Exploring the evolving landscape of autonomous coding

[UPLIFTING STRING MUSIC]

[LASHA] All right. So we are very, very excited about this next portion of the agenda, which is a panel discussion on everyone's favorite topic, which is autonomous coding. I'm very pleased to moderate this session, and I want to introduce you to our panel of Optum experts who will explore the evolving landscape of autonomous coding. We're going to talk about trends, challenges, innovations that's shaping everyone's future. This session is designed to be informative and interactive, so we want to hear from you. If you have questions, as always, utilize the Q&A box. You should see that in the bottom right-hand corner, I hope, and we'll go ahead and get started. We will hopefully have a little bit of time after this panel discussion to answer those questions, if we receive any. Our goal with this panel session is to really just foster a deeper understanding of autonomous coding and also to gather your feedback, and we hope that it will guide some of our future initiatives and some of the conversations that we've been having. So, let's go ahead and get started. So I'm very pleased to introduce our panelists for today's discussion, starting with Temeka Lewis. Temeka is the Director of Automation Capability of middle revenue cycle product. She has over 20 years of experience in the healthcare arena as a health information management and revenue cycle professional. She joined Optum as a subject matter expert for our print publications and then moved to sales within our provider market solutions. In her current role as the Director of Autonomous Capability within our middle revenue cycle pillar, she leads an agile, multi-layered team of AI and NLP developers to optimize performance of AI being used within our solutions. Temeka has an MBA in Health Services Management. She's credentialed as a CCS and CDIP through AHIMA. Also joining us is Lisa Tesno, who you guys hear from all the time. She's the Director of Value Management. She has 39 years of experience in healthcare and has been with Optum for 15 years. She leads a team of subject matter experts that support the AI technology in computer-assisted coding and clinical documentation integrity products. She is credentialed as RHIT, CCS and CDIP. Also joining us is Shannon Weintraub. She's the Director of Strategic Product Management and Automation. Shannon began her career in Optum in 1998 within the physician billing division. She quickly advanced into a leadership role and coding services, where she served for over 17 years. With more than 25 years of experience in the healthcare industry, Shannon brings deep expertise in patient financial services and coding. She's currently serving as Director of Product. She leads strategy and roadmap development for middle revenue cycle automation. Her academic background includes a BS in Business, Finance and Marketing. She also holds an MBA  and completion of the Health Information Administration Program. As you can see, we have some incredible subject matter experts on our panel. Thank you, Temeka, Lisa, and Shannon, for being with us today. I'm excited to share your expertise with all of our user group participants today. So with that being said, we're going to go ahead and get started with our very first question. So, I'm going to start with you, Temeka. What is the difference between our existing computer-assisted coding solution, autonomous coding and automation?

