Augmented Intelligence for Healthcare Operations with Madan Moudgal

Russ Fordyce 🏆 (00:00.866) Welcome back to the Winter Circle, everybody. I'm here today with Madan from Sagility. This is all about healthcare and how we're going to transform our healthcare operations with AI. Madan, welcome and congratulations again on the AI Excellence Award. Madan Moudgal (00:14.38) Thank you, Russ. I'm pleased to be here. We're very excited about the recent win for our nurse's dissolution. And I'm just looking forward to sharing some more information about it with you. Russ Fordyce 🏆 (00:25.014) Yeah, now Nurse Assist won an AI Excellence Award. Why don't you give us a little bit of background on Sagility and the Nurse Assistant? Madan Moudgal (00:32.354) Sure, sure. So think of Sagility as a healthcare operations company. So we provide our services to healthcare companies exclusively in the US and it includes entities like health insurers. We loosely refer to them as payers as well as health systems or providers. And we're a tech led business operations solutions and services company. And we've been in this business for about two decades now. And over the course of that time, we've obviously built a great knowledge pool in the domain in understanding the US healthcare industry and all of its intricacies. And we have brought that knowledge and that domain expertise to build a host of different solutions as we move down this journey of tech transformation. Russ Fordyce 🏆 (01:21.422) Well, I can't think of a better industry that needs it. Healthcare and the payment systems in the US is dreadfully complex. And healthcare overall is kind of, it's got to be one of the hardest industries to modernize. I've been in tech for a long time and they seem to be the slowest adopters. What makes it so complex compared to other sectors, especially when adopting AI? Madan Moudgal (01:24.066) Ha Madan Moudgal (01:29.548) Yes. Madan Moudgal (01:43.659) Yeah, yeah, you know that that's a fair point though. I have to say that I've been in healthcare healthcare technology by the way for my entire professional career, which is I hate to admit coming up on 40 years and while the industry has been a laggard in tech adoption. I don't see that to be the case, at least in the last five years or so. I think the industry is really getting comfortable with the need for more technology and more. cutting edge technology, if you will. That being said, as you can imagine, it's a very sensitive industry and it's highly regulated. And there's a lot of controls put around data privacy for obvious reasons. So that makes it a little more difficult than trying to do technology rollouts in industries that are not as regulated as healthcare. Russ Fordyce 🏆 (02:37.826) And now you've worked yourself in startups and large organizations over those decades. What have you learned that's actually helped in implementing change in healthcare? Madan Moudgal (02:47.446) Yeah, I would say that one, I've learned what are the most challenging aspects of attempting to bring about change in the healthcare industry. Some of it has to do with rather large legacy applications, which are the core administrative platforms that a lot of these healthcare entities use. And these have been developed over, I would say even decades. And so replacing one of these core admin systems is not for the faint of And as a result, a lot of these companies have continued to maintain these core legacy systems. And so anything that you do around transformation has to take into account... this legacy infrastructure and you need to work around the limitations that this legacy infrastructure puts in front of you. Russ Fordyce 🏆 (03:45.516) Now, one of those legacy infrastructure kind of processes is prior authorization. I know you guys are going after that in a big way. Why is this such a massive problem in healthcare? Madan Moudgal (03:57.251) Well, one, it is sensitive, it is controversial. As you can imagine, there have been some recent events, unfortunate events in the industry that have brought light to this particular problem. But prior authorization is also a process that has been put in place by the industry because the industry also has a fairly large fraud-based and abuse problem. And this is well known, well documented. And so what we need to do is particularly around the waste. factor, there are things that need to be done to manage these excesses and ensure that only the right care and the most appropriate care is being provided without restricting the patient's ability to get the care that they need. So it's always a fine line. But it does have significant cost implications for the industry as a whole. And so it's a particularly tough problem that the industry has been grappling with, attempting to find ways to bring about some better utilization, while at the same time ensuring that you know, on the one hand, you're trying to contain costs. On the other hand, you're making sure that you're not restricting the patient's ability to get the care that they need. Russ Fordyce 🏆 (05:27.244) Now, you guys, I was reading through your nomination and the winning nomination. Again, congratulations. You use a term that's augmented intelligence instead of automation. Why is that distinction so important here? Madan Moudgal (05:40.547) Well, simply put, human in the loop, right? We particularly in some of these use cases for which we have developed a solution, we don't believe that you can completely automate these processes because decision making, the critical decision making around these requests for authorization need to ultimately be made by a human. It would be at least currently in the industry, unacceptable that you would