An AI Cure for Provider Burnout, Poor Outcomes, and Low Reimbursement

John Wilkinson (00:04):
My name’s John Wilkinson. I’m the Executive Vice President of Account Management here at Clinical Architecture. It’s my duty here to begin this with introductions. Let me advance the slide and let’s go to Stephanie and then Dr. Charlap and then Michael. Then you can go on from there.

Stephanie Broderick (00:25):
All right. Hi everybody. I’m Stephanie Broderick. I’m EVP of Strategic Initiatives at Clinical Architecture. I’m responsible for partnerships and industry facing initiatives.

Steven Charlap, MD, MBA (00:37):
Steven Charlap, CEO, and founder of SOAP Health, building, the world’s first AI powered primary care physician by September, 2024.

Michael LaRocca (00:50):
And hi, good afternoon. I’m Michael LaRocca. I’m the founder and CEO of Ready Computing, and I’ve had the pleasure of working with John, Steph, and Steve for a number of years now, and we’re excited to tell you about what we’ve been up to at the VA. So thank you for attending.

Steven Charlap, MD, MBA (01:10):
Anybody who’s in healthcare understands the challenge being faced by physicians, particularly primary care physicians these days includes overload of information, ineffective tools, ever expanding body of medical knowledge, which is impossible for the human brain to keep up with. Reimbursement, documentation challenges, decision fatigue. When you go to the doctor only go in the morning in the afternoon, they’re not the same person. And finally, the threat of malpractice. The need to practice defensive medicine, all these taking a toll on physicians resulting in burnout.

Michael LaRocca (01:52):
So thanks Steve. Steve really laid the groundwork out for what some of the motivation was behind the VA initiative that we’ve been part of, and it’s referred to as the AI tech sprint. And the real concept here is to figure out how artificial intelligence could be used to reduce physician burnout or provider burnout. If you, just, to Steve’s point, if a doctor is overworked, that doctor won’t necessarily give the best care possible to the patients. And how could we find ethical uses of AI to help the providers in a sense care for those who care for us, right? It’s sort of the concept. So I’ll give you some of the key goals that were part of this initiative. And just to give you a little bit of background too, it was a technical initiative where over the course of about two and a half months, and we’re still actively in this program, we’re coming to the end of it, but during this two and a half month period, we’re tasked with building and implementing a prototype that’s able to show the end-to-end interactions between some community providers and ultimately the physicians on the other side with infrastructure in the middle that contains all these AI functions.

(03:05):
So I’m going to take you through that in more detail, but that just gives you a little bit of context for what you’re about to see in the presentation. So some of these goals are, well, naturally we have to optimize the time that the physicians are spending during their day, right? It’s almost like we don’t want a doctor working 80 hours a week to do 40 hours of work because there’s all these inefficiencies, and what if we could use AI to help 40 hours of work truly be 40 hours of effective care that they’re giving to their patients? We want to visualize data in a new way right now. Doctors have access to so much information that we need clever ways to make that information accessible at the time of care and new ways to sort of represent and visualize it. We want to minimize or eliminate diagnosis.

(03:54):
One of the key things, and I know Steve, my co-panelist is always talking about this, that one of the biggest challenges in the clinical world is being able to diagnose accurately, effectively, and quickly. And the threat of misdiagnosis is something that we really need to take seriously. So how can artificial intelligence help reduce or eliminate those cases of misdiagnosis and make treatments naturally as effective as possible? So our approach was at a high level is kind of simple, right? Let’s automate whatever we could automate. We want to eliminate all the manual steps wherever those exist, and let’s automate what we can. Let’s get through these vast amounts of data. I mean, if you think of the size of the VA where this initiative is sort of sponsored and where we’re working, they have 130 or more Vista instances, the EMR, they have 170 hospitals.

(04:50):
It’s just a massive amount of information, and now you add the community providers and some of their other exchange partners like the DOD, and you’re just dealing with data sets that are massive. So how can we get through all of this data, both structured and unstructured, and make sense of it through clever AI functions, reduce the data set to what’s available for the context of a particular visit? What would a pulmonologist want to see? What would a dermatologist want to see and use AI to trim that data set down to what’s relevant to that particular episode of care. Infer possible diagnoses. This is one of the really key features of the Clinical Architecture system to be able to look at all of this data and then actually help the doctor by providing a bit of support. You might want to look at this, you might want to rule this out like inferring from the data, what possible diagnoses are. Using virtual assistance, as you’ll see a core part of the SOAP health platform, being able to use a virtual medical interviewer, which in a lot of ways could be more effective than a human interview.