[TEMEKA] Ooh, big question. Not sure it's the first one we should discuss, but okay, here we go. So... whenever I'm describing or defining a new concept I like to use analogies. So, forgive me, I've got one in the back of my mind that I think may help drive our point home again, because we really want you to understand the differences between the three. So, there's automation, and when I think about automation, it's really simple definition, anything that makes your life easier, it's more efficient and you don't lose quality. So my analogy is we're going to talk about, you know, taking a taxi Uber or a Lyft. So, you are in control, right? That is sort of a service that you choose, you can schedule your ride, you pick the type, but you don't have to drive yourself. Someone else is doing the driving for you. So, to me, that's a form of automation, makes my life easier, I'm more efficient, sometimes I work while I'm in the back of the Uber, and I didn't lose any quality. That is equivalent to our computer-assisted, underline the word “assisted”, computer-assisted coding technology that many of you use either in CAC or in OIO. So, there's a form of AI that reads your documentation and then suggest the codes for you, but you are still in control. You send us which documents you want us to review. You map those to do I want CLI to read these or not, and then you've got a standardized workflow where we suggest the codes and you can either accept those, you add them, you modify those that need to be modified or you delete it, right? That is deemed automation. We also call it Automated Medical Coding for those of us who are HIM professionals. So, that is sort of the baseline. Now, when we think about the next level up, I'll go to what I consider to be partial autonomous coding. This is the hybrid approach, right? So, back into my Uber or my Lyft, there's still a driver in the driver's seat, but there's an option to press cruise control, right? So, if you're in optimal conditions you've got a long stretch of highway in front of you, you can turn it on, turn it off, adjust as necessary. You got your foot hovering over that brake for just in case. You're monitoring what's happening while the vehicle is on cruise control. I like in that, again, I call it partial autonomous, are straight-to-build rules, right? Where, again, you're still in control, you're in the driver's seat, but there are certain scenarios that you could use and you can put them on what I consider to be cruise control, right? So, in the pro space, if you've got our application suggesting, you know, simple code pairs, your chest X-rays and all of the diagnosis codes that may be medically necessary in an LCD, you could put those together and say, you know what? I don't need my coders to touch those. Let's partially automate, right, those code pairs so that those could go “straight to bill.” Then on the facility side, your charge codes, your Chargemaster-driven CPT. Those are already automated, right? So what if we built, which we are today, building the capability to grab that charge code, right? Which is your evidence of the service happening, pair it with a diagnosis suggested by our core CLI technology-- And then again, there are certain situations, certain scenarios that you would be aware of, that you want to automate. That to me is our partial level of autonomous where you can take certain service types, certain service lines-- I mean a screening mam is a screening mam-- same CPT over and over with the same diagnosis code, right? There are certain scenarios, certain optimal conditions where you're in control, you're still monitoring what's happening. You could take that information and you can build a rule to say, let's not have our coders pretty much touch those. But again, there's still a lot of monitoring in the background, like a new LCD comes out, new diagnosis codes are not medically necessary. So, what do you want to do? You want to go into your rules. You want to update your rules. Or maybe if they change a payable diagnosis. But again, it's still taking a portion of your coded services and putting that specific service line, that specific type on cruise control, but you are still in the driver's seat. Third level, autonomous coding. So this is where we see fully autonomous, right? So some key words you might hear out in the industry is AKA no touch. So fully autonomous takes the driver out of the driving seat from my Uber and Lyft scenario. So, I'll share a bit, right? I'm a new resident of Arizona, and I can remember landing at the Phoenix airport once before and a car rolled up without a driver and picked up a passenger and took off. Now, at first I thought my eyes were deceiving me. And then I learned that in certain cities they are testing autonomous driving capabilities. The service here is called the Waymo. So, there's no driver in the seat. You know, it's traditionally or typically no touch. But there is a tremendous amount of monitoring that is happening behind the scenes. The cars look different. You can see all the radars and the... the receptors hanging off the side of the car. Like you can really tell that there's an incredible amount of-- I guess I'll call outside-of-the-car monitoring happening to make sure that the Waymo doesn't go off course. That is very similar to autonomous coding. Again, you may hear the words no touch coding. So what happens is there's a small service line or a portion of certain visits that we can lift out of your traditional coders workflow, and we monitor those still remotely. There's a heavy amount of monitoring with our AI solution. We'll talk about that later. But autonomous coding, again, it is what scenarios, what optimal conditions do we find that we could take the coder out of the coders seat or take the driver out of the driver's seat and allow those specific services, under certain conditions, to be autonomously coded or “sent to bill” without any human touch. So those are our three layers, right? We've got automation, give you some efficiency. Partial autonomous, which is like your cruise control. You're still in control, you're writing rules versus fully autonomous, where we use a different form of technology to assign the codes, but there's a tremendous amount of monitoring along the way. So hopefully my analogy didn't turn anyone off. But again, I was really shocked to see a driverless car. But that is how automation is happening just in general across the world today.

[LASHA] Wow. Thank you Temeka. As someone who’s used Lyft, Uber, and Waymo, that's... I totally related to that. Waymo is a trip. If anyone's ever used Waymo, it was quite the experience, to say the least. So they helped me to understand that very much. And I hope that helped everyone else, so thank you. So, building on that, what is the biggest, um, what would you say is the biggest misconception about autonomous coding, right? Is it going to replace coders? That's yeah... I hear that question all the time.

