30 and you're trying to get to bed, you might sort of, you know, you might not do that. This will capture that, you know, automatically and put it at the patient level and that can help inform, you know, the way that we more completely understand our patients.
[00:11:13] Dr. Cifu: Have you noticed, I'm interested to hear you talk about that and I might go in a different way with that given my diagnostic reasoning obsession, and maybe you don't have enough experience with it thus far but I do wonder if what we'll start seeing is kind of clinical clues that maybe you'd miss in a conversation, which then gets laid out there to say, "Huh, you know, I had data points A, B, C, but in fact, there were, you know, data points D, E, and F which kind of went by quickly in the conversation, which maybe I should pay more attention to.
[00:11:50] Dr. Shah: To also consider and inform my thinking, yeah, I think that's the next like sort of version of clinical decision support is, here's what's on my differential based on yeah, what I heard and what sort of fully registered, but here's the model supporting you and saying, again, it's a co-pilot, it's like the GPS, it doesn't take over driving, but it's going to say, "Hey, you know, consider making a right here and a left here," and that might help you round out your differential, think about things in a more complete way. And I think most clinicians would welcome that because it makes you a better clinician.
[00:12:25] Dr. Cifu: I'm not sure if this is something to respond to, or it's one of those things where people get up at a conference and just say something which has nothing to do with... I read recently an essay that was written for something else that I work on where the person kind of was talking about how they do a better job interacting with the computer and their patient in the room. And 10 years ago, this was a big deal, right? We've introduced the EMR. It brings a lot of great stuff but it also causes this problem where there's the temptation to look at, you know, the screen. We have a colleague who spent all this time trying to talk about like how to do this well. And I feel like this is almost another technology to solve the problem that the last technology created, but I do think it's adding something even beyond a solution to a problem that we created.
[00:13:19] Dr. Shah: Yeah. I think you're right. You know, I often sort of talk about this as like, you know, the original sin of the EHR, which was, we were not, like as physicians, as clinicians, we were not involved in sort of the design and the deployment of it. And it took on a life of its own, and a lot of the negative, you know, sort of associations I feel stem from that. This is sort of an opportunity for us, this version of the technology, this is an opportunity for us to, yeah, to sort of fix a lot of those things. And like I said, you know, turn the chair away from that screen altogether, like not necessarily have to form the triangle, like let's actually make a straight line, right? Which is what I think we all want to do. And yeah, sure, we can refer back to the screen to show some lab values and talk them through and like, look at some imaging together and that's nice but you don't have to, you know, you don't have to be sort of factoring that in. But I think it adds, you know, I think it does add like, just like we were talking about, you know, with things like decision support and streamlining orders and, you know, helping us capture information that we might not, that helps us form a more complete picture of our patient and help sort of allocate limited resources that we have to the right people. And, you know, that's a version of risk stratification which we are always trying to do. Like, let's get the right people the right things.
[00:14:36] Dr. Cifu: Got it. The other tool that I know we've used is having generative AI help us respond to notes which I've mostly been struck at how much more polite and maybe verbose AI is than me, but does an incredible job actually kind of responding personally to patients. What have you seen with that thus far?
[00:15:05] Dr. Shah: Yeah, you know, just like this ambient clinical documentation, we're pretty early adopters. We started in October. We were one of the first, you know, tenish sites nationwide to adopt this. So, we did it with our eyes open, you know, it's an early version of it. And you know, I'm in the habit of assigning letter grades for new technology. I gave it like a B minus, C plus, you know, for the first couple of months, we've been on our end doing a lot of like, prompt engineering, like just pushing it a little bit because, and you're talking about you know, the patient advice request messages that come in and that becomes a crushing sort of... An additional crushing sort of after-hours burden for us. And as clinicians, you know, we're often dealing with trying to answer 12-15 messages at the end of the day, and, you know, these are legitimate questions, you know, they don't necessarily need appointments to come in. We don't necessarily have the capacity to accommodate those appointments anyway. And there's some urgency to the questions anyway. And so it's, you know, it's a good medium to do it, but it's just... It's a lot for us to do on our own. And so, yeah, like you said, it helps sort of tee up a response, you know, and what I've had it do as we've adjusted our prompts over the last couple of months, as we've played with this and iterated on it is, okay, you know, for a pretty significant subset of topics, you know, there's good patient-facing evidence-based resources out there that you can reference and say, "Hey, you know what? Tee up a response informed by this." Those things don't, generally speaking, change a lot, you know, like how you manage, you know, upper respiratory symptoms, when you might not need to come in for your back pain. Okay, what do I do about this sore throat? I mean, you know, there's some, you know, pretty good standardized knowledge to share, and that helps tee things up. And then we can sort of take it from there. You might add a line or two, you might subtract something, but hey, if 80% of it is written, that's a huge help. And if it's pretty good, I'd say it's getting to a B, you know, to B range right now, and we're going to keep working on it to get it better.
