Episode Transcript

Welcome everyone to another episode of compliance conversations. 

I am CJ Wolf with Healthicity, and I'm really excited to speak to our guest today, Nicola Sahar. Welcome, Nic. 

Thank you, CJ. I'm excited to be here. 

Yeah. We met, at an AAPC conference. I think it was the regional conference in Philadelphia maybe. 

Yes. 

Yeah. And, got to know you a little bit and and your expertise. And so I'm really excited to have you share a little bit about what you've developed and and how it, applies to kind of coding and billing and compliance worlds. But before we do that, Nick, we'd love to have our guests just kinda introduce themselves a little bit and share anything that you'd like to share about yourself. 

Awesome. Yeah. My name is Nicola Sahar, CJ mentioned, I'm a medical doctor by training from the University of Toronto and actually spent a lot of my medical training looking at, AI and machine learning. I did research at the university to build models that can better understand clinical text and started to, deploy these models in hospitals and, health system settings so that we can better start to structure our clinical data. This was back in, twenty seventeen, twenty eighteen. So before things like CHAT JPT came out, this was the type of research I was diving into. 

And and really looking at the where AI was heading was was really exciting back in that day because, you know, we we did have a hunch that it was going to be very impactful technology, but we obviously didn't know it was gonna take off this fast and this this wide across, many industries. 

Yeah. So so short end of it is I decided to put my medical training on hold and to start a an AI, a medical coding company called  Semantic Health. And our vision for Semantic when we started it in two thousand nineteen was to start to apply this type of technology to our revenue cycle, our coding, auditing, and CDI processes so that we can, make sure that the the medical coders and the auditors are able to review all of that messy clinical documentation, get all the codes, that they need to on the claim, and to do it as accurately and as fast as possible. 

So that's the company we've been building over the last five years now. We've, recently been acquired by AAPC, which is one of the largest organizations that credentials medical coders, and we're very excited to be, working with them and with the APC team to really continue building AI products for coders, auditors, and CDI specialists. So that's a little bit about me. 

Awesome. Thank you so much, Nick. That's it's fascinating, because you were kinda ahead of the curve, if you were doing this in two thousand eighteen, two thousand nineteen, but it seemed like the timing like, everyone that I know in coding has talked about AI, and it seemed like you kind of were ahead of a few years developed this. 

Yeah. 

And it's really taken off. And and so, yeah, we're really excited. So let me just ask you quick. I know, you know, you were in medical school. 

Like, what's your technology background? Like, how do you just learn this stuff yourself? I mean, you're very smart, but tell me a little bit about that. 

Yeah, that was that that's also a funny story. I I wasn't in a I wasn't in a computer science background or degree, but I did teach myself how to code very early on in undergrad. 

And, so I was I was actually in pre med and then picked up Python was my first language, which which really is the best language for AI and and building machine learning applications. 

So in med school, I just continued diving in, and I started doing research at, the University AI Labs. So I kept building things, kept launching stuff, iterating, learning. 

So so really more self taught than, than having a, you know, a full computer science degree. 

That's awesome. Well, yeah, I I just I think that's a great combination, kind of your curiosity, your, you know, your intellectual curiosity, and and, you know, combined with your your medical background, I just think is perfect for for kind of what you've built and what AAPC is doing with it now. So let's talk a little bit about, kind of revenue cycle or, you know, what, you know, what would somebody use this for? Right? 

And who's the use Is it a hospital? Is it a health system? Is it a coder? Is it who like, who's the user, and what's the main one or two purposes of it? 

Yeah. So the the AI that we build is actually packaged into a full workflow software that's that's designed for medical coders and CDI specialists and auditors on the ground. 

And so they're the primary users. They use our software to, number one, if they're coding a chart, for example, if it's an inpatient admission and they still have not or the the the patient has just been discharged and they need to start coding the case, our system will review the clinical documentation and start to suggest, medical codes, so ICD ten procedure and diagnosis codes to the users. 

And they can work with the AI to really, accept, reject, or update these codes. So it's a seamless workflow where where the AI is taking a first pass look at the clinical notes, and then the user is, finalizing the the right codes from the AI suggestions and from maybe their own suggestions. 

