All right, I just have a few housekeeping things to go over. Thanks for joining again tonight. The views expressed during this session are those of the individual presenters and participants and do not necessarily reflect the positions of AAP MNR. AAP MNR is committed to maintaining a respectful, inclusive, and safe environment in accordance with our code of conduct and anti-harassment policy available at our website. All participants are expected to engage professionally and constructively. This will be recorded and made available on the online learning portal and email will be sent out with a link to bring to the reporting and an evaluation and there will also be an email explaining how to claim CME for this session. For the best experience, please mute your microphone when you're not speaking and for questions, please use the raise your hand feature and unmute if you're called upon or use the chat feature to take your question. And just to note that we may not have time to answer all of your questions, but I will turn it over to Dr. Tariq to get started. All right, thank you so much, Kayla. I appreciate it. Thank you everybody for joining. We're excited to start this new community. Basically, I've always been interested in innovation and AI and I would just basically meet people, including the people on this call right now at the APM in our conference or on LinkedIn and people are always basically talking about this and I really feel like there's a platform for all of us to discuss ideas we're basically implementing in our practice and how that integrates into our day-to-day work. So I decided to kind of set this up. It's very informal. Please, anytime anyone has any questions or comments, please raise your hand in the chat or basically just unmute yourself and please be able to interrupt us if you need to. We want to make this into more of a dialogue versus a lecture. I know there's a lot of slides that will be coming up, but at the same time, you know, all of us have used either something like GPT or other, you know, obviously different platforms out there in different ways, but the only way we can get better, especially this early on in our careers or this AI process is to learn from each other. So without further ado, I'm going to actually have our first speaker basically discuss his, you know, basically point of view as a medical student and how he's integrating technology into his process and then we'll go from that to an attending, an early career attending to someone like me who's passed 10 years and actually I just realized they're going to have Chris, you go first. I'm sorry about that. I thought it was Matt. So Chris, why don't you go ahead and share your screen and do a quick intro and we'll go from there. Thank you. Thank you. Yeah, for the great introduction and really excited to share with you some of my experiences with AI specifically in implementing a pilot at our institution at the University of Washington. So initially I just wanted to give a brief background on who I am and kind of where I'm coming from for this pilot. So trained a lot in across either University of Washington or spent some time in Chicago at Northwestern and the Shirley Ryan Ability Lab. But really the I think the takeaway I wanted to share with sharing the slide is that there is a fellowship called clinical informatics that if you're interested in the intersection of technology and clinical practice, it could be a great thing to explore and see if it's a good fit for you. I'll talk a little bit about what that is. But a lot of my background was initially in, you know, cancer research or then cancer rehabilitation. So I think every subspecialty really needs more ways to leverage technology. And so whatever you're interested in clinically I think there's great opportunities for this. And I'll be joining the UW Medicine Department of Rehab next year as assistant professor so excited for that. But what is specifically clinical informatics before jumping into this AI pilot. It's a two year fellowship, ACGME accredited. It's offered at dozens of places across the country and any specialty can do it. So if you're a medical student and you're exploring different medical specialties and you're not quite sure about PM&R, it's a great specialty. I did as well. But I think the informatics is an opportunity after that as well. And you really get to think about how do you leverage technology to improve the lives of patients, providers, the hospital system, population health. So there's just so many ways to think about this technology. And really, I think when people ask me about the fellowship, a lot of times they're like, well, I don't have a background in computer science or other things. And I don't either. And I think a huge part of this is really understanding the technology and how people use it. So the role is really maybe 90% working with people and 10% about learning how the technology works. So what is this pilot that we're doing on ambient AI and why are we doing this? So we're currently evaluating ambient AI scribes at UW, which I'll go into a little bit about what that is. But why is this important, first of all? So specifically, there's been surveys, you've probably seen these about burnout. And that many providers have at least one symptom of burnout, almost half of them in a 2023 survey. Additionally, we know that using the electronic health record more and having more administrative tasks are also associated with lower job satisfaction. Over time, the amount of time providers are spending in the chart every day in the electronic health record is going up with primary care providers in a recent study having almost six hours of their day spent in electronic health record and on average another one and a half hours at night, which is quote pajama time those days. So I think really when we're looking at ambient scribes, a lot of the preliminary studies are finding that ambient scribes decrease the time in notes, they decrease the after hours work, they decrease documentation burden, and they improve the well being of providers through different measures. So I'm going to go through a little bit of how we're looking at this at UW. There's a lot of different vendors, and I won't go through all of them, but Abridge is just one of many. There's also Microsoft has one through Nuance called DAX, there's Nobla, there's a lot of different others. But the way this one specifically works is you have your mobile device that has the Abridge app or the Epic, in this case, Epic MyChart. It can also work on other electronic health records as well. But essentially the application on your phone allows you to record a conversation with a patient, run that recording through a large language model AI algorithm type of chat GPT service that transcribes it into the dialogue and then takes that dialogue text and puts it into a note format. So the workflow is very, doesn't really impact the typical clinical workflow, unlike a lot of the technologies that we try to implement. It really flows well, providers really just have to get consent from the patient, press record, and at the end of it, push the recording to the cloud, and the recording or the transcription and the note template gets dumped into their note for them to edit. So how do we study something like this? The way that we're studying it is through a randomized parallel control trial where we have providers who are using Abridge or the AI technology and those who are not. The specialties we were testing specifically were in primary care, specialty care, in urology, and then also at our cancer center. And we were doing a convenient sampling method for anyone who was interested. So we just did open call, who wants to come and participate in this pilot, and then we stratified them into these intervention groups and control groups. So the intervention group got Abridge, the control group did not. Some of these other inclusion criteria were just based on the limitations of Abridge in the way that we wanted to study it. Specifically, they didn't have accessibility for the Android devices yet. But I think the most interesting part is how do we think about measuring the impact of this? We really want to know, and anytime we're thinking about evaluating technology in the healthcare system, we want to think about every single way it can impact care, not just providers or patients or the hospital system, but all of this all together. And so we really had to pull together a lot of different data sources to start looking at this. And key things we want to look at were provider adoption, how accurate are the notes that are generated by the technology, how does it impact workload like we mentioned, and how does it impact clinician and patient experiences. So the way that we tackle the provider adoption piece is we added in a basically note writer consent option for providers to click when they're signing their notes, so they could say what they were using Abridge for, whether it was for charting purposes only, or it wasn't used, or the patient opted out of using it for the visit. So we wanted to know, are patients deciding not to use it for a visit, and why? And we wanted to make sure that patients felt comfortable declining to use it if they didn't feel comfortable with it. So one of the benefits of gathering this information is we know all the other information about the patients, the providers, and the encounter, so what characteristics are really providers associating with the choice to use this technology? So if we're thinking about any technology, not just AI solutions, what factors are going into the decision of the provider to decide to use