Welcome to AASM committee's present, the nuts and bolts of AI Sleep Center Integration. Brought to you by the American Academy of Sleep Medicine and the Artificial Intelligence and Sleep Medicine Committee. I'm Matt Anastasi, Sleep Team Resources Manager at the Academy and staff liaison for the committee, and I'm excited to kick things off with you today. AI just isn't the future of sleep medicine. It's happening right now. From AI powered auto scoring to clinical workflow tools like AI scribes and personalized mass bidding, the possibilities are endless. Yet making the leap from exciting tech to real world impact, that's where many sleep centers hit a wall. Today's webinar is your blueprint for breaking through. You'll get an inside look at how AI is already transforming clinical practice, learn practical strategies for responsible adoption, and discover how AI can fuel your sleep center's growth, straight from the pioneers leading the charge. Get ready for insights, strategies, and a glimpse into the future because AI in sleep medicine isn't coming someday. It's here and ready for you. Sleep center integration is a continuation of our committee's panel presentation series that began with basic terminology, which was followed by AI beyond the PSG, privacy and security, and our December webinar on the role of AI in sleep health. All four are available free on the ASM.org website under the standards and guidelines header. Just go to the clinical practice guidelines dropdown and click on the AI section highlighted on this slide. Before we kick things off, I'll remind you that your audio is muted and your video and chat are both off. We encourage you to include any questions as we go along in the Q&A section of the Zoom platform because we can see them live there. If we do not have time to answer every question in this hour, the remaining questions in the Q&A will be shared with panelists for follow-up afterwards. Also, you're free to reach out to us at contact at ASM.org or call the number on this screen. So while developments around AI are constantly changing, our panelists will do their best to provide you with the most up-to-date insight. Please keep in mind that their opinions do not necessarily reflect the positions of the academy and their input is intended for educational purposes only. While some of the information in this webinar will address legal issues, it's not intended to be legal advice or substitute for the advice of your own counsel. Anyone seeking specific legal advice or assistance should retain an attorney. Now, allow me to begin by introducing our moderator for today. Dr. Atif Hussain is a Professor of Neurology at Duke University Medical Center and Chief of the Division of Epilepsy, Sleep and Clinical Neurophysiology. And he'll start by introducing our five panelists. Dr. Hussain, the floor is yours. Thanks very much, Matt. It's a real pleasure to be hosting today's webinar and I'm delighted that we have an excellent group of experts who will be talking about integration of AI in clinical sleep practice. I'll start off by introducing all five. Our first is Ms. Amber Hess, who supervises the Neurodiagnostics and Sleep Center at Sanford Health in Bismarck, North Dakota. She also participates in patient care. Amber sits on the Abbrett Volunteer Leadership and Development Committee, is the Education Chair for CSET, and is the President-Elect for the Montana Sleep Society. Our next speaker will be Dr. Soubella Zia, who is a sleep pulmonary and critical care physician, founder and CEO of a Silicon Valley-based digital healthcare company called Telemedora. Her interests include AI, telehealth, and healthcare innovation. Next will be Dr. Nancy Collip, who is the Director of the Emory Sleep Disorder Center and a Professor of Medicine and Neurology in the Emory School of Medicine. She has served on the ASM Board of Directors, including as President. She was Editor-in-Chief of the Journal of Clinical Sleep Medicine until very recently, and her term was from 2015 to 2024. And she is the Section Editor of Sleep Medicine for UpToDate. Next will be Dr. Tim Lee, who is an Assistant Professor at the University of South Florida in Industrial and Management Systems Engineering. His research interests encompass health informatics with a focus on connected healthcare systems and prognostic analytics. And finally, Dr. Ramesh Sachdeva is a practicing sleep and pediatric critical care physician who has served in the national role as Medical Director for Quality Initiatives for the American Academy of Pediatrics. He has served as faculty in the Center for Ethics, Medicine and Public Policy at the Baylor College of Medicine and as Adjunct Faculty in Law at Marquette University Law School. So with that, we will get started with Amber. So Amber, I'll ask you a couple of questions and it'd be wonderful if we could get your take on these. We'll start with the impacts of AI scoring in your center. So we've been using AI for quite a while now. When we started, we were a four-bed sleep lab running seven nights a week, as well as