That's the first use case. And, really, really, that leads nicely into the second use case, which is around auditing or checking our coded data. And the second use case is another AI workflow that takes a look at not only the clinical notes, but the answers that the coders have put down for the codes. So Okay. 

Which codes have been coded. And the goal with that AI workflow is to check the accuracy of the codes. So things like, are we missing codes? Can we be adding different codes? 

Can we be modifying codes to be more accurate with the clinical notes? Can we send clinical queries, for example, to fill in documentation gaps? And so that AI is more in charge of spell checking the claim. 

That's the second workflow. 

That's awesome. So is it so then I would imagine, you know, coders and revenue cycle professionals, like, that work for health systems that have hospitals in them, but also physician offices. I mean, are do both of those kind of settings? Yeah. Okay. 

Both settings, we've primarily focused on a lot of hospitals based settings, but we are expanding out into physician based settings. But re when we first started the company, it was it was really focused on inpatient settings. 

K. Yeah. That that's really fascinating. And, you know, I know you're you have connection to Canada. It is is it applicable like, I know different countries have different health systems and maybe different, you know, reimbursement methodologies or, you know, government sponsored plans. Is it useful in I mean, it's obviously useful here in the US, but is it useful in Canada and other places? 

Yeah. We actually started out in Canada because, that's that's really where I I did my medical training. So it's Right. It started out in Canada. 

We we launched in the US a couple of years after. And more recently, we've now launched into the Australian market. So it's really applicable in any market that requires coding to happen. So it's it's really not about the payment system as much as whether the coding is actually required and whether there there are people on the ground doing the coding. 

The payment system helps you to make a stronger business case for Right. For the software because it's not the coding is tied into payments, but you don't need, you don't need the coding to be tied into payments. 

Gotcha. So and the coding sets that we're talking about here are ICD ten. Right? 

CPT and also Yeah. ITPIX, or is it just one of those coding sets? Or tell me about that. 

We've started out with ICD and CPT. We're expanding out into HICPIX. 

Okay. 

That's awesome. Yeah. Fascinating. And and so we have a lot of coders listening, to this podcast and auditors. 

And some of the first things whenever we whenever I go to a conference and people are talking about AI, people always ask, is this gonna take my job? Is this gonna take my job? And I I don't think so. I think it probably helps you streamline your job. But you tell me, what how have you addressed that question or that concern? 

Yeah. I think I think that's a common question that I get all the time, first and foremost. Yeah. 

Is, you know, is AI going to take everything over? And my my answer to that is no. And especially in coding and auditing and CDI, AI is more of like a copilot. It's it's really designed to help you in your day to day, make the right decisions, make them faster, but it's not going to automate things. I think at the core, coding and CDI and a lot of the revenue cycle work, it's very intricate and nuanced, and there's a lot of gray area and and art to it. And Yep. The AI is not at least the AI that we're designing is not designed to automate, but instead to augment. 

Gotcha. Yeah. I mean, I I use this silly example. You tell me if it's true, but, like, calculator came out decades and decades ago. 

And did it mean we didn't need people to do math? No. It just helped you graph things and and compute quicker. Right? 

Exactly. It helps you unlock kind of a new, ability to do things in math, for example, with with that tool. 

And so AI, you can think of as the same way in in the coding space and the revenue cycle space. You're gonna be able to do more. You're gonna be able to do different things with the clinical, documents. 

And it's just really the beginning of this space of AI in general in health care. So, it might it might really unlock, like, a lot of different use cases at the clinical level, at the operational level. I I would say those are still early right now, but they they will be, coming out over the next decade. 

Gotcha. And so, from a technical standpoint, it's reviewing electronic medical records. So there must be some sort of interface. Are can it do anything with handwritten records? I know, you know, that's becoming less and less, kind of how people document in medical records. But tell me a little bit about how it actually interfaces record. 

Yeah. So we can talk through any type of standard standard API in health care. So h l seven interfaces are, like, the most common ones. 

We will use that to talk to the electronic health records like Epic and Cerner and pull out or send back data. 