it or not? And so that's been a really interesting question for us to start diving into. When we think about the second option of note accuracy and usability, there's this whole process that the data goes through. And so at each one of these points, from the recording, to the transcript, to the AI draft, to this final signed note by the provider, there's opportunities for data to be lost or misrepresented, either through the AI model or from the providers themselves. So, for example, going from the AI draft to the signed note, we can track what percentage of the AI draft is actually retained in the final note, meaning, did the provider delete 50% of it and fill it in with something else? Did they delete 100% of it, or did they keep most of it? The other way we can see how usable it is, is the rate of documentation errors. So it's hard to know what the baseline accuracy rate of a provider writing a note, because we don't have a recording to compare it to before this. So, you know, what percentage of the note is actually accurate? But we can ask patients to read their notes and respond with any errors they identify and get a sense of, you know, how often are we seeing errors in our notes and use that as a baseline comparison. So we can know from the recording to the signed note, how much of the content is accurate. Finally, I guess the third one here is looking at workload impact. In Epic itself and different electronic health records have tools to assess use or to track how providers are using the technology and Signal is the one for Epic. And so, really, you can look at like the number of clinical encounters a provider has, how many work RVUs they're generating per encounter. One of the key things is, you know, same day closure rate. So as you're writing your notes, people want to have those notes closed at the end of at least in a few days. And the sooner the better. So is this technology by having a draft generated for the provider actually impacting the rate that they're having a same day closure? Finally, we were talking about pajama time and time in notes per day. This really lets us look at those variables and see how it changes from providers who are in the control group to providers who are in our intervention group. So, and then lastly, we are surveying our providers throughout the pilot asking them for their professional fulfillment, as well as the task load of generating of documenting their notes. So we'd seen in the literature that, you know, professional fulfillment index is very improves actually with the use of a bridge or any other of these AI technologies. Also providers were reporting that their task load of actually documenting the notes was less. So these were some of the metrics to allow us to look at these. And this is a publication out of NEJM that was out of this year, a little earlier this year, looking at some of the initial results for usage and these metrics. And in the graph on the left here, what's interesting is the gray line are non-users, the green line are low users for the AI technology, and the blue line is high users. And you can see that over time, the amount of minutes per note, which is on the y-axis here, drops for everybody, but the high users by the very end of the time period were actually having the least time in notes compared to non-users, where non-users were maybe more steady and low users were also more steady. Interestingly as well, it was higher for the high users at first, which suggests there's a learning curve to how do you use this technology? You know, is it actually as useful at the beginning? There may be some may feel slower at first, but as you learn it, it pays off. Additionally, they looked at two different vendors in this study, which I think is also interesting to report that depending on the vendor, they noticed a difference in number of notes generated where vendor two seemed to actually generate, people use that one more and generated more notes compared to vendor one. So when we're looking at these different technologies, whether it's for Ambient Scribes or others, consider the different vendors and the services they're offering and try to look at the literature to see which ones are actually having the biggest impact. Lastly, for patient experience, this is something that we are sending patients via phone or email for anyone who is engaging in these, in the visits with a bridge or with our AI technology or not, just to learn, you know, how are they assessing their clinical visit? Because you could imagine there could be benefit for the patient, and you know, how are they assessing their clinical visit? And the study I was just referencing also looked at this, where they showed that there was actually, patients were reporting that they were feeling better about the time the provider was spending speaking with them, looking at them instead of the computer screen, and then the quality of the visit and their overall comfort. And so I think we see here that about half of patients really felt like there was a benefit from using the technology. So it's important also, one of the things that we want to look into that wasn't listed here was the patients who are declining to use the technology, what is their logic? What do they care about? And how can we better serve them, either by making the visit comfortable in other ways without using AI, or addressing their concerns about it, so that we can make sure that everyone feels safe with the technology that we're using. So that's the end. I tried to cover a lot of information in a fairly short period of time here. But happy to take questions, or maybe we'll save them for the end. Yeah, great question. So we elected to study specifically in outpatient to try to have a more homogenous group. We discussed doing the emergency room, but again, decided just to focus on outpatient. There are a few rehab providers in the pilot, but it is mainly primary care. And then oncology and then surgery. So we specifically chose the surgical urology group because they actually already had scribe services in many of their clinics, whether it was virtual or others. And so we wanted to know, did they feel the content generated by this technology was comparable to the technology that was generated by a human scribe? So that's my second question, then. Are the AI scribe systems out there, are they trained on MSK? Are they trained on physiatry lingo? Does it matter? Yeah, it's a great question. And I think the templates that we have available with these different technologies, at least with a bridge, is mainly designed for primary care. And so that is something that we're hearing from our subspecialists is it's not quite as good at generating problem lists for specialized care. And that's what I've heard because I've reached out to our rehab colleagues who are in the pilot, and they're like, I got to delete a lot of stuff because sure, the patient brought up something else that's not even like their hypertension or something, but that's not what I'm managing this visit. I don't want to have that as a problem list that I'm covering. And so it doesn't necessarily filter all those out for the specific specialty. So that is that is some of the feedback we're getting that the content isn't necessarily customized to the subspecialty quite yet. They may want to build those templates. Yeah, and this is going to be a problem for people who want to start doing this. And I think the way that at least I've done it, I think other colleagues same way is that I've had to pretty much train my GPT to think that way. Especially Doximity is that you are attending physiatrist board certified in brain injury and think this way. So I think it has a better chance doing that. But I think over time, obviously more as it gets more data, and that's going to get better. Other question I have is obviously you're, you're in the academic setting and there's a lot of people here that are prior practice independent physicians. How would you think about going through because a bridge, I've talked to them, they're not very affordable for prior practice, you know, and there's so many different softwares out there. How would you even judge which one to use? Do you pilot it? Do you like ask them? What kind of questions do you ask them? Yeah, that I was a part of our like RFP process as well where we were looking, we did an open kind of call for all the vendors and we had about a dozen who initially kind of filled out our questionnaires. I think the, there's a lot of things to consider. Some of those are around like security and data retention. Like one of the things that was a big concern for us was giving data to the vendor, you know, because like data is really important and we don't necessarily want to share our patients' data. And so protecting our patients' data was really important. But from the functionality perspective, I would say the pricing structure is a big thing to discuss, whether it's individual license or are you paying per use per encounter? I think we're trying to discuss that and we've seen how that affects different vendors and different, we've seen different vendors have different pricing structures. So getting creative with the pricing structure could be a way for smaller, you know, non-academic settings to be able to use the technology. Maybe there's specific visits you want to use it for, not for every single visit, so why pay for a whole license? Because that is another thing we've heard from our providers who, if they have highly templated visits where it's like, I have a great template, I already clicked through everything, it takes two seconds, why would I use this technology? It actually