doing 30 to 40 home sleep studies a month. And I scored every single study that came through our lab. So very time-consuming on my end, as well as also running the departments and all of that. Since we've implemented AI, we have actually been able to increase to a six-bed sleep lab that runs seven nights a week, doing about 100 home sleep studies a month. We've had no impact as far as turnover time. We still keep it between 24 and 48 hours before it's in the provider's hand. That's fantastic. That's quite some growth. Can you demonstrate a scoring workflow from an upload, download, event review to report generation? Yes, I can. So let me grab that here. There we go. So I'm gonna show you on the home sleep end because on the in-lab side, nothing changes at all. Once my techs upload their studies to our server, AI works in the background. It comes back to us within seven minutes, no change from my provider's reports, anything. Home sleep, we do through the workflow or through EnsoData's database page. So this is what it would look like when they come up. We can go in and put in the tech who did the test, the provider who's going to read. And then, gotta move in a sec. We're gonna click here and we're gonna pop in. I just threw some info in here. And we'll come to this fun little screen after it stops squiggling there. So at the top of my page, this will clear up in a second, you will get your hypnogram. You do get wake and sleep on these, which is great because they do have actigraphy. Because we are not using brainwaves to look at this, whoop, I gotta move stuff. There we go. I do like to look at a two-minute screen because then you get a better idea of your respiratory. But it's very simple to use as far as I can just click. It's just scrolling through for me here. It's just gonna run through. But you would put in like you would any event. So you would highlight it. You can change it by right mouse clicking. It's very straightforward, which is wonderful. Once we are all done scoring, you get all the way to the end, there is a generate report button and you get a beautiful report. One great thing working with them too, they can make this report show anything you want it to show. So I do get AHI, I get oxygen DSATs. You get all of the information you're used to seeing. My providers find it very easy to use because they can actually select their own name from our database page, where it says up here, it says my studies and that will list the studies they need to read. They don't have to worry about sorting through a bunch of them. And that's fantastic. Thank you for sharing that. Was that the AI algorithm that we saw running and you were manually scoring on top of that? So the AI algorithm had run and then I'd go back through just to verify everything was good. It takes just, on a home sleep study, less than 10 minutes to get through it. Fantastic. How did you get the buy-in from your department leadership for staff and how did you demonstrate the return on investment? So when we started looking at this, it was a very big thing that we wanted to be able to do more. We were booking out very far, as far as our home sleeps and our in-labs, but I didn't have the time availability to be able to add any more. I was spending four to five hours a day just scoring. So stuff was to my doctors. So when we proved we could double the amount of home sleeps and add those two more beds without it impacting our output as far as people getting their results, doctors getting their studies, it was very easy for them to really buy into it. Staff as well, because nothing changed on their end, as far as what they were seeing, they were just fine to go forward with it. Okay, so we have time for one question. We have a lot of panelists to get to and then at the end, we're gonna have a freestyle discussion and get to as many of the rest. The very first question is a question that I'm sure a lot of labs are gonna have this for you, Amber. Are you able to get a report for both 3% and 4% or do you need to get separate reports is the question? You can get just a single report and they will build that for you. So if you want it to say, this is my 3% AHI, this is my 4% AHI, they will build that into your report for you. Thank you very much. Let me move on here. Well, great. With that, we'll move on to our next panelist. Dr. Sobella Zia is gonna talk to us about integrating AI and the business of sleep medicine and promoting one's practice. Thank you, Dr. Zia for joining. How has AI helped with business development in your practice? So, first of all, thank you, Atif and AASM for having me as a panelist for this webinar. By adopting a patient-centric AI-first approach, I've managed to keep business operations lean because AI has allowed me to reimagine workflows. So let me share some examples of how AI has helped me in business development. To number one, I use an AI-powered tool to enhance the cybersecurity. I like it because it allows for 24-7 vulnerability scanning without any downtime for maintenance and is cost-effective. For marketing, Gen AI has