(05:54):
I don’t want to steal too much of Steve’s thunder, but it’s very fascinating and naturally making data available. Also, post-care. What are some things you might want to do? Researchers and analysts that want to look at the data post-care. So that’s kind of our approach. Now, you might wonder why our three teams came together. We’re really between the three of us. We’ve done a lot of work in the interoperability space. We understand how to connect systems. We know how to normalize data structurally and semantically. We know how to deal with all those sorts of challenges. We also have paid a lot of attention to our use of AI and making sure it’s ethical, being responsible. We’re guided by NIST’s trustworthy AI framework. These things are really important to us. NLP experts, again, one of the absolute core competencies of Clinical Architecture, just the absolute core competency, and this project depended quite a bit on it. Everything we do has the human at the center, human-centered design. We know that you could spend all the time and money in the world, but if you don’t have a system that’s adopted by the users, then it doesn’t benefit anyone. So everything that we do, we make sure that it’s designed right to make people’s lives better. And we have a proven track record at the VA.

(07:14):
So I’ll take you through the technical solution a bit. So there’s a lot of boxes on here, but let me just guide you through it. On the left side, you have the community partners. Just think of that as the data sources for this particular tech sprint. And on the right side, we have two kinds of targets. We have the clinical data warehouse at va, which is really for the researchers and analysts, and we have a number, a collection of the physicians that are ultimately the biggest beneficiaries of this AI framework that we’ve put together and the key target user. But those are all on the right side. So you kind of have this left to right orientation because the data’s coming in one side and eventually out the other side, but there’s a lot of magic that happens in the middle. So I’ll take you through each of those. So Ready Computing, we are really the systems integrators, if you will, a part of this initiative. So we stood up software to sort of act as that interface engine that took the inputs, produced the outputs, and then we helped sort of navigate mitigate all the traffic between the sub-services.

(08:19):
Clinical Architecture was the first sort of piece of software in the process where you’ll see the usage of AI. There’s a lot happening here, and there’s three major steps. The first thing is being able to normalize the data, and that’s structurally and semantically. We might get CCDs in or HL7 messages or X12. We could get a variety of things and to be able to normalize that structurally and semantically is the most basic step and the most important step that we have to do to make the data usable. Then through their service, we recognize codes out of the unstructured data. That’s really important. I mean, some of the data we process are things like a coded allergy, a coded medication, but a lot of the data is also unstructured. It could be a progress note or a comment or something, and there’s a lot in there.

(09:07):
So using the CA technology, we are able to recognize coded concepts and the unstructured data, and then finally, the inferences are made. You remember on the previous slide I talked about trying to eliminate misdiagnosis, right? This is one of the key steps in that because it’s actually suggesting possible diagnoses based on an AI sort of driven review of the content. So that content goes through this AI function, and it comes out with what are probable diagnoses that you might want to look at. If you’re the physician, you might want to rule something out or you might want to confirm something, and it’s done with all the justifications for why that suggestion happens. So then control eventually comes to the SOAP Health software. And in that you’ve got two major functions. You’ve got the screening, what you might think of as the subjective part of this work where a patient or a veteran is actually using an AI driven survey, and it’s an AI bot or a virtual assistant that’s doing the interviewing.

(10:12):
So imagine instead of talking to me or your doctor or whoever, imagine if you’re talking to an AI driven virtual interviewer that goes through and gets a lot of content, often more effectively than a human interviewer, and gives the veteran a chance to not only validate what the current veteran health record is that’s there, but also add new information that’s captured from the survey, so it’s very effective. And then ultimately optimizing it. You might remember I said that we want to not overburden the providers, which just immense massive amounts of data. So this is a step in the process where data is actually sort of trimmed down and tailored for that particular context. So just quickly in summary, the results of this of running data through the solution is AI driven SOAP notes, right? SOAP notes being the Subjective Objective Assessment Plan notes, something very common, being able to generate and support the use of those SOAP notes, optimizing the encounters to make them as efficient, as effective as possible, and ultimately also enriching the data with all the subjective data that we were able to capture from the veteran.

Stephanie Broderick (11:29):
Thanks, Mike. So as Mike was explaining, there is that step within the process, which is the recognized step, and that’s where Clinical Architecture invoked our SIFT product. And SIFT is in back-cronym. Basically we backed into it. It stands for Semantic Interpretation of Free Text. It’s a clinical NLP product. And essentially what we did was we received a diverse set of records from community providers, patient encounters, complex medical documents. What’s shown here is an example of the types of documents that we were asked to process. The goal was to expand the knowledge for a physician of that veteran and to improve the continuity of care, as Mike is going to explain, there were several tests that we had to go through, several different milestones, and we’re still working through this process, but as part of the second gate, we were given six documents to process.