slows me down. If you have visits like that, then paying per encounter may be a better way to navigate it. Yeah, thanks. One thing I'd add quick, there's an idea that like paid models are way superior and I haven't looked at the context of clinical visits, but responding to MyChart messages, we've done a lot of research that open source, free, locally run models often have comparative performance to paid models. So there could be opportunities for not having to pay for licenses, not having to front a lot of fees and being able to locally run things that have pretty comparable functionality. Yeah, the buyer build question is a great one. And I think if you go back X number of years to when we were all deciding which EHR to get, I think a lot of institutions elected to build their own initially and eventually transition to buying. But I think at first, especially like building something to bridge until the technology becomes more affordable makes a lot of sense, especially for how effective a lot of these free LLMs are. I just had a question from the academic center perspective, part of our resistance that we're having with sort of transitioning to this is sort of the QI and research needs of a progress note and the very like hanging on for dear life to note elements and clicking things because of like the downstream SDEs and sort of the data value and fully recognizing that, Cosmos, if you're an Epic customer, like that's blowing up, like there's other data opportunities, but like I was curious sort of if you had a deal with that as you were rolling this out or if you had provider resistance, like did you change templates? Do you have an ambient template? Did people like merge this or sort of how did you approach those mechanics? Yeah, that's a great question. And I think change management is a huge challenge with these types of rollouts. And the way we approached it was we had an ambient smart link. So if you're familiar with kind of the Epic lingo, smart link is what you can pull into your note. And so people could add those smart links to their personal template. So if they already had a note template that they'd spent hours designing and their years using, they're like, I love this template, I don't wanna change it. They could just plop in the ambient smart link to their HPI section or whatever. And it would pull in just that piece of the recording. So not everyone used the full note template. I don't think anyone really used like the abridged note template. They just pulled in the pieces to their note. That way they could retain those like dropdowns and those data elements. Did you find that there were certain sections that were more amenable to ambient? Like for example, a lot of our providers like don't think it would help with their physical exam because they don't wanna speak in front of the patient versus HPI and plan may be gold for this. So I'm just curious if you had that too. 100%, yeah. HPI and plan are the biggest usages. I would say people even use the HPI section sometimes more than the plan section because they have like to what we were just saying, I like to write my problem this way and it's drafting it in a way that I don't use that like jargon, I use this other jargon for my problem. And so, yeah, it's a great point. I think the physical exam sections, you do need to say a lot out loud in the room and that can be awkward for the patient right there. There are features coming out from the different vendors I've heard where you can almost do a short phrase that the AI would pick up on. So you basically say normal physical exam for elbow or knee or whatever. And it would just drop that template into your note. You'd still have to edit it and everything because maybe there's differences that you just wanted to put in the template. But I think for the most part, the physical exam section is such, it's also such a kind of personal way we like to write it. And our dot phrases are pretty good that it's not that big of a time save. I think the biggest time saves are for the narrative portions of the note where people are able to have that. Anyway, I've been talking a lot, but- Yeah, we have some amazing topics to follow this. I appreciate it. I will keep on moving. And Dr. Matthew Cowling, I think I'll have you go next if you don't mind sharing your screen and talking about just doing a quick intro by yourself and going from there. You're muted in case you don't know. You had it on. You're still muted. There. Okay. Okay, and then? I want to share your screen again. There you go. Perfect. And then let me just flip it here. I think your audio is a little bit off too. So I can't hear you too well. Can you hear me now? Yeah, I can hear you. Okay. Yeah, you might have to move closer to the microphone. Sorry. How about there? Is this better now? Much better, yeah. Okay, perfect. All right. So, all right guys, thanks for having me. And thanks for that, Chris. That was a great talk. I'm going to talk about real-world application and AI and how I use this for documentation. I first started using AI a couple of years ago and I've been trying to use it every day. And continually improve and get better. And there's some basic things I think that I'm going to talk about today. There's a lot more advanced stuff out there but just some simple things that I implement and use in my practice every day. I work in a sub-acute rehab and acute rehab and do a telemedicine prosthetics clinic. So every single day I'm using AI in some way, shape or form. So the reason this is important is because as we just previously mentioned, physician burnout is at an all-time high. There's all sorts of issues with regulations and guidelines which are constantly changing for getting prosthetics, for doing injections and imaging and wheelchairs. And it's really hard for us to stay on top of all that stuff and then have to do the work of getting the documentation together and then seeing patients as well. So all of this just really leads to an incredible amount of burnout for physicians. And some studies have come out showing physiatry as one of the highest burnout specialties. And I think one reason for that is because we use so much work in the DME space and dealing with patients with such complex disabilities. So AI is a sort of a solution to this. And it's one that I have found to make things a lot easier during my daily work. So I'm going to do some real world use cases here. I'm just going to show you guys the five most common ways that I use AI. The number one is going to be for prosthetics and DME. I also use it for insurance appeal letters, bullet points for peer to peers, prosthetics, and all sorts of different applications on the day to day. The two, I'm going to talk about two different AIs. One I'm going to talk about is Doximity GPT. This is the oldest one. This is probably the one that you guys are the most familiar with. It's HIPAA compliant and it's accessible to anybody who is practicing in medicine. PAs, pharmacists, MPs. So it's pretty nice. Medical students can utilize it. And it's gotten better and better over time. This is what Doximity GPT looks like. This is just the prompt screen. As you can see, you have your area where you put your prompt here. And then below that, there's some suggested prompts that you could use. The one I'll talk about is open evidence. So I don't use open evidence that much. I think that open evidence is one that I use more clinically. So if I want to look something up about a particular patient, let's say that I want to know baclofen dosing or for a patient with a renal impairment, or if gabapentin is interacting with another medication, then I'll use that on the fly. I have it on my phone and I would just use that while I'm rounding. With the Doximity GPT, that's one that I use more regularly for documentation and justification. So I use a pretty simple workflow. When I get to my computer, as you guys see, I'm kind of toying around with the screens here, but I have a two screen set up almost everywhere that I go, but this would work just fine on your laptop. One in one browser tab, I'll open my Doximity GPT on one, I'll open my EHR on the other. And then what I'll do is I'll go through the EHR and I'll find either the progress note, my HNP or a pre-admission screen or whatever I need to get from that. And then I will copy that. And now I like to get as much information as I can to put that into the AI. And I'll paste that right into Doximity. And then I'll load a prompt above it in Doximity, the prompt that I want, and that can be as specific as you want. I use a lot of very general examples in this slideshow, but then once my GPT pops out what I want, I edit it on there and then I'll either copy and paste it and I'll put it back into the EHR or I'll fax it or I'll sign it and give it to the case manager or whoever needs that information. So multiple ways to use it once you have a completed product. First one I'll talk about is a power chair justification. So another cool thing that you can do is you can put your information into Chad GPT or into your Doximity GPT, your HIPAA compliant, and you can have it de-identify and make that information not identifiable. And then you can move it into a non-HIPAA compliant and then you can do different things with that if you want to kind of create different tools or look more into the different conditions. So this is a 71-year-old long-term care resident. And for the sake of time, I'm not gonna go through each of these clinical scenarios, but more so I just wanna show you guys what you can do with this. So our scenario is a patient with MS who is essentially a wire dependent and we're trying to