become an essential part of our brainstorming sessions, from generating marketing assets to finding SEO-rich hashtags. From social media management to website development and maintenance, I've found it quite time-efficient. And of note, it did not eliminate any roles. On the legal side, it allows me to understand contracts better and helps me to negotiate them better. Unlike Dr. Sajdeva, I'm not an attorney, so I need to understand the French that the attorneys write in all these contracts. So I just ask the confusing parts to be clarified by the Gen AI, and it has reduced the number of back-and-forth emails to the legal. So he likes it. For the growth planning, it helps me with crafting a business strategy based on comparative analysis of our competitors to monitor trends in digital health and to identify new opportunities in sleep medicine. On the administrative side, I leverage AI as my executive assistant. It has written some basic internal procedures and policies for me. It has generated some training materials and videos for team members, and it does email management for me as well. And for the business meetings only, I use AI Notetaker, which not only captures the transcript, it also summarizes these meeting notes for me and generates actionable items, which can be shared with all the attendees. And it allows me to be more present in the meeting and to focus. And I'm looking forward to hear what Dr. Kolop has to say about the AI scribes on the clinical side. So speaking of the clinical aspects, how you've used AI remarkably to bring in patients, do you use AI on the clinical management of patients as well? Yes. So on the clinical side, I found AI incredibly helpful in developing educational materials for patients that are clear, engaging, and tailored to different learning levels. Without sharing the PHI, I have used it for polishing the after-visit summaries and the care plans for patients and their family members. I have used it for polishing our call scripts for the messages and for drafting letters of appeal for insurance companies when they deny it. So in a very polite, yet firm manner. So those are the things that I've done on the clinical side. And overall, I would say that AI has become an essential part of telemedora. It allows me to be more efficient, to stay afloat, to communicate better, and to focus more on patient care. That's fantastic. You know, one of the things that we hear with AI all the time, especially as a small business owner, a small practice, or a practice owner, is when AI comes in, staff are worried that they're going to be let go because you're going to replace them with AI. How have you managed that? Thank you for this question. Yes, there is a concern that the AI is going to be taking away jobs. What I tell them that we are in a transformative period and we are re-imagining workflows. So everybody, I see a role for AI in every, for every team member, from AI receptionist to AI sleep coordinators, to AI sleep medical educators. And it has only allowed them to be more efficient and to be 3X more productive. So, yeah. It's just, Atif, it's just like, you know, remember when we had the auto, when the automobile industry was launching and then the cars were coming in, the jobs of the horse repair, horse cart repair workers were taken away, but yet there was, it was the dawn of a new automobile industry, mechanics, and much, many more jobs. So we are seeing the same paradigm shift here. Yeah, thank you for that. But I'll remind you that both you and I were not around then. So it's a fantastic example, but I don't remember. Matt, any questions? Yeah, I do want to keep moving, but Dr. Zia, if you could mention what, which HST device are you using with Telemedura? There's a couple of questions about that. Sure. So there are a few types of HST devices that we are using. One is the one from WatchPad, Itamar, and the other one is Sleep Image. Those are the ones that we are using. We have used a CleaveMeds device as well, but yeah. So those are the three devices that we are using, but more and more Sleep Image and Itamar. Excellent. Thanks very much, Dr. Zia. We're going to move on and talk to Dr. Collip. We have scratched the surface of AI in the clinical space, and we'll talk to Dr. Collip about AI scribes. Dr. Collip, how do you use AI scribes in your clinic at Emory? Yeah. Well, thank you for inviting me. When Subela asked me to be on this panel, I'm like, I am the least AI sophisticated person here. But she said, well, you use AI scribe. So sure I do. And so our medical record has now embedded an AI scribe tool into it. And so it's available to anybody that uses our electronic medical record. So I use it in clinic for both new and returned patients. It pretty much records the conversation with the patient and develops a HPI and an assessment and plan. You can also use it even to dictate if I'm working with a trainee, for instance. I can just let it transcribe that conversation. In telemed, I'll have my computer do telemed, and then I use my phone to do the scribe. And it works pretty well. And