[00:17:11] Dr. Cifu: It's cool listening to this because I didn't think of this before, but you know, one of the things that the electronic health record did, which was terrific, is it really improved patient access. Right? It's very easy. And that's good because there used to be too much of a wall and too much hassle between patients and doctors.
[00:17:33] Dr. Shah: That's right.
[00:17:33] Dr. Cifu: The problem is that if someone wakes up at two o'clock in the morning and has an itch, you know, "Yeah, I'm going to put this in this," and you know, maybe we needed a little bit more of a barrier. And so, here's again, a tool to sort of fix some of those original problems that like, okay, maybe we can have, you know, a supervised computer respond to the kind of low-importance questions and then maybe help us with some of the higher importance questions.
[00:18:02] Dr. Shah: I think that's absolutely right. It's a personal assistant at some level, right? Like a good personal assistant. And it's not doing all your work for you but it's streamlining some of it. So you can focus on some of the... You know, we talk about working at the top of our license in healthcare a lot and it helps us do that. And actually, you know, our nurses who often are the first ones looking at this, they have access to this tool as well. And so they can also tee up responses and that helps them, you know, "Okay, I'm going to write something that requires some medical decision-making, not just, you know, the salutations and, you know, the reassurance.
[00:18:38] Dr. Cifu: So my last question is one that I just can't not ask, you know, thinking a little bit about the future, and maybe I'm not talking about the future of retired Adam Cifu going to see his doctor, but, you know, me a few years down the line, kind of what do you think is next that's going to be in our practice assisting us pretty routinely? And it may be, you know, it may be the improvements and the evolution of the things that we've already talked about, or it may be new things. What are you kind of looking forward to, I guess?
[00:19:12] Dr. Shah: Yeah, you know, I think it's going to evolve really rapidly. These are early use cases, and even though they're really good and impressive, I think they're going to get, you know, an order of magnitude more better pretty quickly. And like I said, on the order of months to, you know, a year or so, just think of the year before, you know ChatGPT came out, you know, we had no sense of that. I mean, unless you were really, you know, sort of in the AI research world, you had very little sort of insight that something so transformative was coming.
But I think, you know, rounding out some of this stuff, like I think the documentation tools will get much better and it's going to capture all this structured information directly in the clinical conversation. And so like the things that, you know, sort of crush our spirit as clinicians, like coding queries, and like all this stuff, all this focus on documentation, which feels so backwards, like, you know, reimbursement-related, you know, the payers need this and that. I think a lot of that is going to by and large go away. What I really want to... You know, like our goal here is that the only time that you spend thinking about or documenting, you know, on that note is the time you spend with your patient. And it's going to happen almost automatically. And you're going to review it and you're going to sort of finalize it, but that should take seconds and you should have a high degree of trust in it. It's going to help us, you know, support decision-making. It's going to help us capture social determinants of health. And then, you know, the other part of it that, you know, one of these things, one of these metrics that I think is really fascinating and is novel with this ambient, you know, clinical documentation technology specifically is, you know, you heard me say, I press start when I go in the room and I press stop when I leave the room. It tells me very directly how much time I spent with my patient as a clinician. And that's like, you know, we've tried all these surrogate measures in sort of the optimization of ambulatory operations kind of world to figure out how much time we need to spend with a patient. I often tell my patients when I'm running behind like, "Oh, it's this morning that, you know, I'm really sorry I'm running late, but I'm not doing oil changes."