get them a power chair. In this particular scenario, I did a lot of work on the backend already for this patient's power chair, but my input would be my progress note. What I listed here is just some bullet points because I didn't wanna paste the entire note on the page, but this would be the progress note that I'm copying from the EHR. And I'm gonna take that information and I'm going to put that into the Doximity GPT. And then I'm gonna put my prompt above it. So for this one, I would say, create a justification for a power wheelchair using the following progress note, make it HIPAA compliant with CMS guidelines. And so I would do that. I would then hit enter and then it would pop out something like this. And this one is not a note as I talked about, but this one is more so just like three paragraphs that I wanted to put at the bottom of my note. And it's just going through explaining why the patient needs the power chair, how their conditions that I entered into GPT would lead to them needing a power chair. As you notice it added in some things here, like the patient is too much strain with a manual wheelchair or the use of walking aids. So they need this to provide necessary support, high risk for functional impairment. They add in all these nice phrases that really help us out when it comes to justification. I also use this a ton for my prosthetics clinic. So as you guys know, prosthetics are changing a lot, the regulations, and we gotta constantly try and stay on top of it. This was a 73 year old male with a below the knee amputation, K2 level. So once again, same process. I would input either my progress note or even the note from the prosthetist that they give you if you're reading through that or your HNP or whatever information you have. I'm gonna ask it to write me a medical necessity letter or justification for a left BKA prosthesis. Here's where you can get kind of clever with this. So you could put for a patient with United or for a patient with Medicare, and then it will sort of tailor that exactly how you need. You just have to play with your prompts. For the sake of this, I just made it very simple and it pops out something really nice like this very quickly. So you can even do this while you're in the visit with the patient, copy and paste your note in two seconds while continuing the visit in almost completely synchronous workflow. And it'll pop in the functional goals, the rehab plan, the functional level. What I really like about this is the projected timeline. And of course you can edit this, but I mean, it puts in here K2 level expected in three to six months, which you guys know is very important for now having that expected timeline for the K level with prosthetics. It lists our supportive factors and then essentially our conclusion being the prescription and the prosthetic is essential for the patient. Other full use that I'm using, this is one of those where I was talking about how you could de-identify it, but I don't even think I needed to for this one. A lot of times if there's a patient that is very complex and I'm talking with therapy about a specific brace and maybe I don't know as much as I would like about that particular brace, I'll take all their clinical information and I'll put it into Doximity GPT and I'll ask it what brace it would suggest. And then we'll kind of go back and forth about which brace would be ideal. And then once I ended up selecting one, you can have it help you create the justification. This one was a patient who had normal pressure hydrocephalus, vertebral compression fracture. I know very well was admitted in and out of the rehab hospital in nursing homes and was really struggling at home. We had some issues with him initiating swing phase and he was getting this really intense extension moment when he was going from sit to stand, which was really impairing him. Once he was up and walking, not too bad. So we wanted to get him a KFO. So I just pasted this into the GPT here, write a medical necessity justification for a KFO using the following progress note. I noted he was in long-term care in a skilled nursing facility. And then it pops out something really nice like this. So it'll organize it for you, lists out the brace that you want. This one I really like in the justification. It goes through each component, the stability and support, improved gait, fall prevention, and how that brace is going to help really nailing down the justification. It includes our plan here. We're going to refer to the orthotist for fitting with follow-up with rehab. So very nice. This comes out in less than 10 seconds. This is a common scenario problem we've all got into where you're rounding on patients and a patient who just had to say, hey, can you write me a letter for work or an excuse for my plane ticket? And you're thinking, oh, I have 20 more patients to see. I really want to do this, but it used to just take so much time. And now this is so easy to do. So I just say, yes, absolutely. Let me take care of that. You could either go back to your computer or if you're rounding with any kind of laptop or device, you can have that right up on there with the two tabs, like I had mentioned. You're just going to input your progress note, ask it to draft a formal work excuse letter for the following patient. And then you get a nice letter like this and it'll have your name on it. And one thing I really like about the Doximity GPT is that it has your name built in, it'll have your phone number, your office information, your fax. And as you guys probably know, I'm not sure if you use this or not, but there's also a really nice fax, e-fax and e-sign on Doximity, which is incredibly helpful. So you can actually create the documents in Doximity GPT and then fax them and sign them and do everything right from there, just from your phone. So it's an incredible feature that I really enjoy. All right, so this one, you could really get complicated with, but this is just peer-to-peer bullet points for a patient in inpatient rehab. And how I would do this is, I would look at what clinical information I have and put it into the GPT, but you could really use the pre-admission screen, the HNP, any information you can get your hands on. This is a 60-year-old stroke patient who was denied inpatient rehab. He had a right-sided weakness, expressive aphasia, he needed three disciplines. And your, once again, scenario where you're rounding on the floors and you go, hey, can you do this peer-to-peer in the next hour? We need to get this patient in or we need to figure out what's going on. So it pulls you away from your rounding schedule and you don't want to spend hours trying to figure this out or digging through the chart. You really don't even have time for that. So just having easy access to that pre-admission screen, being able to move it over into your HIPAA-compliant GPT and get some quick points. Obviously, you can get way more complex than this, but I just listed out some basic ones for sake of this talk. So we know now left MCA stroke, there are the deficits, hemiparesis aphasia, patient's max assist, needs OT and speech. They are unable to safely return home, but they do have family support, but not 24-7. So just some basic talking points there. This is how I would put it into GPT. I would ask it to summarize stroke patients' HMP into bullet points for a peer-to-peer discussion, emphasizing the need for inpatient rehab. You could really get clever with this if you wanted to determine whether the patient they thought was more appropriate for a SNF or home, you could say, you know, summarize this patient's HMP and help me determine the disposition, inpatient rehab, skilled nursing or home, and list out pros and cons of each or something like that. You can get as complex as you want. You just gotta kind of learn the prompts that you wanna use. And I have several that I use every day. So it would pop out, like I said, something like this. The patient requires three therapies. They can tolerate three hours a day. They're unsafe to discharge home. They have cognitive deficits impacting safety and follow-through. IRF admission is required. And I really like, so they have the delay in rehab increases risk for complications and long-term disability. Just some short talking points that you could bring up during those peer-to-peers. All right, so let's say that that same patient got denied the peer-to-peer, so they got denied rehab. Then they were going to go to the appeal process. You can actually put this right below that on your thread because Vaximedi will save your threads on the left and say, okay, the peer-to-peer was denied. Can you create an appeal letter to submit for this patient? And then it just pops out a nice letter like this, just summarizing everything that we just talked about. And they added this line at the bottom. I respectfully request that this denial be reconsidered to allow the patient to access the standard of care supported by CMS guidelines and stroke rehab best practices. So some nice terminology, very professional, they're very well done and very quick to do with zero burden on the physician. You can go back to rounding, you just get this printed off and handed to the case managers. All right, so those are just some simple ways that I use AI. It's obviously not perfect. You always have to edit things. You always have to look and make sure that there's accuracy, the tone sounds right, it fits regulatory standards. It's not a replacement for a physician, but it is something that we can use in adjunct to make us more efficient and help us learn. It's obviously not an excuse for poor documentation. But it does have a