the microphone's incredibly good, shockingly good. I have made the mistake of putting the phone in my pocket, and it still seems to get the conversation. That's fantastic. What do you think is the biggest value add of the AI scribe as you use it? Oh, 100% it's my documentation. I bet. I estimate I probably spend 50% less time on documentation now. The other advantage is you immediately have a note, and it keeps the note. It keeps the information for two weeks before it erases it. So if you don't get right back to your note that day, it's still there. Even though I still keep my own notes, it allows you to have kind of a note to go by. Even if you might miss a note for two or three days, you can go back, and it's still present in the patient's chart. That's great. I imagine your billing people are a lot happier since all the notes are getting closed in time. Oh, yeah. Talk a little bit about the limitations. There must be some issues with AI scribe? Yes. Well, there's a few. I mean, it's not perfect. So first of all, it's best to ask the patient if it's OK that you're using it. And sometimes that creates a little bit of discussion with the patient, so that might take a little more time. Most patients are usually fine with it once you explain it, but it does take a little bit of time. You definitely have to edit the notes. I mean, they're not perfect. The current one we use is kind of, you can tell it's made for primary care and not necessary for sleep practice. So it will list kind of all the problems you may talk about with the patient and not focus on what you're focused on. So again, getting back to the editing, it's not very trainable at this point. I think they're working on that. It does miss some things that you think are important, but it didn't think is important. The other thing is sometimes it may take what you said and put that into the patient's notes. For instance, sometimes I talk to my patients about the weather or something else, or like a wedding, and it puts that in under the patient's voice instead of your voice. And the last thing I would say is that the notes don't sound like my notes anymore. They're a little more sterile. But the idea, the concept, the reason for notes is still there, so. Thank you. Matt, any questions? Yeah, there's a couple of questions about specific software. So I assume, Dr. Kopp, you're using a scribe that's part of Epic. Is that what you said? Yeah. We used Emory Transition to Epic about 2 and 1 and 1 half, 3 years ago, time flies. And the one that is embedded is a bridge is the name of it. And so it's right on the Haiku app on my phone. So I can just click on it and start, go to work. OK, excellent. Thank you so much. Great. Now, having learned a little bit about applications in clinical practice, we'll transition and talk about use of AI in clinical research. So Dr. Li, if you can tell us a little bit about how AI can help clinicians conduct research while they're in clinic, while they're seeing patients. Tim, you're so muted. Can you hear me right now? Yes, perfect. Thank you, Dr. Hussain, for your introduction and your question. As a researcher, I think I am very excited to be here as a researcher rather than a physician. And I'm very happy to answer your question. But before answering the question, I just want to make sure we are on the same page in terms of the AIs and AI concept. As a researcher, my research focus mainly on the developing AI models as well as utilizing available AI models. So from the developing AI model perspective, I mostly focus on developed different data-driven models and also integrating the AI's algorithms into different wearable sleep monitorings from research site. So when it comes to conducting the research, I think it is very important to think about what type of AIs are there available for me, especially when I consider different tasks of research. We usually think about AIs as like chat GPTs or open AI germinase or cloudy, et cetera. So this is actually one first type of AIs we call problem-based AIs. Mostly those AIs model utilize large language models or natural language generation models in order to provide kind of prompt-based response. There are other type of AIs, including analytics-oriented AI, search and retrieval AIs, knowledge graph reasoning AIs, and also automation AIs. So those models are good at different other tasks, including extracting information, finding ranking information, linking different inference and relationship, et cetera, and also providing workflows automatically. So let's talk about what is going to help for the clinical research. Mostly each model is going to be excellent in specific tasks. For example, the first more likely going to help us to brainstorm the idea generates different hypothesis. The second type of AI is going to help us to analyze the data, performing pattern recognition of the data. For example, the PSG data, different kind of slip stage or actigraphy data, et cetera. The third one is more related to helping us to perform some very prelims literature reveal and then help us to arrange