[TEMEKA] We do. And so here's what I'll say. When people hear autonomous coding, it instills a little bit of fear because some people are of the mindset that AI can do it all. And I am here to tell you, only having been working with this specific part of technology for the last year and a half, that all of it has limitations, right? And so, while there's going to be, again, those subset of scenarios, those optimal conditions that you can use in autonomous, there's still limitations for other areas. Now I will say, those evolve relatively quickly, I believe. You know, ChatGPT 4.1 was out earlier this year. We just released ChatGPT 5. Well, not we, but OpenAI released it in August. So, it was only four months between the most recent versions. But back to my Uber scenario. Waymo, right? If I was an Uber driver and if I was being positioned, right? to kind of be outed out of a job, what would I learn? I would learn everything that that Waymo could do and everything that that Waymo could not do. So, what I've learned thus far is Waymo can't handle inclement weather conditions. So, as an Uber driver, I might say, you know what? I'm going out during monsoon season. There's three months in Arizona where Waymo won't be as popular because of the complexity. Waymo also can't drive very long distances. So if I'm picking up at the airport, I'm taking everybody that needs to go outside of Phoenix, right? So, I am positioning myself either to work in tandem with this new AI or to work where it has limitations. So the same thing with our autonomous coding solutions. It should be on your top-of-mind to learn. What are organizations doing about autonomous coding? How are we implementing it? What specialties are on the roadmap? Which visits are we targeting? How does that data or that monitoring look? Like, could I pivot myself from, you know, maybe a simple visit coder to someone that reviews data and analytics? And APC? In the January... In January of 24, they released a completely free module, it was eight modules long, called AI in Medical Billing and Coding because they wanted all of their existing HIM professionals to get a good understanding of what autonomous coding looks like. AHIMA same time frame, about a year and a half ago, January, they released the AI Resource Hub, right? Where all of the HIM professionals now have access to webinars, right? They've got things on there that says, what do you do when AI joins your team? How do you pivot as an HIM professional? What additional roles or careers could you be looking into if you are afraid that you might be automated out of a job? So, biggest myth is that AI can do it all and that it does not have limitations. I am here to tell you that it does. You as an individual though, will have to answer that question for yourself. Like, you know your skill set, you know where you want to go next, you know what you are capable of or feel like even sometimes learning. I think back to the transition from i9 to i10. Like, I knew several people who said, you know what? I'm out of here. I'm just not going to learn this, right? And again, that is totally up to you. But I would also say be encouraged, right? We've survived all of the other tech transitions. We've gone from manual books to using encoders to using CAC to now autonomous. And so be encouraged. Don't be afraid of the new AI. Learn everything you can about it and use it to your advantage.

[LASHA] Excellent. Well said Temeka, thank you.

[TEMEKA] Mh-mh.

[LASHA] So, Lisa, what areas would you recommend exploring to initiate autonomous coding.