measurable impact. There's faster approvals, more time with patients, less documentation fatigue. I mean, that's the big thing is you can spend more time with your patients, doing the things that you like to do and really adding value. Instead of spending that extra time coming up with a work note, you can call the family and tell them how their loved one is doing and what their prognosis is and really do the things that bring joy to the part of being a physiatrist. And the way that I think about AI is really, if you go into a garage and you're a mechanic, there's a computer you can plug in to see what's wrong with your car. So if I just go into the garage and I plug it in, it may tell me that a part of the car is broken, the clutch or something like that. And I can say, okay, and it may even tell me how to do it. But I'm not a mechanic. I don't know how to take the car apart. I don't have the tools to do that. I don't know how to assess what else is wrong. So with this stuff, we're really the mechanic. And we're using AI to tell us what's wrong and help guide us through this. But if an average person just uses this, they don't have those same tools. Like I can't go and fix my car just by something telling me what's wrong. So very similar to that analogy I'd like to use that I think makes sense to a lot of people. And that once again goes to using AI as a support, not a decision maker. Don't blindly submit output. Use your clinical knowledge, use 14 years or plus of education to be able to use this tool when it really, really increases your efficiency and also your knowledge base. Make sure you review everything before you submit. So how do you get started? I mean, we're all just trying to figure this out, right? So sign up for Doximity GPT, sign up for open evidence, prep with common denial types, build and share templates with each other, find out what prompts work for you and practice frequently. I started off by just trying to use some AI every day and now I use it every single day more and more and more. And it's been an incredible resource and I'm really excited for where it's going and how it's going to help our field. I say, don't wait for the system to change, change how we work within it. Physicians who are able to use AI are going to be so much more efficient at delivering healthcare. And that's all I got. Thank you so much. Very, very practical advice and I love that. A couple of questions for you. One is that, have you noticed hallucinations and how many of the notes do you have to review and how do you get around that? Yeah, the biggest issue with the hallucinations, or issue that I've had is you have to make sure that you're resetting those fields when you're typing in information. So if you have a thread going, instead of just pasting information below that, you need to create a new conversation. Otherwise, sometimes it'll misgender the patient or it'll take information from a different paragraph and it'll think that that's the same patient case. That's the biggest issue that I've seen. Always reviewing stuff and you trying to avoid acronyms if you can, because sometimes it will choose the wrong acronym. For example, at UW, it may choose Washington instead of Wisconsin, depending geographically where you live or just little things like that, that can cause errors. I agree. And actually what I've done is that if there's a prompt that I create that I'm having issues with, I'll ask GPT directly, this is a prompt that I'm using, would you recommend a better prompt? So then it'll tell me what prompt to use instead. And so basically you can have it learn or basically teach you. Other things are that a lot of times I'll ask it, are you sure? And then it'll go back, review everything and say, you know what? Actually, I'm not sure. I think I made a mistake and this is what it is. But I love the idea that you have about resetting it because it has a memory. So the memory goes back to previous conversations and brings that up. And that makes it challenging. But what I do is I basically, like you said, I basically put a new prompt in, a fresh prompt, and that basically makes it fresh and clean. But I'm sure in the future it's going to get better with that. Great. Anybody else have any questions or comments? Obviously we can do the end as well. There's one hand raised, I believe. Oh, yeah. Maybe I missed it. Yeah. Dr. Hassan, please go ahead. Hey, everyone. I have the incredible pleasure of working with Dr. Tariq and Dr. Cowling at Madrina. Much newer physician. I've been doing it for about two years and exclusively in nursing homes and a physiatry consultant. And it's been an incredible experience, very fulfilling. The thing that I learned quickly at the nursing home is there's an incredible amount of data and you're not going to get the data from the patient or potentially the nurse. It's a lot to kind of aggregate every day. And to do an excellent job every day for every patient, especially with the volume that we're expected to see, you kind of have to be a superhero or use AI, which turns you into a superhero. So, you know, I've been a technologist for a couple of years now. I had a digital telehealth company and now I'm actually beginning a startup from scratch. I'll tell you, we can talk offline about it. I'm not going to make this into an infomercial, but it's roundsmarter.com. So you can take a look at it. My contact information is there, but just a 30 second summary. A little bit about, you know, like kind of how we're using AI here. You know, like obviously we know the power of dumping documentation into a chat and then having that serve as the context for, you know, like whatever you want it to do. So I, you know, the initial stage of this was like, you know, putting the medical records into chat GPT and saying, Hey, this is the patient that I'm dealing with. This is, you know, like my physiatry type, you know, like plan, you know, this is what I want to generate. And so essentially what I did with round smarter is I was able to upload all of the data from the largest market share EMR in nursing homes, which is point click care. And I was able to kind of standardize the report template. And, you know, I migrated into my HIPAA compliant data Lake, which is just a database, a series of interconnected tables, connect that with our therapy database, which is a net health. If you've had any experience with that. So now you have, you know, the patient's functional goals, their progress every day, you have everything that's changing in the EMR, right? Any new medications, diagnoses, labs, radiology. So now, you know, when I'm, you know, like custom coding the AI pipeline for around smarter, you know, it's, it's, you might've heard this term called rags, you know, retrieval augmented generation. So now I'm, you know, for any patient context for that specific data service, I'm pulling in all of the relevant clinicals, you know, that I want to know about when I round in addition to, you know, it's not a bridge or Sully.ai, which are like the two market leaders. And I think Sully raised like $250 million, like a bridge, you know, $30 million in the last round. Right. So this is bootstrapped AI, you know, like AI company. So what I'm trying to learn now is where is that, you know, like that, that edge of like hallucination, right? Like how much can you expect AI to do with, you know, with like medical grade fidelity, right? And what I've learned is you can't just dump massive amounts of data, right? You need to kind of like structure the pipeline where AI will give you an output. You know, there's deterministic coding, right? Which is Python. And, you know, like the, you know, like the old programming languages, the advantage of using that in parts, I mean, if you guys ever work on a project like this, just remember, you know, like you want to deterministically code anything that's possible. For example, like when I upload all of my medications, we all know that their FDA publishes these datasets. These are all of the medications that are renally cleared, hepatically cleared, right? These are the drug interactions. Why rely on AI every time, like on execution to try and figure that out, right? Where maybe it has a 5% chance of hallucination. So now I'm labeling all of the data as it goes into the cloud so that when it retrieves it in the rags, now like it's going to already have the metadata attached to it. So when it's, you know, producing the documentation, you know, like I don't have to rely on that to get it right. Like, so the advantage there is like, you know, like when I'm, you know, seeing these patients and, you know, Dr. Kelling and Dr. Drake have a thousand times more experience than me, but, you know, they're on like 20, 30 medications, right? So every day am I checking, you know, like their patient's creatinine and their kidney function and like looking if this is appropriate for their GFR. So really like the way that I've kind of designed the pipeline is now it's going to stand the diagnosis list for any ICD-10 code that's within the like kidney failure range or like within the hepatic issue range. And then it's going to, the second part of the pipeline is going to look at all of the medications across the FDA database and identify the renally cleared and hepatically cleared medicines. And then when it generates my plan, it's going to actually have a physiatric, you know, functional relevance of those medications and their side effects that we're monitoring. I actually learn more about the patient after I generate the plan. So. Yeah. Thank you. I appreciate that. That's really important because, you know, your perspective as a technologist and a