all of the state of the earth. And finally, it can help to verify the result as well. So those kind of tools are going to be able to map into different tasks. Well, thank you very much for that. I mean, one of the things that I've been very interested in is how AI can help in finding patients for clinical trials. That's always a challenge. But can AI be used for that? Definitely, yes. I've been collaborating with different sleep centers and research with different physician, focusing on clinical trials and observational studies, performing perspective and retrospective studies. So I utilize AIs in multiple ways. For example, I've been utilizing AIs in order to narrow down the electronic health records information and determine the specific patient population in epic systems or to specify the sleep phenotypes or sleep comorbidity that I'm interested in. So AIs can be able to be excellently helping me to identify those subgroups and identify them and recruit them if I want to recruit them in clinical trials. Also, it helped to generate different outcomes, simulate different scenarios so that I can be able to optimize my study protocol at the beginning as well. So it helped me to determine different sample size, different endpoints, what would be the intervention durations, or help to optimizing the cost of the clinical trials by introducing different steps. Also helpful, the data collection as well as the post-trial analysis that align well with the previous answer that I provided. Thank you very much. That's very helpful. Just in the few moments we've got left, if you could just point out some of the blind spots that we should keep in mind as we consider using AI in conducting research. I think there are a lot of different blind spots that I encountered so far. But two main things that I usually think twice or three times, data bias and generalization of the model that I come up with. Usually, the model work really well with this data set, with this population. Turns out it doesn't work with the others and vice versa. So this is very important to think and to re-evaluate and to validate the model with different set of data set and with different scenarios. And also, one of the blind spots that I also encountered is the data quality and the standardization of the data that I can be able to use for secondary or retrospective analysis. We have different sources of data come from different places with different standardizations, criteria. So this should be considered carefully. Otherwise, the model, the data, or the analysis that we get from those models will not be valid. Great, thank you so much. Dr. Li will be part of our freestyle discussion. So if you have questions for Dr. Li about research, feel free to put that in the Q&A. So I'll move on to our last panelist. Last but not least, Dr. Sachdeva. Thanks very much. And thank you. So we're going to round off our panel with a discussion on legal and ethical issues. So Dr. Sachdeva, thank you very much for joining. Talk a little bit about the ethical challenges that come up in your mind. And what are some of the top areas of concern that our audience should be aware of? Great, thank you, and thank you. You seem to have frozen on my end, Matt, also on yours. Yes. Dr. Sachdeva, if you can turn off your video and just audio, you may have a limited bandwidth that will only allow the audio. Otherwise, we can get back to you. Live webinars, there's always one glitch. Well, we can still transition to our open forum. And we'll come back to Dr. Sachdeva when he is able to join us. Sure, we have a lot of questions. Sure, we have a lot of questions. So let's go back to Amber. There were a couple of questions I had to skip as we were moving through. Hello? Oh, let me come back. Yeah. I'm sorry. There's a sudden glitch with the internet, but I'm back on. So I was just commenting. The other growing area of concern from an ethical standpoint is privacy. So let's say we have using a wearable device. What happens to the information from that device, particularly as it moves to different areas? So the issue of privacy and, in turn, informed consent of the use of the secondary uses of those data in the AI analysis become the real ethical concern. That's great. Talk a little bit about what clinicians should be most aware of as this legal landscape surrounding AI evolves. Yeah, that's a fascinating question. And the short answer is that this entire area is evolving from the legal landscape. Again, as pointed out on this slide, the remarks about the legal comments are only for education purposes. They're obviously not legal advice. But having said that, this entire area is growing. But the two broad concerns, at least in my mind, that come up, one is basic medical malpractice. So let's say the AI algorithm provides us a certain recommendation. As a clinician, if we use it and there's an adverse event, effect, outcome on the patient, where's the liability? With AI or with the clinician? Conversely, if we don't use it and there's an adverse event, the concern of omission, then where does the