[LISA] So, first I want to let everybody know that I understand that this huge initiative and industry shift to autonomous coding can be really overwhelming. So, I have a lot of perspectives, I think that you should consider, where this could potentially fit into your organization. And I've grouped them into two categories. The first one is how to determine what services to consider for autonomous coding. And then other, umm, considerations that it's not too early to start thinking about. So, how to determine what services, umm... could potentially be your initial autonomous coding. Explore your CAC coding agreement rates. There's a lot of data at your fingertips. You know, what services have a really high level of accepted codes per case that you're currently seeing within CAC. Number two, you know, exploring where it may be a good fit for our organization. You know, evaluate your low complexity, low risk, but high volume of patient services that are currently being coded in HIM. And within this perspective, it would probably be a good idea to prioritize services that are already partially coded outside of HIM, like the Chargemaster, example that Temeka just gave. And HIM is only coding diagnosis codes. You know, there's probably opportunity there. And, you know, also, you know, cases that are fully coded by HIM, where are your low complexity? Um, like same day surgery, ancillary services, or a simple short-length stay in-patient cases that don't have surgery. You don't have to do all of ancillary services. You don't have to do all of same-day surgery or all of in-patient. You can pick pieces and parts that would be... umm, a good area to dip your toes into this. And when you're looking at your documentation by your providers, where do you have clear, concise documentation? So what service lines have the most robust, coding-friendly documentation? Uh, where is it very easy to find all the information that you need, meaning that technology can also easily find it and you're not doing queries and you're not, umm... having to make a lot of decisions. It's very, very coding-friendly. And if you have service lines that you don't feel like they're really there with their documentation, where can you get there easily? Um, have that coding-friendly, robust documentation with maybe some template revisions or physician education. And another area is to consider where do you have minimal abstracting requirements? Simple abstracting requirements, um, such as data service or operating physician, can be obtained by autonomous coding solutions without the coder intervention. But if you're abstracting abnormal things? Umm... autonomous coding technology may not be able to grab that for you. Things like birth weight, consulting physicians, Apgar scores, or other organizations-specific data collection practices. It's time to seriously reconsider if that abstracting is still relevant or needed at your organization, because we all know sometimes HIM practices are continued because we've always done that way, when you really don't have a need for it anymore. There is, you know, HIM used to be the source for all data collection. And there-- with all the electronic information, a lot of that data can be collected elsewhere. So those are some things to consider to... think about where can I potentially start thinking about where autonomous coding fits into my organization. But there's also other things that it's not too early to start thinking about for your success. You know, number one, like Temeka said, research and understand and learn what it is, what it can do, what are the limitations and what's realistic for autonomous coding. And that's on your own. You know, just, you know, any of the industry experts or the resources that we trust. What are they saying? What are case studies that other organizations have published? And, you know, where were they successful? Where did they have challenges? And apply that, you know, as applicable to your organization. But really understand also your leadership automation goals. You know, when you understand what they're trying to achieve, whether it's, you know, um, percent of services that are automated or... you know, staff savings, understand what they're trying to achieve, so you can create an implementation plan that's aligned with your leadership goals to ensure your success. It's not too early to start thinking about your coding staff. You know, how is this going to impact them? Communicate those organizational goals with your coding staff and include them in the plan. You know, how-- who are going to be my super users to advocate the autonomous process? Who's going to be testing the technology? Who's going to be providing vendor feedback and training their peers to facilitate successful adoption within your organization? You know, start thinking about what does the autonomous coding workflow look like for us? What should be our autonomous coding best practices? You know, from my research, what have I learned about having acceptance criteria to go to production with my vendor? Um, and also, you know, don't forget about your organizational compliance program. How does autonomous coding fit in there? Um, and, you know, this was a huge shift for all of us. And I, you know, as working with my peers and partners in this company gave me a sense of comfort, of where the industry is going. The research and understanding the experiences of other organizations has made me really, um... understand that this is doable and this is an area where you can be successful.

[LASHA] All right. Thank you, Lisa. You made some really, really excellent points, great recommendations. It's important to really think about how you can take bits and parts of different patient types. You don't have to go, like, headfirst on across the board. So that's important, and thank you for calling that out. You kind of touched on this a little bit, but what would be, like, some recommendations for managing an autonomous coding operation?

[LISA] Well, the biggest piece I can give you is data is your friend. Once you are in production and autonomous coding it's really important to understand your patient profiles pre-autonomous coding and understand what they look like post-autonomous coding. You need to create relevant key performance indicators for your autonomously coded cases. We have KPIs for our existing coding solutions. So what are my KPIs? You know, what are my autonomous coding rates by service line that I've implemented? What are my denial rates by that service line? Are my denial rates improving? Are they decreasing? You know, what does that mean? What are my compliance rates by service line? Meaning that you still need to, at a minimum, review the coding provided by your autonomous solution according to your organization's compliance plan. So if you look at, you know, 10% or a sample of 50 or whatever you're doing for your human coders, you should implement that into your compliance plan and always be monitoring what your autonomous coding accuracy looks like. And you need to have, you know, a vendor communication line that you can communicate those findings back to them, so they can use them for model improvements. So, that all comes from your data. Um, again, look-- understanding your patient population coding trends, you know, by your service lines, you know. Am I seeing my ED procedures suddenly showing procedures that we don't offer in our ED or things we typically see coded by our human coders are not, you know, being coded by autonomous. Are you seeing any new patient’s diagnosis trends or procedure trends across these services that are not seasonable or can't be explained otherwise? That-- it’s so important to understand your patient population at your organization, what your data looks like pre- and post-autonomous coding and understanding if any of those changes are warranted. And last, it's really important to monitor your regulatory agency guidelines and their workplans. I expect that work plans are going to change as autonomous coding becomes more prevalent among organizations and you have to understand their guidelines and work plans to ensure that you're not putting your organization at risk.