physiatrist, that's super unique. You know, that doesn't happen. Usually technologists come in to tell us what to do and how to think, because we know, like we look at the holistic approach of the patient and that's not necessarily how a lot of this is set up. So I appreciate your comments there. So next we'll have Matthew Allen, if you don't mind sharing your screen and please go in through your slides and just do a quick intro. Totally. And I'll be quick because I want there to be time for lots of discussion. I am a fourth year medical student at UC San Diego School of Medicine. So I'm going to talk a little bit how I use this as a medical student and as well about how I think this will transform research, medical research. So I want all of you, you can put it in the chat if you want, but think of a clinical question that you wish you knew the answer to. You wish there was evidence to answer this clinical question, but there's not. I just kind of have that in mind during my presentation. So sometimes we don't know the answer to clinical questions because it's really hard to keep up on the evidence. You know, it's rapidly, rapidly progressing. I don't know if you guys remember doing flashcards, memorizing things in medical school. I think at this point we've realized we're not going to be able to remember everything or keep up with everything. Sources of truth, like up to date, oftentimes, you know, they haven't updated it in a while or there isn't an article on the question that I want answered. So sometimes we need to do more research, but that can be time consuming to conduct. So this all goes to say that research takes too long to impact care because we might not know about that research. And then we have all this data. We're producing mounds and mounds of medical data, but we oftentimes don't analyze or gain insight from it because it is so time consuming. So we've already talked a lot about open evidence, so I'm not really going to go into it too much. I do use it a lot as a medical student. It's incredibly useful. And what I love about it is that it points to the medical literature. You know, you're not getting internet links, you're getting journal articles. It's highlighting what's new research. It's highlighting what comes from high impact journals. And so it keeps you very based in the medical literature. But what I kind of want to focus on today is how will AI accelerate how we do research on all the data that we are producing. So I'm going to give you guys an example from a tool called Mediloop. Mediloop makes use of what is called agentic AI. So AI agents means like a lot of things we've talked about today are large language models that will answer our questions. But you can train large language models to do different tasks. You could train a model to be a statistician. You could train a model to be an expert in, you know, research design. You could train a model to be an expert in research ethics. In, you know, writing an IRB. You know, there's a million different AI agents that could be created that can augment our ability to engage in medical research. So think about your question. Like, let's say you typed it in. So here's an example question. How many patients who experienced a stroke received rehabilitation therapy within 90 days following the stroke? So the first AI agent is going to take this question and then turn it into a more fleshed out research question. So in this case, it's going to define stroke as cerebral infarction and it's going to define rehabilitation therapy as OT or PT within 90 days following the infarction. So at this point, there's human intervention. You know, the first AI agent has performed its role. It's translated your question into a more nitty gritty research question. And I could have, I could change things here. You know, I could say, actually, I also want you to include speech therapy in this definition, except move to the next step. The next agent will come up with a data extraction plan, right? It's going to identify ICD-10 codes or, you know, it will be familiar with the database that you're using and it will come up with a data extraction plan that you can then accept or edit. It'll come up with a statistical analysis plan, right? You can make changes. You could even converse back and forth about the advantages or disadvantages of a particular statistical plan, move on to the next step. And so as you can see, this is kind of what's been presented today. It's AI working in concert with physicians to augment their productivity. So, you know, at the end, it can give you a visual distribution. Like normally a research project like this could take months. You know, part of what could make it take a lot of time is access to PHI is very, very protected, as we all know. And an advantage of agentic AI is there are ways to let AI agents into sensitive data sets and not let humans in that are safer. And so the AI might be able to get access to the data set that you're not able to get access to. And then it can gain insights from that data and present them to you. And what this will do is rather than taking months to answer an initial question that's maybe a minimally impactful question, we can iterate, right? Now that we've answered question one in five minutes instead of three months, we can move on to, you know, how does this change if we change the cutoff? How does this change depending on insurance factors? You know, it will suggest future research questions for you and you can iterate on the research so much faster. Another advantage besides iterating is that it's reproducible, right? One issue with research is that most of it is new and very little of it is confirming prior findings. If you lower the threshold to analyze data using agentic AI, that means me as the next person, if I see a paper that I'm skeptical of, I could attempt reproduction of that paper, you know, using agentic AI. And rather than that taking me six months and me having to, you know, cut my work back half time, I could reproduce their results that same day and see if I come to those same conclusions. And so I really think this is going to exponentially increase the speed of biomedical research and it'll make it so that we have more insight rather than just tons and tons of data. One word of caution, you know, garbage in, garbage out. These tools are incredibly powerful but we do have to be thoughtful about the data we're feeding them because the best AI system will not be able to give us good insight if the data quality is poor. And expertise, I think this has been a theme of this whole talk, expertise is more essential than ever because it's going to become harder and harder to pick out the issues in AI outputs because they'll be very sophisticated, very convincing, and so really that puts more responsibility on us to use our expertise. Like has been said before, comply with your institutional rules, but don't be afraid of this. We can't fight it. If we try to, we will fall behind. You know, if we as PM&R physicians don't embrace this, other specialties will and our specialty will not have as much impact. You know, if we as individuals don't adopt it, others will and we may have less impact. So those are my thoughts on how AI will impact biomedical research. Feel free to check out Mediloop. There's other similar tools. I don't have a financial relationship to them. I just think they're cutting edge and cool. So thank you. Thank you, Matt. Matt, I want to just ask a question, I think, just because you're the only medical student here as far as I know, but how much is GPT usage common in medical students at your institution and you know places? It's extremely common yeah I mean I would say almost all my classmates use it we use it to answer clinical questions you know throughout a day on the wards you can also use it and say like give me five practice questions on this concept I'm having a hard time understanding and it will do that to you or for you or you know here's my paragraph of my presentation for my attending how could I improve this presentation what things should I tweak about my plan for this patient I think us medical students have to be cautious because we don't have the same expertise as you more experienced physicians but that is not stopping us from from using these tools very very heavily. And are you noticing that your attendings that you work with are not using AI as much or they're adapting to it slowly? My anecdotal perception is more medical students use it than attendings but quite a few attendings use it and often if they notice you using the tool they will be interested in it and we'll try it out and then the next week say wow you know that open evidence actually cites really good literature and I found a couple cool papers through it you know. Right and the reason I'm asking is because you know I mean we started this community just recently and one of the things that I've talked to my other colleagues other specialties urology orthopedic surgery you know cardiology you name it they're so far ahead of the stuff they have like you know they have startups they have basically like innovation communities they have innovation leaders so you know I feel like PM&R is really falling behind it's nice that you know someone like you as a medical student is already thinking about this and we have people in different levels of expertise here but this is such a significant fear that I have that we're gonna fall behind and other especially you said neurology and just general internal medicine will grab the stuff in this information and move ahead so yeah I I agree with you and I think our specialty needs it even more than others