(12:28):
Five of those were procedure notes. One of them was a patient summary. So ready computing, grab the docs, they generate the text through Amazon Textract, pass it to us, and then we put it through our SIFT product along with our Clarify, which is what summarizes the results. And this just is an example of what’s happening within the SIFT product. So we basically are looking for different headers within the document, and then based on that, running the document through what we call SIFT arrays that are very, very targeted towards the domain that’s being represented. So the way we process lab is not the same as the way we process medications, not the same as the way we process chief complaints. And so you can see how the data is coming out. So we’re identifying lab results with dates, and the data then is being put into a JSON output and then provided out. And as the information is being detected within the results, then we are coding the results to specific terminology.

(13:54):
So we can process a number of different sections of the note, and these are just a sample. So labs, meds, procedures, procedure findings, history and chief complaint. We also make use of synonomy, a lot of synonomy to do the appropriate targeting. We also are able to handle negation, so not having the presence of something is really, really important. And then finally, the information is passed back to Ready Computing, and then they’re providing, they’re taking the information and distilling it down and categorizing the information. It’s de-duplicated and formatted for easy consumption, and then this information is passed into the VA’s EDW and then passed on to SOAP for follow up with a veteran.

Steven Charlap, MD, MBA (14:50):
Okay. So as we’ve been discussing, SOAP stands for a subjective objective assessment plan. One of the most important things we can do in healthcare is validate data with patients. Only a patient knows what medications he or she is taking. Only a patient knows if they’re taking the medication as directed. Only a patient knows whether or not they are or not diabetic. Many times additional conditions are added to medical notes to support billing, but in fact are not present. Giving a chance for a patient to validate is exceptionally helpful. So we’ve built the perfect medical interviewer. It took many years to build this. It is patented, clinically validated at Stanford and elsewhere. It does a 100%voice-based interview of the patient on any web-enabled device. It can be started on a desktop, finished on a mobile app. It’s a mobile app design with a progressive web app.

(15:49):
It has been proven to be a better interviewer for several reasons. One, it collects more truthful information. A study show that more than 50% of Americans lie to their doctors about things potentially embarrassing. There’s a mountain of signs that show that people answer more truthfully to a digital human. The brain lights up differently. It uses intuitive imagery to help people of low literacy levels understand what’s being asked of them. And finally, it reflects answers back to the user. It’s also always accessible, infinitely patient, sometimes funny. Second, we built the world’s most comprehensive risk assessment, 500 algorithms that look at family risk, lifestyle risk, social determinants of health risk, et cetera, and does a differential diagnosis. That differential diagnosis both generated G novo, as well as fed into a large language model to be strengthened. And finally, it generates today a specialized formulation of a note for the doctor, making it easy for the doctor to read it, edit, and finalize it.

(16:56):
As I mentioned, we are building the first AI powered PCP, but it can’t be built in a day. These are the three components that are already complete. This is what I mean by a special format for the physician. This is a screen that has key data. We both collect systematic information as well as open-ended information, allowing a patient to go through a systematic checklist of symptoms, as well as open mic response to a question such as, what do you think is causing your symptoms? What other concerns do you want to share with your doctor? What request do you have? And so we presented initially key data. Then we do risk view, which is a risk profile, reference diagnosis, diagnosis, profile, and then the SOAP note, which is a traditional looking SOAP note that can be pushed as discrete data into discrete data fields in the EMR. We will finish Athena Health Integration in 10 days from today. And then finally result file where you can have your lab results and your diagnostic results, et cetera. And then that’s what the final SOAP note actually looks like. It’s a judicial note looking SOAP note that can be pushed as discrete data into discrete data field.

Michael LaRocca (18:12):
Yes, I’ll take you through just a little bit of scheduling. So this is an active initiative that we’re in. It kicked off in January and the 19th just this year. And as I mentioned earlier, it’s about two and a half to three months to get through the whole process. So far we’ve gone through Gates one and two and we’re sort of awaiting gate three, right? So I’ll take you through each of the gates. The first gate started off with the questionnaire fairly detailed one about sort of our general approach, the IT infrastructure to the applications and services that we would implement in that infrastructure, but it was essentially a deep technical questionnaire. Then we had to make sure through this gate that our software could be accessed from a government furnished piece of equipment, right? So as long as a VA staff member using GFE equipment, then if they could access the software, then we’ve done something, right?