have a quote unquote virtual agent or a bot making a decision around the need for care, right? And so we believe that... they will need to be a human making the final decision. And so when we talk about augmentation, we're talking about making the decision process for that human more efficient and ultimately something that can deliver a better experience to all involved. And that's where the augmentation comes in. Russ Fordyce 🏆 (06:46.594) So where do you draw that line between what an AI should do and what the humans still have to own? Madan Moudgal (06:51.714) You know, it comes down to ultimately having a fairly detailed understanding of the domain and the process, knowing where the regulations stand with regard to what is allowed. for to be done by a virtual agent versus what requires a human's involvement. And by the way, that regulation landscape, it continues to evolve. There are state entities that have gotten into passing laws for what can be automated and what cannot. Of course, the CMS plays a big role. in defining some of those regulations as well. So we take all of that into account and looking at the entire process. And we find that there are plenty of opportunities for automation without having to remove the human from the process in its entirety. Now, the other approach that we take is we look at the complexity of the case that is being considered. Now, there are certain transactions that happen in healthcare where the information that's being sought, for instance, is fairly straightforward. So you can think of those as very simple use cases and then there are others requests that are more complex. So certainly the ones that are more simple and for which you know that the information that needs to be provided. does not have a bearing on whether or not the person is going to get the care, then those are certainly good candidates for automation. In the cases where a key decision is required, whether it is around access to care, whether it is around the availability of providers within a network, et cetera, where you likely need a human to be involved, in which case you would move more towards an augmenting of that human's ability to perform that service. Russ Fordyce 🏆 (09:01.612) Now, Sagility's built this nurse assist capability. Give us a little bit of preview, like an elevator pitch on nurse assistant and kind of walk us through what it's doing behind the scenes when you're processing a case. Madan Moudgal (09:14.56) Sure, sure. So, Saturity has a large team of clinical experts and RNs who provide operations and clinical support to some of the largest payers and providers in the US. So, nurse assistants and similar solutions to nurse assistants are co-created with these operational teams to ensure that they address real-world problems. And the output match the desired outcomes and standards. Nurse assistants is delivering significant value primarily because it can help interpret clinical guidelines and find evidence against those guidelines from the documents that it structures and summarizes. So the nurse's solution will extract relevant data points from case documents, and these could be medical record documents, including things like diagnosis, treatment details, pre-existing conditions, et cetera, and summarizes and structures them. And then it will use the information that's extracted, compare it to the clinical guideline related to the service for which an authorization is being sought, and will then provide a recommendation, and then the RN at the end of the day makes the decision. Russ Fordyce 🏆 (10:27.682) Well, what's really interesting to me, think you guys are one of many now that have really chosen to use kind of a domain specific model versus kind of this general purpose AI. And I liken it to, you you can go to Walmart and buy a tennis racket, but you're not going to win Wimbledon with a Walmart tennis racket. You might get pretty far, but you're not going to win Wimbledon. So walk us through kind of your decision making process on why you kind of built your own model. Madan Moudgal (10:54.166) Well, the models are ultimately need to... Madan Moudgal (11:03.366) you need to look at relevancy, right? So there are general purpose models, as you said, the Walmart type models. even today, these LLMs are getting better at incorporating medical terminology and medical context. But ultimately, what we believe is that there's a need for some specialization and an ability to focus in on the types of conversations, on the type of interactions that happen. that are very specific to the case at hand. And that's why we believe that ultimately the models that we develop basis these interactions that we take into account are going to be more reliable and more trustworthy. It's ultimately a question of trust. How much do you trust the system? And that's why we believe that it's important to focus in on these small language models or clinical language models because they will give you that higher trust factor. Russ Fordyce 🏆 (12:01.134) Well, I think you hit the nail on the head, at least from a patient standpoint, because healthcare is really about accuracy, and that's how you're building trust. Is that how you're building trust in the AI systems as well? Madan Moudgal (12:14.124) For sure, for sure. Accuracy is paramount, as you can imagine, right? So we cannot get it wrong when it comes to ultimately the interpretation of a lot of complex medical terminology, how it needs to be positioned, especially around recommendations. Because over time, a human in the loop is going to use these recommendations to make a decision, they want to over time have that confidence that they can rely upon the information that's being