You know, like today I can review my morning and say, okay, I spent 18 minutes with this patient, 42 minutes with this one, 25 minutes with this one. And they all have 20-minute visit lengths. So what we can do with this data now over time, a couple of visits, I can say, okay, let's develop a predictive analytics model. Like how much time do I expect to need with this patient? Let's have like some dynamic visit lengths. And that can change the way we think about like access and capacity and like, you know, we can accommodate more patients, or have the right amount of time with the right patients. So I think that's also, you know, an interesting, you know, sort of manifestation that's coming fast. I think people are factoring that in.
And then, you know, in terms of diagnostics, I think we'll see a lot more, you know, manifestations or applications of predictive analytics that are powered by AI. Our traditional sort of models are what we call deterministic, right? You put the same set of variables in, you get the same answer every time. These models, LLMs are more probabilistic. You put the same stuff in, you might get different answers, but there's some nuance around it. And that can help us think about these things in a more nuanced way. If, you know, it's not about like a red line threshold, it's about, okay, what's the likelihood, you know, what's the nuance we can actually factor in things like pre-test and post-test probability.
And I think it'll bring some more nuance and it'll inform, you know, sort of clinical decision-making in a much more robust way. And that's also exciting. I think there's going to be more.
[00:22:42] Dr. Cifu: Really exciting stuff. I've been suppressing my, you know, brave new world fears. So let me just ask one question, which came up as you were talking, because I think one of the things we miss is often the conversations which happen before we even come into the room, right? Where there's conversations with a medical assistant or a nurse, you know, early on, where there might be important concerns that are raised, which then actually don't even come up with us. You know, the way we're structured now is, you know, there's consent at the front desk, but there's consent in the room. Do you think we'll get to the point where this is sort of built into the rooms and my fears, which hopefully you'll assuage, is does anything kind of last as a recording of these visits, or is this kind of processed and then disappears?
[00:23:33] Dr. Shah: Yeah, those are great questions. I mean, it's a really insightful comment because this is what we're working towards probably in the intermediate term where it becomes less about like me as a physician has this tool on my phone and press the start and stop at the beginning and end of my specific interaction with the patient but moving it to the patient level. And so, you know, yes, we press start when the MA comes in, get some details, they might drop, just like you said, you know, a detail that might not come up later but is important. And then maybe the medical student comes in and then, you know, we'll get some nuggets that they may or may not present to you.
And then, you know, maybe the RN comes in and has a conversation but you know, all the members of the care team and then it actually stitches it all together into one sort of, you know, note. And I think that's the direction that this can and should go because I also, you know... Look, you know, our residents, our APP colleagues that do a lot, they share a big sort of documentation burden and that should absolutely be a part of this.
And so what I think we'll end up seeing is that done at a more holistic at the patient level. The second part of your question, you know, what happens with these recordings? We've been very clear, you know, the approach that we've taken is the recording lasts only until the note is finalized, right? It's there to help generate the note. Once you press sign on the note, the recording actually goes away. And it's not kept and the longest it's kept is 30 days. I think that's how it should be. This is PHI. You know, there's going to be plenty of folks that want to do research on this and that should be done with eyes very wide open and there should be separate consent for something like that to hang onto.
[00:25:14] Dr. Cifu: We don't need to start building data centers to save all of the patient-doctor communications. Sachin, thanks so much. I think this will probably be a kind of yearly check back in on the podcast because we do also have to really get into how this, I don't know, bumps up or assists the actually clinical excellence in the doctor-patient relationship, which are all important things. So I hope you will accept a return invitation.
[00:25:43] Dr. Shah: I would love to. I would love to.
[00:25:44] Dr. Cifu: Thanks for joining us for this episode of The Clinical Excellence Podcast. We are sponsored by the Bucksbaum Institute for Clinical Excellence at the University of Chicago. Please feel free to reach out to us with your thoughts and ideas via the Bucksbaum Institute webpage.
The music for The Clinical Excellence Podcast is courtesy of Dr. Maylyn Martinez.