liability rise? And the courts really don't have a good answer on this. And the fundamental question becomes is, is AI a product or a service? And that relates to the second area of a growing area of concern. And I think only time will tell as some of this gets litigated through the case law and the courts is that, is AI a product itself? And historically, the courts have leaned towards looking at AI more as a software. But AI is its own thing in a way. It's not really a software, particularly with training sets and autonomous AI and things that are on the horizon. So this is evolving. But the fundamental points are issues of medical malpractice, what are the standards of care, particularly where some of these algorithms may be embedded very deep into the data sets. And of course, this issue of classic medical malpractice or product liability. So that's very interesting. And I think you raise more questions, perhaps, than there are answers. But then it does also, for our clinicians, introduce the issue of vicarious liability. So if you're an owner of a sleep lab or owner of a sleep practice, and obviously, perhaps, you're not doing the scoring, but your technologists are or the physicians that you've employed are using AI in generating notes and other things, perhaps. What is the liability here? Is any of that known yet? The short answer is it's not clear. But it may be worth it to take a step back and say, what is vicarious liability? So vicarious liability really comes from a doctrine called respondent superior, which means that as an employer of a hospital or a sleep lab or any entity, actions being done by the employees, even though the employer is not directly involved in the actions, there's an imputed responsibility and, in turn, accountability and liability to the owner. And a lot of the cases here actually come from even before AI in the area of drowsy driving and what happens in those situations. But the liability goes from the individual, obviously, who is interacting in the AI space with the patient, then the owner, and potentially, the entity itself. So the liability goes all the way up. There are some cases in the robotic use, surgical robotic use, like da Vinci robots and so on. Looking at this, and we can certainly take a deeper dive into it in the Q&A, but with respect to AI, this area is, I think, very important, but no clear direction at the moment. I think that's the best way to put it. Thank you. Matt, do you have a question for Dr. Yeah, so there was one about liability, given that there are questions about where liability may lie, this concept of vicarious liability. Given that AI can make mistakes, you know, I'll throw this out to Dr. Sachdeva first, and maybe to all of you, does it create worry about your practice, the fact that this is something, you know, in some cases new, in some cases, these are new considerations? Yeah, you know, I think that's a great question. And in my opinion, just like any technology, there's worry, but there's also a clear benefit of the use of the technology. So really the issue becomes is how do we minimize the risks and maximize the benefits? And from a, more to your questions, like you take some of the other situations like drowsy driving and so on of this vicarious liability, imputing liability to an owner or another person, and not the employee directly, a lot of this issue comes to this notion of foreseeability. Was the adverse effect foreseeable? And that's where the standard of care comes in. So I think another important consideration is what can we do to mitigate these risks? Well, you know, in the areas of privacy, there's this concept of dynamic consent to make sure there's adequate consent, and it's not just a one-time consent. Limit the amount of data being shared and supplied so we minimize the risks. In the more of the vicarious liability issue of making sure that the output coming up from the AI algorithm, to make sure that we're not just simply following it blindly, we are using our clinical judgment, and that goes to a previous question, will AI replace human beings? And I think probably not, because always that human touch provides that clinical context and provides the check and balance. So I think as long as reasonable steps have been taken, we minimize the risks. But again, I don't think one can completely eliminate risks from that end, but we can minimize the risks and maximize the benefits. So it's very nuanced area that we're working into now. Great, thank you. So I'll just summarize things that we've talked about so far. As you can see, AI supports multiple facets of a sleep medicine practice. Certainly sleep study scoring can be streamlined with AI, and it can improve efficiency and reading costs. Business operations, as Dr. Zia has pointed out, can be remarkably enhanced with AI-driven tools. As Dr. Kolop has pointed out, AI scribes are currently being used in many centers and can facilitate documentation, reduce burdens of paperwork, and enhance patient care in various ways. AI tools can accelerate and enrich clinical trials, as Dr. Lee has