provided, the guidance that's being provided by the AI. And so that comes from a highly accurate model that over time is very consistent in providing the kind of recommendations for decision-making. That's where the differences come in, terms of the models that we believe that are more suitable to these types of processes. Russ Fordyce 🏆 (13:11.949) What? Russ Fordyce 🏆 (13:18.958) Well, and I imagine in kind of that domain specific model that you're building in the LLM that you're building, it's kind of easier to put in those safeguards. So what sort of safeguards are you kind of putting in place to ensure that you're compliant and that you're avoiding errors? Madan Moudgal (13:35.267) Yeah, the safeguards, we refer to them as guardrails, but yes, you're right. So there are a number of things that, for instance, one around making sure that patient information is being kept confidential. That's one of the other challenges in that you don't want to be... relying on an LLM to which you need to send information that is protected by PHI, personal health information. So we ensure that these guardrails are in place to account for that restriction on sharing PHI specific data. We also incorporate various regulations that inform us as to what is. possible and is not possible within the context of that particular process. So those are also incorporated into the guardrails we put around the AI that drives this particular process. Russ Fordyce 🏆 (14:37.346) Now, one of the reasons that you won the award was that you've reduced review time significantly. What does that mean in terms of patients kind of in real terms? Madan Moudgal (14:46.252) So when a request is submitted for an authorization of a service or a procedure or test, it will be forwarded on to the payer for consideration, of course, from the provider system. And based on the nature of the request, there are some that are considered expedited requests. meaning that they need to be decided upon in an urgent manner, and those are non-experienced. Based upon the classification of these requests, there was various handle times in which you need to respond. So the average handle time can vary, can be from a couple of hours to maybe 24, 48 hours, kind of in that timeframe. A large amount of time is spent in the traditional process, of documentation review. So when a nurse practitioner today is considering an authorization request, he or she has to go through a voluminous amount of information, particularly information that's buried within a medical record. They also then have to consult the clinical guidelines for that particular procedure and then in a sense marry the two and then based on that decide whether or not this particular request has merit and then make a decision. Now, when it comes to the review of the medical record, the nurse assist solution can significantly reduce the amount of time there are. These documents can run hundreds of pages long at times. It's like looking for a needle in a haystack because you're looking for specific references to certain medications or certain diagnosis, etc. within the medical record. Madan Moudgal (16:34.452) So the GENI solution that we built will annotate the document, will summarize all of this relevant content, and point the reviewer towards the specific information that is relevant for this particular case. And as I said earlier, we'll also consult the clinical guidelines and based upon the information that is found in the medical record, we'll provide a recommendation for the nurse to consider. And the nurse can either, you know, accept the recommendation or make a different decision. And by the way, based upon the action that the nurse ultimately takes, they will then inform the model what has happened and that will help improve on the model over time because that feedback loop is happening. We found that we can reduce the average handle time by up to 50 % in some cases, just based upon the solution that we have developed. And as I said, can take, a case could take up to an hour, 30 minutes to an hour for it to be addressed in the traditional process and we can cut that by 50%. Russ Fordyce 🏆 (17:43.552) Now, as you roll this out, imagine organizations kind of have to adapt their processes as well. Where are they seeing the biggest benefits? it cost saving speed or better outcomes or do we know yet? Madan Moudgal (17:56.213) It's early days, but I would say certainly speed is for sure. And then with speed comes the benefit of lower costs, right? The expedience too, I think is going to ultimately be improved because now you're not having to send a request in and then follow it up with another query to find out the status of the request. So. Certainly all of those have, you know, play into each other, but for sure to start with the speed is going to, and then the efficiency that we can bring about, but reducing the amount of time taken by the human in the loop to make these decisions is going to have a big bearing on this process. Russ Fordyce 🏆 (18:38.766) I imagine that one of the biggest kind of challenges, you know, is you've got these, you know, legacy systems in these gigantic organizations. How did you approach that, you know, kind of tapping in and integrating all these legacy systems? Madan Moudgal (18:56.598) Yeah, that's a great question. So for your AI solution to be effective, mean, there are a couple of elements that I find are critical. One is you need access to the data. And it's not just access. The data in some sense has to be curated. of the legacy nature of these applications, data is held in disparate systems. You have to know what If you look at a particular repository, let's say a table of information, it may give you