And so the data itself that we can, read and analyze with the AI can be digital data. So typed clinical notes, which are pretty common these days. They could also be PDFs. 

The PDFs can be typed. It can be typed and scanned or handwritten. 

Handwritten is obviously a little bit harder depending on how clear the doc after documents. 

Sometimes it's not that great. So Exactly. It's a hit or miss with handwritten notes, but we can take a look at them and at least try to decipher them. 

Yeah. Okay. Yeah. That that makes a lot of sense. So does it, like so, like, let's say I'm a hospital system. I implement this. 

Yeah. 

And in year one, is it learning my system? And then, like, in year two, it does better. In year three, it does better. Or do you see peaks of, like, performance, or does it ever level off that you've Yeah. Are aware? 

We try to do a lot of the training upfront by looking at the data from your from your hospital, either that you've collected in the past. That really helps train the the AI to understand how things are how things are coded really at your organization. 

We also monitor the AI and apply training to it once it's live. So throughout year one, we'll be adding more and more data. You know, the new patients that are coming through the hospital will also be used to train the the system as well as the feedback from your team. So the goal is really to get it to a place where your team can start to use it from our historical data and then fine tune it and tweak it and and really customize it to your team's preferences as they give us feedback over year one. 

And that continues to improve over years two and beyond. 

But it's really a it's it's really kind of an iterative process that that involves the data that you've collected, your team. We also have a team of coders and auditors on our side that, provide feedback as well to the AI to to do more training. So we've got a few layers of quality checks before, before the results start flowing back to the hospitals. 

And we lost you, Nick. 

Hello? 

Oh, now I got you. 

Hello? Hello? 

Yeah. I'm back. Yeah. We can hear you. Yep. 

Did I cut off? 

Just for a minute. But you can just pick up from where you were talking. 

I I was just saying we've got, like, a layer of quality check with our own internal team of auditors and coders that will, essentially give feedback to the AI and course correct it, and that helps to also increase the accuracy pretty significantly before, the hospitals get their hands on the results. 

Yeah. Well, this is so cool. I could talk all day, but let's take a quick break, and we'll come back. We'll ask a few more questions, because I I think this is very fascinating. So we'll be back we'll be right back, everybody. Just give us a second. Let's do it. 

Welcome back from the break, everybody. We're we're talking to Nick about AI and coding and all sorts of things. And we were talking before the break, about how there can be incremental improvement maybe from year one to year two, as as the as the machine learning, you know, improves. But, one question I wanna ask too is so, you know, it's you said it's like suggesting codes, and then the coder kinda has to verify. 

What are you seeing as far as and I don't know if this is hard data or just anecdotal. As far as coders saying, this picks the right codes eighty percent of the time or ninety percent of the time. Do you have any of that kind of data? 

Yeah. Usually, we track, like, agreed, disagree rates, or partial agree rates. And it depends on the codes. It depends on the hospital. 

But you can expect, like, eighty percent, agreement with our suggestions. 

And then the twenty percent, are either partial agrees or just disagrees. We take a look at every one of the disagrees and really try to understand what happened and Mhmm. Either update kind of the approach internally if there's enough of these examples Right. Or or really try to reconcile. 

Sometimes it's, like, complicated coding or gray areas, so we try to dig into it with our team and the and the coders on the hospital side to to better understand these mistakes. 

Yeah. That's why I was just gonna ask about gray areas because a lot of coding is pretty straightforward, but a lot of us in the field know that there are sometimes differences of opinions. So you might get, like, one hospital system that says, oh, that and I don't know if you guys do e and m coding or evaluation and management coding where you have the multiple levels. Like, you might have one coding team say, oh, that's a level four, and one says it's level three, and they're both adamant about it, and and they wanna stick to their interpretation. Does does that come up? 

A hundred percent. And sometimes some hospitals like to do things specifically, specific to them. So they've got their own directives or own kind of policies on specific target conditions. 

And so we do take those into account where possible. Sometimes we don't know about all of them, upfront. So that's that's where part of the mistakes come from. 

Yeah. 

But but that is another, like, consideration that you have to incorporate into the systems is some hospitals wanna do things. Some coding teams wanna do things differently That's different. 