pointed out, and hopefully that speeds up drug discovery even. And finally, as Dr. Suchdeva pointed out, AI is something that we must be careful about as we venture into the legal and ethical issues surrounding its use, and we've just started to scratch the surface. So with that, I think we will open it up to the Q&A section, Matt. Sure, so it looks like there's a number of remaining questions, and I will try to direct them to the recipient that it was meant for, but feel free after that, we have a panel who has a lot of different perspectives, so feel free to jump, to make that a jumping point for your ideas as well. So the first one goes back to Amber, it's about scoring. Let's see, so, do you need to manually overread the AI scoring, and if so, have you found a lot of significant errors with the software? So yes, we overscore every single study that comes through the lab. You have to add that human component to it, because AI is not perfect. As far as extreme things that pop out, I mean, a lot of times it likes to mark post events, so you have to take those out. That's really not that big a deal when before it was taking me an hour to score a sleep study because there wasn't a ton done in it. Now it's pretty much ready to go. We're in about 90% agreement with ENSO scoring. So we do, we go through and we do checks to make sure that we agree on what's going on and all that data gets sent back too. So it's definitely a time saver, even if I am having to pull the posts out. Yeah, and that's totally in line with accreditation standards that allow for auto-scoring or AI scoring. Some auto-scoring is artificial intelligence and some of it is used, is rules-based. So frozen algorithms, but whatever one you use, it must be over-read. It must be reviewed before an interpretation according to the standards. Matt, can I just jump in and ask you to just make a quick comment about the accreditation process or the certification process, I should say, that ASM has for these automated scoring programs? Yeah, so last year we concluded a pilot program to see whether this was feasible to expand into full PSG in-lab type ones. The pilot was limited to stage score only. It was limited to a two-year certification. And basically what we did was it was a voluntary industry program where the industry company would sign up. They would give us their software to install locally, cut off from the internet or from the cloud onto an ASM computer. And the software would score 100 records. Now the output was compared to a pilot group of gold standard scores, 11 of them, that were really experienced. And their results were compared to the auto-scoring software results. And the auto-scoring software results that met a threshold, a standard for that pilot were given auto-scoring certification for stage. And you can find that on our website. There were three companies that passed that threshold. EnsoData, Somnometrics, and Philips with their software. But that was just a pilot, just limited to stage. We're developing a full PSG auto-scoring certification program that would be long-term and that would have, it would be more robust, learn from the pilot. It would have more records, more scores. I think we're looking at 15 scores. It would have more diversity in scores, diversity in records compared to what we had. And that's gonna take a long time to score the records, to open up the application period, but we should start seeing results, not 2025 sleep, but before that should open up 2026. So we're making progress on that, but that's a much bigger undertaking than the pilot. And that will give, we get a lot of questions from members. We're considering a lot of software for automated sleep scoring. Have you tested any of them? And we can finally say, we have done some limited testing, but this full PSG would be really robust testing because the FDA threshold, the FDA approval is somewhat of a low bar because they're not actually given a third-party testing to any of the software. It's data that's submitted by the companies. So this would be the first opportunity in our field where we're actually doing, providing oversight over software in a third-party way that's neutral. So we do actually coming off of that, have a few questions on AI scoring. So let me get to a couple of those. Let's see here. So there's some with asking about which software is recommended for AI scoring for lab studies. Amber mentioned, and so data, I mentioned a few that were certified for stage. Does the rest of the panel have any other that others that they've used or that they would recommend? So maybe just open it up to the panel. Are you using AI scoring in your labs? Ramesh, you came off of mute. We just, this is the interview. We use the auto-scoring for Watchpad and for our HST, but not for polysomnography at this point in time. Yeah, for the sleep image as well and Watchpad, they are auto-scorings. And do either of you have, know the companies that you're using? The auto-score program company. So Watchpad has inbuilt auto-scoring and sleep image also has inbuilt auto-scoring. None