(19:09):
We had to have a robust framework for error handling that goes without saying, our response times had to be under two minutes, and the expected output that came out the right side of that diagram had to be what was expected. Gate two took it another level, deeper. First thing we had to do, I mentioned the NIST trustworthy AI framework, which is something really good and important to read. We had to make certain attestations and respond to certain questions and another questionnaire just ensuring that we’re ethical users of AI. We had to make sure that data that came out the right side had clearly indicated correct data and as well as incorrect data. So just making sure that the data was in fact accurate. In cases where it wasn’t accurate, we pulled that out and that was reviewed by clinicians. And then finally that the content is in fact coherent and making sure that a doctor can read the content and make good use of it, that it was comprehensive and coherent. As I mentioned earlier, where this is the stage that we’re at gate three, the still sort of yet to finish, but we’re getting close to final submissions. And then ultimately, if everything works out right, we’ll be on the podium there.

Virtual Assistant- Video (20:43):
Hi, I’m Jeannie. I’m the virtual assistant created by SOAP Health to revolutionize healthcare through the power of artificial intelligence. As the perfect medical interviewer, I conduct automated patient intakes before appointments. My conversational interface guides patients through a comprehensive interview covering their health history, medications, family history, lifestyle factors, and more. This not only saves physicians time, but also ensures no critical information is missed. I can capture more complete patient data than traditional paper forms. My advanced natural language processing skills allow me to have fluid natural dialogues with patients and understand complex health information. After the automated interview, SOAP Health’s risk view feature analyzes the patient data I’ve gathered to identify any potential health risks or red flags. This gives doctors actionable insights to improve preventive screening and early detection of diseases. Finally, I generate a smart SOAP note that pre-populates all the essential documentation for the patient visit, integrating directly into the doctor’s electronic health record system. This automates up to 90% of the documentation process, saving physicians two hours per day that can be better spent on direct patient care.

Steven Charlap, MD, MBA (22:04):
So let me just summarize what this whole partnership is about. We are able to pull the highest quality, most trustworthy, most accurate, and most comprehensive direct interview from the patient. Ready Computing is able to extract existing historical data from community sites as well as from the electronic medical record itself. Clinical Architecture is able to take that data, normalize it, and push it through SOAP to run it through the patient to get the patient’s validation further of that data. Combined, we are creating what we call the precision patient profile, the most accurate comprehensive database ever created in the United States beyond some of the governmental efforts.

John Wilkinson (22:59):
Questions. Does anyone have any questions? I guess I have one. As you work through this, obviously a couple of months, what are some of the lessons learned from three companies with products that are standalone and doing well on their own to make them work together and work together as a team?

Michael LaRocca (23:19):
Maybe I could start. For me, it all starts with a concept and understanding how the services that each of us offer would interact with one another. And from that concept drilling down into a proof of concept and a prototype and that we were able to do in a really short time, we worked well together. We have known each other for a bit, but our teams work well together, and that to me was just a great experience getting that off the ground.

Steven Charlap, MD, MBA (23:45):
Yeah, I’d say the biggest challenge from if to be transparent is customer discovery. The VA did not give us as clear instructions as one would’ve hoped for this project, and I think that made us need to be more creative than one might have to be if one had very clear instructions from the customer.

Audience Speaker 1 (24:12):
What sort of clinical use cases would you guys focus on? The VA gave you specific clinical use cases?

Steven Charlap, MD, MBA (24:22):
I can? Well, the data that we got were diagnostic reports like ultrasounds of the abdominal aortic aneurysm, and a typical patient case study of a primary care patient. But the reality of this technology could be applied to virtually any medical specialty other than maybe anesthesiology in some customer discovery we’ve been doing outside this project, we discovered that even surgeons who we thought wouldn’t be as interested are in fact interested in this type of capability because the whole point in this project was burnout. SOAP’s product alone can knock 12 minutes off a patient encounter. If you see 25 patients in a day, that’s 300 minutes. There is no product in the market that claims saving doctors right now, five hours. Okay. That’s from a work efficiency. But now from an accuracy perspective, we can significantly improve accuracy and increase compensation because we identify medical complexity missed by the doctor.

(25:30):
So we’re actually increasing what the doctor can bill, and we found a 17% average increase in revenue, plus some new HCPCS codes that could be billed that went into effect January, 2024, and a reduction in no-shows. Because once patients engage with an application that’s taking an intensive history, they don’t want to miss that appointment. So we actually saw in a practice with 14% no-shows drop to 0% no-shows. So saves doctors’ times allows patients the luxury of doing this from home reduces no-shows increased average revenue per patient, and most importantly, most importantly, reduces misdiagnosis, reduces malpractice exposure and improves patient outcomes.

John Wilkinson (26:23):
Any other questions? Thank you all for attending. Thank you for your hard work. Thank you all. Thank you.