in some ways I I'm doing an away rotation at the University of Utah right now and they had a research symposium today talking about how a lot of our outcomes in PM&R are soft in the sense that they're functional right there how is this person able to navigate the world how has their function and pain been improved you know we don't have a hemoglobin a1c and so we need sophisticated data analysis methods to prove the value of what we do and to measure our outcomes and so really we should be like so relieved AI is here because now from a patient free text talking about how they've been for the past few months since their spasticity injection you know we could gain insights from that from large language models on a large scale that we couldn't have before so we need this more than other specialties in some ways I think I agree I agree and you know with me I'm definitely obviously years ahead of you but you know I work from you know as a provider chief clinical innovation officer for you know more than 1,200 positions and then I'm also working with insurance side of a Medicare Advantage medical director and I work for tech companies I kind of my fingers a lot of different things but without a doubt talking to hospitals and payers and even you know directly to CMS you know we have to show our value we always talk about that the only thing to value is certain metrics length of stay or total cost of care readmission back to the hospital you know obviously avoiding surgery you know you know all the utilization of different resources we have you have it so one example the doctor calling myself are involved with that her husband as well we do telemedicine for example for P&O patients so it wasn't just a matter of like doing telemedicine and getting access to patients but what is the exact metric that comes along with that and that basically was reduction in time to getting a prosthesis which results in reduction of falls which results in a reduction of ER visits and that's a direct impact to patient care not just the telemedicine part of it so we'll talk about more examples now so I'll throw in my slides hopefully we'll finish in about 10 minutes only 10 10 slides obviously if anyone has any questions about these topics please please feel free to bring them up now okay hopefully you guys can see it all right so my name is Dr. Tariq I am the chair of the innovation AI community brand new and I'm so glad to see 30 people on the call that's amazing especially the first call we have at least one talk at the annual conference this year and you know hopefully we get to meet as well and I think hopefully we'll keep this going maybe make it a every other month or quarterly kind of meeting in which we discuss what we've learned and how we're implementing AI into our practices I'm the chief clinical innovation officer Madrina and no involvement disclosures so for me I should probably minimize this okay so why does it matter to physiatry to innovation is the complexity of care there's a lot of care that is being delivered in and there's a lot of transitions that are happening between acute hospital to LTAC to sniff to home and back and forth and you know how do you actually manage that where the majority of the cost of care is happening there's obviously a data explosion with EMRs and different AI models now and wearables and how does that all fit into the PM in our practice because a lot of times other specialties are they see grabbing onto that and making their own where we should be the the ones kind of talking about the function and the sports performance and brain injury and you name it how do you actually you know make meaningful impacts like reducing readmissions and how do you make personalized care you know how do you basically look at a complex patient because all of our patients are very complex of multiple comorbidities so many different dynamics happening then how do you actually personalize their care how do you actually give the right medication at the right time and the right DME and right injection all the different things to talk about and obviously overarching all this is value-based care how to how do we physiatry fit into value-based care and how are we supposed to show our outcomes so what is clinical decision support you know basically it's using software to aid us in deciding you know about patient care basically if the patient needs more care less care and we see the overarching behind that is rule-based alerts that could be as simple as someone's blood pressure is elevated it could be more complex like a trend of lab that is actually going up or you know for example I'll give you example the nursing home setting you know you think about 98.7 it's a normal temperature but actually the average nursing home patients temperature is 98 point I think one or something and that's lower than average obviously older patients and there's other reasons behind that but a typical nurse and nursing home might not think of a 98.7 or 98.9 or even 99 as a temperature to be even concerned about but we know considering the baseline how much elevated that is one degree is significant for these patients and that's how you can actually use certain models to see a this person might have a UTI I might have an aspiration pneumonia might be developing a wound that's like one example of using rule-based alerts or it could be things like declining function someone's progressing a certain way it's a let's say a stroke patient orthopedic patient all of a sudden you notice that you're kind of plateauing what is the reason behind that it could be pain it could be some other nerve issue and basically that's things you can do to create alerts then you have predictive models you can have predictions of based on the patient's diagnosis medications demographics like what is the estimated length of stay in a nursing home right now we're basically basing length of stay in a nursing home or acute rehab or even outpatient or number of therapy sessions they get or the type of procedure they get just based on like art of you know basically medicine or basically based on my my previous experience but you can actually use certain models which we use in our practice to actually predict someone's outcomes obviously you still have to use clinical judgment there you have to have that part of it but you're actually being able to predict if someone is hospice appropriate or palliative appropriate or what is the estimated length of stay in the facility and if they extend the length of stay what was the reason behind that and obviously the model keeps on learning but we have certain ways to do that and obviously have different type of diagnostic tools as well so how does physiatry fit into this we have you know things that can help out with discharge planning with therapy targeted therapy and basically outcome prediction and then using those models to actually help us justify our value you know did I make an impact as a physiatrist in changing the outcome of a patient that was probably going to go to a nursing home but now they're going from the inpatient to home which is the best place to be so one of the box on the models we use are basically of seeing if someone is appropriate for sniff avoidance can actually skip the sniff and go straight home like a sniff at home model or who is more likely to have a wound based on their mobility based on their nutrition status and those kind of things and how do we then prevent that from happening versus you know the model of basically like you know not preventative but reactive care so one of the things that we do we use is a care port which is by a company well sky was obviously a cure for BC is a there's many different softwares like this out there but it basically connects with epic and Cerner and meditech and eclinical works and all the different EMRs for home health and for sniff like PCC so they basically connect with everything and also connect with HIEs not saying they have like 100% connection but they have a significant in a lot of markets and what they do is with a large amount of data they have with every single readmission hospital visit they're able to basically use their models to figure out you know what's the best rehab setting for a patient the patient and then obviously identify patients that have a high risk for readmission based on the number of ER visits they had the number of sniff visits they had length of stay in each facility so that is all like you're able to use that information even before the transition of patient you're able to max the intensity of rehab so are they able to tolerate three hour therapy a day or is it one hour or is it more outpatient so be able to kind of use that and it's used very frequently and inpatient rehabs and sniffs and hospitals majority of time softwares like this are used for the referral management side of the facility or the hospital where the social worker discharge planner might be using it to figure out which facility a patient needs to go to but as it is advancing you know a lot of the inpatient rehab hospitals and post-acute networks are able to narrow out the post-acute network into facilities that have the best outcomes versus like you're gonna go to the nursing home that has the nicest Facebook page or has the best reviews you're focusing more on like what has the best outcomes for which diagnosis do they have a dialysis center we can actually have you know a end-stage relent disease patient do they have a pulmonologist that you can at a nursing home or inpatient where you can send a patient for as you know COPD exacerbation so those are some of the models you're using and you know my own practice