the data that you need, but it may not be in the right context. So you need to know what it is that you're looking at, which comes from the metadata, right? And that is usually information that's held in the minds of the data analysts who actually work with that data. Now, if your AI is not informed of all of those nuances, it's just going to take the information that's been provided and say, yeah, I believe this without necessarily understanding the context. So in order to make your AI more effective, what you need to do is curate the data so that it is, in a sense, prepared for the AI to take action. So that's one effort. So when we go into some of these large enterprises, one of the things we do is we sit down and talk about the need for this data curation. Some of it will already have been done because there's been a lot of effort and particularly in the last 24, 36 months to build these data repositories. So we can tap into that. But then there's also the other thing is we work very closely with our clients around AI governance. So data curation is one, governance is the second element where we make sure we understand. Madan Moudgal (20:47.53) what their governing policies are with regard to the use of AI, right? And so that we can make sure that we're in lockstep with them and we're not being non-compliant with that. And then the third element I would say is change management, because to your point, ultimately it is a transformation of an existing process. And so you need to make sure that you've considered how this is going to impact all the others that are in the flow of this in this workflow and what they need to do differently. Because if they continue to do the things that they're doing, then it's going to be disruptive and will not give you the outcomes you're seeking. So you have to work through all of these three elements. I think in my view, I would put data curation, AI governance, and then lastly, the change management as being the three elements that. you need to tackle and tackle upfront so that ensures the success of your solutions, particularly in a large enterprise. Russ Fordyce 🏆 (21:50.83) Your career is interesting because as you said, you've been doing this a bit and you've been at that intersection of technology and healthcare, which is, you think about, you know, 30 years ago, the healthcare system and what it is today. It has completely transformed. You know, having recently been in a hospital with a loved one, you you do, you have books like this of, you know, a paper. that follow this person around. And I don't know how any doctor would ever consume that or nurse practitioner would really ever consume that in a timely manner. But you've been at this kind of forefront of innovation and risk management. How do you really kind of maintain that balance where you really want to drive this thing forward, but at the same time, know the outcomes really have to be accurate and trustworthy. And how do you really manage that risk management piece? Madan Moudgal (22:42.042) I think one of the things, at least from my experience, is that you can't rush it. You have to take your time. And I understand the desire to move quickly, especially given the pace at which these technologies are advancing. mean, you can't catch your breath nowadays because everything is happening at warp speed. But particularly when it comes to deploying these capabilities within an environment to your point, you have to take your time. You have to make sure that you do adequate testing. Some of this basic fundamentals of the software development lifecycle still are applicable. Now, you may change the way in which these the steps in the SDLC, the development lifecycle are performed. Like for instance, the testing step could be more automated than it ever was in the past, again, using GEN.AI as an agent of that transformation. But you still need to do that step. It's a very important step and you cannot assume that just because you're now in the GEN.AI world that you can miss some of these critical foundational steps that, you know, technology professionals learned 30 years ago, right? They're still relevant, they're still not just relevant, they're still critical. So you do need to take some time. And the other thing I would say is work in increments, right? It's important to do POCs. It's important to test with small batches, important to test with small scenarios and get a level of confidence in your solution before you deploy it. more broadly. So those are some things that immediately, you know, sometimes I feel like I'm being the old foggy in the room trying to keep people's enthusiasm in check. But I just think that those are very necessary elements. Russ Fordyce 🏆 (24:44.354) Well, you know, the impact of going bad is pretty bad in this environment, right? It's different when you're writing a blog or something, but in financial and healthcare terms, you know, well, it's life or death, right? Do you think healthcare will ever get fully automated if we look out, you know, 10, 15 years, or will it be kind of augmented? Madan Moudgal (24:58.444) Yes, for sure. Madan Moudgal (25:06.978) I think it's going to be augmented, know, fully automated. know, there was a funny story just. Even 20 years ago, there was talk of the Nirvana state, which was real-time claims adjudication. Why do you need to wait until after to figure out how much you owe on a particular claim, right? So real-time claims adjudication was considered something that was on the end, it was gonna happen in the next year and so on. It's been 20 years and it's still not there. And some of it has to do with just our healthcare system and the way it's structured. It is, and I don't