Or more specific. Exactly. 

Yeah. 

Yeah. I mean, because I I talk to coders all the time. We we can't even agree amongst ourselves sometimes. 

So that that makes a lot of sense. So, Nick, the other kind of, folks that are listening are compliance officers in health care, and and I Yeah. Wanna just tell you may already be aware of this. You know, the the US Department of Justice, publishes a guidance document. 

It's for their own attorneys that when they're investigating the company, there might be, you know, coding allegations of improper coding or something. And so the DOJ might go in and, you know, investigate their compliance program and this and that. And they just updated their, that guidance document. They've updated it a few times over the years. 

But just a couple of months ago, they just updated it, and they updated with a language about AI, about not saying that you can't use it, but they're just saying, look. Technology is advancing. And in compliance, we do risk assessments, and so we're looking at Yeah. You know, new risks. 

And so if people are concerned about AI and they might just be afraid of that phrase or something, but that that's kind of gonna be a focus, I think, in compliance a little bit. Have you interacted, like, with compliance departments as you've as you've been implementing? You know, as compliance officers, we're usually like, you know, prove it to me, right, type of attitude. 

Have you ever had that kind of interaction with compliance folks? 

Mostly when we're getting set up, I would say two ways. Mostly when we're getting set up to make sure that, they're they understand kind of how the product works and the workflow and the software works and the AI works as well so that they can they can really assess, like, more rigorously, any risks, any compliance risks with the actual technology and and the deployment strategy. 

And then secondly, indirectly, I would say through the coding teams. Like, the coding teams that that we work with will work with compliance, and they might, you know, pass on some of the audit results. For example, if there's a lot of audit findings that we are surfacing up that might be more compliance related, they may not pass them on. But we don't directly work with the with the compliance teams once we're live. It's mostly the coding teams. 

Yeah. That makes sense. But I I'm I'm thinking, like, also future, and maybe your company moves there or there's gonna be other companies out there that because in compliance, we're often looking for the needles in the haystack. Right? 

We're looking for you know, we know that the majority of doctors do things appropriately. 

Everyone's honest and but there are a few bad apples. And so we're constantly looking for those needles in the haystack, and I couldn't just imagine that AI could probably help us identify some of those, like, the patterns, right, of of poor coating or something. 

That's what we do on the audit side mostly, which is identify kind of cases with undercoding or overcoding, and we surface these opportunities to the coding team. And some of these opportunities might get passed on to compliance teams if they're, like, systemic patterns or issues that we're seeing. 

Yeah. Yeah. It's fascinating. And I just came back from a conference in Washington DC where I spoke, with Brian Burton who is here at Healthicity, and it was the the HCCA's health care enforcement compliance conference. 

So we had a lot of enforcement agencies there, Department of Justice, OIG, CMS, some state Medicaid. And and they were talking about they're like they weren't really explicit, but they were talking about using AI in their efforts to uncover Mhmm. Fraud, waste, and abuse, and those sorts of things. So I like you said, I think this is like a major shift. 

Right? It's like when a computer was created or it changed our lives. And then the cell phone, the smartphone, you know, it's changed our lives. And I think AI is gonna have a change in our lives and in all these industries. 

Yeah. 

It might it might be even bigger than the cell phone to be honest. 

Yeah. 

Or the or the Internet. Like, this is this is the next big kind of technological shift, I would say, and it's still very early. It's it's a little bit different than the Internet or cell phones where they just kind of came out and things shifted because Yep. This this is more of an iterative technology that just gets better and better with each iteration. 

Yeah. So fascinating. 

You mentioned CDI. 

Are the, for those who are listening, you know, clinical documentation improvement. 

Yeah. Are is the technology different for CDI as as for coding? Or tell me if if if that's any different or what some of the things are about that. 

CDI is really about surfacing up opportunities for us to clarify documentation so that we can be more specific with the coding. So it falls under our audit workflow when we are, when the AI on the auditing side is reviewing clinical notes and the codes themselves, we're looking for opportunities to ask questions about the clinical notes back to the physician to basically say, hey. We noticed, you know, a few signs or symptoms or a workout for this disease. Maybe it's sepsis, but you didn't really document it in your chart. 

