Healthcare Innovation: Allison, before we dive into this study, could you describe the research work you do at Trilliant? 

Oakes: We publish our annual “Trends Shaping the Health economy report. We also published weekly studies that are intended to be research on a particular topic, generally leveraging our in-house data. Within healthcare, data access is often so siloed. But we have a combination of data — a national all-payer claims database, our health plan price transparency data, and also our provider directory, which we curate. That allows us to really understand that a particular service was received by this patient at this location, it was rendered by this physician, and on the commercial side, this is the associated negotiated rate as well. So we're able to get into those questions of utilization, quantity and price, which allows us to get at spending.

HCI: Who pays for Trilliant’s services — health systems?

Oakes: Health systems and hospitals are generally our primary customers, but we also work with payers and with life sciences companies. The thing that we bring to any of those stakeholders, but especially hospitals, is that complete visibility into what's going on within their market. A particular hospital or payer has pretty good visibility into what's going on within their own four walls, if you will. For instance, Penn Medicine within their own electronic health record knows what's going on when their particular patients come to them. But if they're going down the street to Temple for something else, Penn doesn't have any idea about that, right? So we're able to give these different stakeholders insight into their market dynamics and can also help them with national benchmarking and understanding broader trends in utilization and where the puck is headed.

HCI: You mentioned a national all-payer claims database. We have written about the creation of state-level all-payer claims databases. Is that where you get data from? 

Oakes: Different states have set up their own all-payer claims databases, but combining things across states can get really complicated. Instead, we’re aggregating data from a handful of national clearinghouses and also getting data directly from CMS as well. We spend our time as a business collating all of those data sources and getting them cleaned up and put together into one data set so that we're able to look at utilization for commercial, Medicaid, traditional Medicare and Medicare Advantage in the same data environment, if you will.

HCI: Well, let's turn to you recent research report about ambient AI adoption. When we talk to chief medical information officers, they're thrilled about ambient AI, and the reaction from the providers has been amazingly positive, but we haven’t asked about this increase in coding intensity as an issue. Why did the Trilliant team want to look at this?

Oakes: We’re always interested in the role of new technology and new interventions in terms of how they impact healthcare value. We think of value in terms of the amount of money we spend in this country on healthcare and what our outcomes look like. So with any new technology, we're curious about whether this seems to be improving the value of our health economy or potentially making it worse. That was our motivation in going after this topic. We focused on six different hospitals and healthcare systems across the country that had made a public announcement that they were implementing one of these AI scribing technologies. Over that study period, we did see that coding intensity increased across all six of these systems.

HCI: Do the timing of those things correlate?

Oakes: The different health systems didn't all implement them at exactly the same time, but our attempt was to get this broad signal of how it's potentially changing coding practices on the provider side of things.

HCI: Are the underlying causes of the increase in coding intensity not clearly understood? Has this not really been studied closely yet?

Oakes: I think we're just starting to put the pieces of the puzzle together. But if you think about what the goal of an AI scribing tool is, it allows for clinical documentation to be captured more thoroughly and accurately. It’s pumping more information into the patient's electronic health record. I think the promise of it is to automate processes, but the nature of these AI tools is that once the model learns the rules it's probably going to be less error-prone than humans when it comes to following established regulatory parameters as it relates to billing. 

When a patient goes for an outpatient doctors visit, whether they're a new patient or an established patient, there are these different billing codes that get used, and they vary in terms of being lower intensity or higher intensity. We wanted to get a sense of whether the proportion of high-intensity codes was changing over time. Within these six systems that we know implemented AI scribing technologies, we see that the proportion of high-intensity codes do increase over this period of time, and it's pretty significant. For new patient visits, we defined the high-intensity codes as the two most intense out of the five that exist, and we find the proportion of high-intensity codes for new patient visits increased by 12 to 20 percentage points across the six systems. High-intensity codes increased by 7 to 12 percentage points for established patient visits. And importantly, more intense CPT codes are ultimately more expensive.

HCI: So can it be assessed whether the scribing technology is just getting things down more accurately than the humans used to or whether it's actually overstating the clinical complexity?

Oakes: That’s a little bit of the nuance where we don't necessarily have hard evidence at this point to say that it's one thing or the other. However, taking a step back and looking at it logically, the nature of these AI scribing technologies is that they are rules-based. So our sense is that they are probably just improving the accuracy of provider-based billing rather than there being some major issue of fraud going on here.

HCI: I think your research paper mentions that some payers in their earnings calls have grumbled that maybe there's fraud going on here. 

Oakes: Absolutely. I think we find ourselves in the midst of an AI arms race on the provider side of things and the payer side of things. Historically, from a technology point of view, payers have been the tip of the spear as it relates to that — especially in the Medicare Advantage and risk adjustment space. There was a recent Kaiser settlement for more than $500 million and an Aetna settlement of million dollars related to fraudulent coding specifically. So I think it's a little interesting that they're the ones calling it out. 
But I think odds are this systematic increase in billing intensity that we see across a diverse set of hospitals — geographically and in terms of size — suggest that historically providers have been under-coding these visits. Our hunch is that these AI scribing technologies just follow the rules that much more accurately and consistently than human providers and human-driven billing departments had been. 

HCI: Your paper also mentions that one of the benefits here is that this is all recorded, so if you did think there was fraud, it's auditable in a much better way than previously.

Oakes: That’s one of the things that we're proposing or emphasizing. As new technologies come to market — AI scribing being a good example — transparency as it relates to how it works and why it potentially is leading to different outcomes is important to understand. If payers really think there's an issue here potentially related to fraud, we should be able to look at these AI scribing models, what exactly they're doing, and essentially run an audit to understand: do we agree that this visit that was billed at one code should be coded as a different type of visit? Or does the AI scribing model need to be better tuned in one way or another? I think the transparency element is important here and will continue to be so moving forward as these technologies continue to get that much more sophisticated and are used that much more frequently.

HCI: That raises policy, AI governance and business practice issues, right? 

Oakes: Yes, absolutely. And I think an important thing to think about as it relates to healthcare value here is that the CPT code and the billing intensity of the visit has changed, but the patient's experience of the visit itself has not. The treatment that the patient receives, the conversation that they're having with their physician — none of that has actually changed. But a visit that used to be billed at one code that cost $100 is now billing at another that cost $130. When you multiply that at scale, it can have a very real increase in the amount of money that we're spending on our healthcare system without improving anything related to experience or quality.

HCI: Well, the health systems would say that the patient experience is better in that the clinicians are more attentive during visits instead of having eyes down typing. 

Oakes: That's true. It might be a better experience, but it's probably not a better clinical outcome for the patient. There is also the important aspect of reducing administrative burden on the physicians and addressing the burnout issue. 

There is potential on the billing and revenue cycle management side of things. Maybe we can reduce some of our administrative spending if we have tools like this. As we know, we spend so much money on the administrative side of healthcare in this country. That could potentially be another positive thing, where even if we're billing more intense codes and some spending goes up there, maybe we're able to offset that and have some efficiency gains in what it takes to be billing all of these procedures. Maybe there could still be a little bit of an ROI argument there, but we haven't gotten that far yet.