what I use care port for is that care port assigns a patient a readmission score or readmission risk and then I use that risk to just leave have much myself and my team focus more on those patients and try to dig deeper with using AI models to see what is what am I missing you know what is the issue here and then you know creating the plan accordingly so not you know having the family you know discussion early on about a complicated discharge versus like you know let's see how to do in therapy the old model does not work other stuff that we use again I'm only talking about a few items here a few things is so many other things layers and layers of innovation that we do is basically monitoring patients again proactive care versus reactive care one of the things that we use quite a lot is a company called circadia they're based out of England and in the LA area but they created a contactless monitor and this uses ultra wideband ray basically frequency which is Wi-Fi waves basically and is this a device that you know you see shoots out these waves and there are 6,000 per second to high number of ways that are being shot at to the to the patient but these waves don't penetrate water so they hit the body which is composed of mostly water and reflects back and what is basically measuring is an FD approved for this for a respiration and heart rate and they did this over like you know compared to EKG and all the stuff that to go through for FDA but what it's doing is basically looking at the change in the fluctuation in your respiration and as you're looking at your pulses a carotid process pulses and heart rate and is giving you like life feedback on someone's heart rate without touching a patient and that's the beauty of this so the patient is in the bed they can check that you can even have a duvet on top it could be even behind wall you can go through walls too so this is the future where you're gonna have not just wearables and some wearables you need those too but having contactless monitoring in different rooms or different places in the in the facility and then you also have some wearables you have to use for example blood pressure you obviously have some you know wearable blood pressure cuffless blood pressure cuffs now you have rings you have different wearables you can actually add on there's patches that you can visit patches that we tried in a pilot that can actually look you can listen to someone's heart rate and listen to their lung lung sounds it's basically you know get the auscultation for that remotely and which is pretty amazing so these things are you know basically being invented and they're getting FDA approved but how do they fit into the PM&R practice is all the challenge you know is this data gonna cause alert fatigue and what is the data gonna be used for necessarily for the PM&R perspective so for us using this you know during COVID time as well it was amazing because you can actually catch COVID you know or at least the symptoms of you know certain conditions including UTI and pneumonia COVID as well at 72 hours before and how it's doing that is basically creates a baseline you know for the patient the respiration is 17 breaths a minute the heart rate might be you know 80 at rest and a subtle change which is not gonna be noticed by a nursing staff at rest is significant enough for us to kind of start digging where then at that point you ask the patient what's going on yeah I don't feel well today you order an x-ray you order a COVID test and hence you catch things and hence you prevent the readmission so very proactive care and obviously non-intrusive and it's which is amazing about it so we have dashboards where we can actually look at hundreds of patients at a time the challenge with this is you know basically as value-based care kind of keeps on expanding is a reimbursement for that you know and how does that fit into the inpatient setting and how does it fit into outpatient and alert fatigue and other stuff that goes along with that privacy and on and on other stuff to talk about is hinge health for example there's other obviously competitors for this as well including sword and forgetting one more name sorry but some of the leadership actually in these groups it happens to be human are which is great but at the same time like these companies are massive now massive multi multi-billion dollar companies I think hinge became publicly traded recently for quite a lot of money so what they do is they predict surgical versus conservative treatment based on in the millions of points of outcomes they have they can track pain they can track function adherence overall reducing unnecessary imaging and specialty referrals streamline care for back knee shoulder and pain you know basically I just kind of threw a lot of stuff at you guys but how is it doing that is basically using some models including the the picture you see there where you have a phone in front of it the phone creates certain dots it can tell you tell you how much how you're moving your gear get up and go test your you know obviously range of motion and then based on that is giving you certain rehab exercises and then the millions of points of data they have and then obviously knowing how someone does they can analyze you know if someone's range of motion is certain amount or some of the demographic is certain amount if they might need a specialist to see them so these things that were already happening and if you're not on the table you're on the menu so we have to be involved either we start this stuff or we actually get involved with these things because MSK care is really our bread and butter but other people have taken it and they run with it but we definitely have to be involved with them other other thing is sort I mean talked about AI and motion sensors for PT you know obviously there's different codes that see a CMS has as well for being to be reimbursed for this telemedicine being the most basic one but there's another one called remote therapy monitoring or RTM that you can actually get reimbursed as a physician for basically giving your patient therapy you know exercises and then they do those and then you basically you know basically have a staff member that collects information and response back to them and so anyway things like you know cure AI which was AI imaging for spine and joints is a lot of innovation happening in radiology and predicting things and catching things even before radiologists notable help for smart documentation example of me I saw and that health and doctor now these are a couple of EMRs that are more specific to wound care where you can actually use a lot of tracking and predictive models to see if a certain wound is gonna get better with a certain amount of graft or debridement or what what the what the need might be so AI is going to penetrate into every single part of our practice now it's a matter of like how are we gonna grab on to some of these things and you know fit them into PM and ours and then also tag along so because eventually like some of these a lot of these models will will be part of the value-based care and then if we're not the ones using this they're not gonna care about PM and ours role in this so overall you know inpatient you know care port subacute also care for plus circadia outpatient you can do in your spine zone and other things are cure AI and notably and again I have many more examples but I wanted to keep it just brief for now so benefits for this a lot of things that I like to do is anticipate decline enhances coordination it personalizes the care they need eases documentation a lot of times well so you know obviously we talked about ambient AI the challenges are integration issues the big one is alert fatigue data privacy without a doubt is a big issue and then you know trusting AI do we actually trust it I feel like I'm at a point now that I don't have necessarily blind faith in it I still find issues but I definitely certainly trusted more than me in certain tasks because I'll miss certain things and it's not going to and then it's always like oh yeah I didn't think about that but I definitely still feel like I need to be the captain and that needs to be the co-pilot so what can you do you can think about using we always talk about a lot of different things here there's certain free tools obviously your institution might have certain things as well but you might be already using a lot of it but don't just limit yourself with the low-hanging fruit of ambient AI I think the ambient AI people ask me all the time it's like what's the low-hanging fruit I'm like ambient AI it's the most simplest and actually it helps it has one of the biggest ROI because all of a sudden documentation is better but that don't stop at that you know we talked about agenda yeah we talked about research we talked about obviously using the proximity but that's just the beginning this is the beginning layers of where we can be I think using different models of how you improve care how do you present at the value-based care how you present at the hospital how do you end up getting all the consoles as a physiatrist like those are things you have to use AI for you know work obviously focus on functional outcomes big part of what we do is physiatrist work with IT and admin the best readability try to do a pilot if you can demand we have specific metrics that's something that a lot of AI companies lack and keep patient-centered and that note I am done

AI in physiatry

clinical decision support

predictive models

contactless monitoring

AI-driven PT assessment

patient outcomes

ambient AI

Mediloop

Circadia

human touch

innovation communities