of them integrate with Anso data or other PSG scorings because they want to keep that in-house. Okay, great. Other questions, Matt? Yeah, on AI scoring, has anyone noticed that uses it or whose lab uses it, have noticed an improvement in reliability and accuracy over time, given that it is a deep learning approach that would change over time, the more that it learns? So we're on year three and a half of using Anso. And yes, it has over time, I'm changing less and less. It's not something you notice until you book back a year. Also it's like, oh yeah, I haven't been marking as many of those or I haven't had to change this as much. So over time it definitely does improve. Hmm. And there was a question about using rules-based algorithm versus AI algorithm for scoring. So Amber, I'm assuming since AI would learn, there's some advantages there. There are. AI, when I go in and I look at an AI scored sleep study, it feels like a person scored it. It doesn't, it feels like it's more natural, I guess. For some reason, when you look at an auto score, I feel like I have to change things and move things. And I don't see that as much with the AI scoring. Okay, we have a couple for, to start off with on scribes for Dr. Collip, perhaps. So how much of the scribe note is just quoting versus organizing? How much organization of the conversation is done into the format of a medical note? Yeah, it's amazingly good. I mean, like I said, some of the things the scribe finds important are different than what I find important. And sometimes it misses stuff and sometimes you have to rearrange. But I mean, it is intelligent enough to know kind of what to put where. And so I've been very impressed with its ability. Now, not everybody in our practice decided to use it because I think some of the nuances of certain diagnoses just aren't really captured by the scribe. It doesn't deem them important. But I hope that as it improves and becomes more trainable, that will also improve. And Dr. Collip, is ability, I forget the name of the one that you mentioned, but is it able to flag or screen out HIPAA protected information? Like if a patient comes in and talks about a neighbor or talks about a family member as part of their story, their background, does it flag that or does it automatically put that into their chart? No, yeah, it's called a bridge. And it doesn't put in a lot of peripheral stuff typically. Like it kind of, again, amazingly to me, it kind of sticks to the script. Because when I have lots of patients I've had for a number of years, like I'll chat with them about their kids or whatever, their dog or, and it's smart enough to not really put that into the note. It really kind of sticks to the medical part of things. So yeah, it's pretty good. And there is another specific question from an attendee about how it handles an assessment and a plan versus history. And does, the question is whether it misplaces that information or how it handles that. Yes, it does. I mean, like I said, it all needs to be edited and edited. But again, it's pretty good. Like sometimes there's some redundancy in the HMP and the assessment and plan, or some things are misplaced. But more importantly, it seems to capture, like when you have a discussion with the patient, it seems to capture most of what your plan is as long as you discussed it with the patient, right? So it's, again, it's pretty good at differentiating what should be in the HPI, what should be in the assessment plan, but sometimes you have to rearrange it a little bit. Okay, we have time for one more question and then I'll put our contact information for follow-up on the final slide. And again, any of the questions we receive after this will be saved to the attendee who asked it and we'll follow up after the webinar. So the last one is on AI scoring and it's about whether AI has improved its accuracy when a study's quality is compromised by artifacts. Yeah, I see it a lot of times. I love, that's when I really love the AI because if you've got someone who's got a VNS device or constant muscle, or you're just getting some weird artifact, like in Inspire, and you can't technically really see what's going on behind there, AI can pick that out. So it can tell there was a K complex in there. So we know we're in stage two when I could look at it and be like, I think they're awake because I can't read anything. So yeah, it's been very beneficial for those. Great, great. So thank you so much to our panel who worked really hard to plan and for your time today. And thanks to everyone who participated. Great questions. So stay tuned for upcoming AASM committees present webinars offered by the AASM on the first Friday of every month at this time. Now in about two weeks, there'll be an email sent out with instructions on how to access this webinar recording on demand. Thank you, everybody. Have a great day and great weekend. Thank you.

AI in sleep centers

AI-powered auto-scoring

clinical workflow tools

AI scribes

patient care delivery

sleep medicine

AI efficiency

ethical considerations

data privacy