Can you help us figure out whether this patient had sepsis or not? Yeah. And Yeah. So the AI tries to look for these types of clues scattered in the chart. 

It's a little bit different than coding because with coding Yeah. You're really looking for things that are documented, that you can code, that are very clearly, like, diagnosed or present in the patient. Right. But with CDI, it's more of like, you're investigating for things that might be present, might not be present based on clinical indicators or other things in the clinical notes. 

So the AI has to work a little bit differently to be more of an investigator versus, like, a coder. 

Gotcha. 

Yeah. And that kinda leads into, like, when clinical documentation improves. I'm thinking so I also teach, in a patient safety, master's degree program and and, also quality of care. I'm thinking AI, and I don't know if your tool gets involved in this, but AI can probably help with quality issues and patient safety issue. 

A hundred percent. We don't do as much of this, but, there there are a lot of, like, quality metrics and performance measures that are being tracked and are increasingly more important for hospitals. Like, for example, is a big one. 

We've looked into expanding into that space. It's definitely possible to start to track and flag these these measures or automatically kind of characterize them for each patient. 

But we haven't done, we haven't done a lot of of work in the space yet. I think it's an exciting space. 

So Yeah. 

I remember, like, over twenty years ago when the Olympics came to Salt Lake City, I was working for a large health system. And in preparation for the Olympics coming, we a bunch of coding and technology people got together, and I was on this work group. And they were talking about this is twenty five years ago. They're talking about, well, if we monitor all the hospitals in Utah and, you know, monitor the ER coding and see if somebody comes in with a certain, infectious disease, maybe we could catch, you know, an infection, and and and catch it from spreading. 

Right? Like, they were gonna monitor coding to see if there's an uptick in a certain kind of, you know, infection being reported or something. And that was so many years ago. I'm like, AI could, like, do that with with data all over the place and probably find things like that. 

A hundred percent. You if you have access to enough data and historical data and maybe even live data, you can start to pull out some of these trends and predictions about changes in the patient population that, are more nuanced. 

But you do need a lot of data, and you need to be tracking it over time. 

Yeah. Yeah. Like, some so for for it to really be effective with, like, public health, it it would have to have access to, like, all the data. You know? We're not in a in a shared, electronic record across the country. Right? 

Yeah. 

And so it would need access to all of those different kinda data inputs, I would think. 

Yep. Exactly. 

Well, Nick, we're getting kinda towards the end of our time here. I've asked the questions that I think I know about, but I know very little about it. Yeah. Is there anything that I didn't ask or that, you know, a kinda final parting message or thought that you'd like to share? 

Yeah. I I just wanted to say, you know, thank you for having me on, CJ. I think, like, AI is very important for us to talk about in health care. It's still early, I would say. 

Like, there's a few other industries that have more penetration, more mature maturity in in how AI is being deployed. But I think that makes a lot of sense for health care. We have to be a little bit more, rigorous and methodical with how we evaluate where we wanna deploy AI and how we want to do it. And the revenue cycle and things like compliance are two really good starting use cases, to to allow us to really see the potential and and the benefits of AI. That's what we've been focusing on, and I think there's there's really gonna be a lot more changes to, health care overall as AI as a technology gets better and better. And I'm I'm excited. 

And, and, you know, I think, like, these types of conversations are important for us to really, like, get on the same page about building safe, responsible, effective AI in health care. 

Yeah. Absolutely. Well, thank you, Nick, for for sharing your expertise, and I encourage everyone to go check out, Semantic on AAPC's website and and give it a look. And, and this has just been this has been great. So thank you so much for taking the time and willing to and your willingness to share. 

Thank you so much, CJ, for having me on. 

Alright. And thank you to all our listeners for listening to another episode. 

As we always invite you, if you know of somebody that you think would be a good guest or there's a certain topic that you wanna hear more about, please reach out to us. We wanna cover topics, and speak with individuals that that you think would be beneficial to everybody. So, until next time, everybody. Take care. 

This transcript has been auto-generated. Please forgive any errors.