[LASHA] Excellent. Thank you, Lisa. So, Shannon, um... We want your input as well. So, how is Optum leveraging artificial intelligence in coding workflows? Can you speak to that?

[SHANNON] Yes, absolutely. Uh, let's say, it's just a really exciting time right now in technology and software with all of these AI improvements. And, you know, we have to look at it as an organization and say, how are we going to adopt that and pull that into our already existing application and infrastructure? And we've taken what I think is a very thoughtful approach to that. Optum's strategy around AI centers around scalability, explainability, and intelligent optimization that will deeply integrate with our clients operational systems. Specifically, Optum is going to use what we're currently calling our fusion architecture as the backbone. We're calling it fusion intentionally, because it's a hybrid approach. It doesn't use just one view of technology. All of you know that we've had our CLI, our Clinical Language Intelligence technology for a number of years to assist with that coding process. That is still a core component of and one of the components of fusion.  But fusion-- The fusion platform will bring in not only that CLI, but, you know, machine learning and deep learning, LLM-type technology to power autonomous coding across multiple system components such as facility coding, professional coding, etc. We're going to ingest that clinical documentation and automatically code that in either a full or partial format, as Temeka was talking about earlier. By the way, Temeka, I love your analogy. It's fantastic. The first time I've heard it, I loved it. And that architecture will not only be able to support those variety of case types, but it will also support case level confidence scoring, auto accept functionality, routing of high-confidence cases directly to billing. So we're looking at it not only from a code assignment standpoint but from a workflow standpoint as well. We're also going to focus on the integration of that, as I mentioned, machine learning and deep learning across the various specialties such as emergency department, hospital list, office visits, radiology, pathology, moms and babes, and short stay. We have a very ambitious roadmap that we're going after. And yes, you heard me talk about those acute and ambulatory specialties. We're tackling the entire spectrum of what your profession handles. We're looking at it holistically to bring the best possible, you know, use of this technology forward for our clients. All of our models at Optum are trained and validated within our Optum's proprietary data sets, and they're all reviewed and approved by our internal machine learning and AI review board before they're released to our client base. So we're very serious about the accuracy and compliance of our models. We will also maintain and support rule-based automation as part of our strategy or our custom configuration. That being said, you know, we've felt for a long time, and I think Optum puts the responsibility of deciding how workflow happens in the hands of a client. What do you want to go forward? What do you not want to go forward? What do you want to be seen by a human? What types of encounters are you comfortable pushing through automation with this technology. We're going to make sure that we configure rules or modules that allow you to make those decisions, and we're not making those decisions for you. It's not going to be a black box. It will allow that specific organization to automate what they're comfortable with and what makes sense for them at that given time. It will include things, not only automation, but things like wait times, documentation completeness, routing logic for exceptions, etc. So again, we're taking a really holistic approach of not just automation but automation in conjunction with workflow. We also will continue to monitor all of our-- our outputs from a quality perspective, just as we've always done. And really our overall strategy and objective here at Optum, as we're incorporating and building out our AI technologies, is to make the best operational impact possible for our clients. At the end of the day, are we going to put forward processes or technology that are going to significantly reduce you know, manual processes, workload? Are they going to improve throughput or are they going to enhance auditability or traceability, within your documents? And so those are the things that we're looking at to make sure that what we're putting forward is going to be leveraged within your organization and meet your objectives. So that's really how we're approaching AI as a strategy at Optum. It's a pretty holistic approach. And, you know, I invite further discussion and dialogue.

[LASHA] Excellent. Thank you, Shannon.

[MUSIC SURGES AND THEN FADES OUT]