want to go down the path of whether or not it's unnecessarily complicated, and that's for other people to debate, but it is what it is. Our healthcare system is the way it is for various reasons, and we need to work within the confines of that system. so I think that you can certainly automate a lot of areas. And the other thing is within an existing process, if you take a process from one point, A to B, and you focus on the outcome of that process. So it's just not a one point within that workflow, but the end output of that entire workflow. That, if you think about it in the context of an agentic model or an agentic framework, which is all the talk right now, there are already various agents at play in that flow, right? Some of them are actually being performed, some of those steps are being performed by human agents. Some of them are being performed by bots, let's say an RPA bot is acting upon a particular step, and others maybe by an agent that's powered by an LLM. So as you look at that entire workflow and you look at the individual steps in it, each of these can be considered a candidate. Madan Moudgal (27:07.884) for an automation and some of them may not be candidates for automation because a human is required to make, let's say a decision on that particular step, but others certainly will be. So you have to go through that process and say, okay, as I string this, break this entire flow down into its independent steps. I can then go through and say, I can automate here, I can automate here, I can use a different type of automation over here and now create a workflow that's more efficient, more accurate, provides a better experience. I think that's essentially what we need to be doing more and more of. Russ Fordyce 🏆 (27:44.706) Now this nurse assist is obviously transforming kind of that role. What areas, you know, other than this are in healthcare, do you think are ready for this kind of transformation? Madan Moudgal (27:54.851) Besides, besides nurses. So, so nurses is particularly adept at helping where clinicians are involved and clinicians are making decisions and where clinical documents are being considered or being reviewed. So that's certainly an area beyond that. If you look, for instance, on the administrative side, for instance, there's a lot of engagements that happen, interactions that happen in the healthcare space, right? So between a member as the insurance segment calls an individual rather than as a patient as referred to by a provider. So member slash patient is calling in or a payer is reaching out to that member or patient for a particular let's say for engaging with that member around their care. So each of these is an interaction. So those interactions today, a large part of those interactions are performed through of the voice channel, right? There's some that happens through web chat. There's some that comes through, it or not, faxes still, faxes, I think is the only industry that still uses faxes at the volumes that we do. So all of those interactions can, are being looked at and, you know. Russ Fordyce 🏆 (29:16.994) Lawyers. Madan Moudgal (29:27.19) there's the industry's identifying some of those simpler interactions as I referred to earlier that can be fully automated where you can deflect that request to a virtual agent to answer your question. And over time, I've gotten questions from people who said, isn't a consumer going to be upset that their... request is being routed to a virtual agent. And what I have seen is that as long as the virtual agent is answering the question and doing so accurately, I don't think the consumer minds. That's the key thing. At the end of the day, all I need is I need to know that my question has been fully answered. And if it saves me having to be on holding on the phone for the next human agent to be available to me, I'd be I prefer to just get the answer I need from the virtual agent. So Those are certainly areas where AI is starting to take hold. Then there's also a huge push in the claims accuracy area. So claims processing is a key function of the paid industry. given the volume and the dollars involved, it's one that requires a high degree of accuracy. And of course, it needs to be done in a timely manner. So there again, automation is needed. There has been, the course of the last couple of decades, more and more automation has been applied. So auto adjudication rates are in the 90 plus percentage. But there are still errors that happen. So there are services around what are referred to as payment integrity, where you're essentially trying to ensure the accuracy of the payment that needs to be made. Madan Moudgal (31:28.45) So payment integrity is a service that can save payers a lot of money. Can they save the system as a whole a lot of money so that it prevents a lot of waste, unnecessary payments or overpayments, for instance, that sometimes get made and for various reasons, right? So there again, we have found that capabilities like prediction models, and even models that are bases based on LLMs can have a role to play. Russ Fordyce 🏆 (32:04.642) Well, Madan, this is fascinating. can't imagine, you if we look out 30 years from now, where we'll be. I agree. I don't think we'll be in a fully automated world, but it's going to look a lot different. Madan Moudgal (32:14.466) That I'm convinced of, yes. I don't know whether I'll be able to recognize the industry that I started in so many years ago. And even the next five, it's going to be transformational. Russ Fordyce 🏆 (32:26.178) Well, I appreciate the time and congratulations again on award and all the hard work that you guys have put into it. Madan Moudgal (32:31.597) Thank you, Rasa, enjoy the conversation.