Layer Health’s AI Chart Review Driving Clinical Research, Registries
A Boston-based startup company called Layer Health has developed a large language model (LLM)-powered data abstraction platform to extract clinical data from patient medical charts for data registries, clinical research and care optimization. David Sontag, Ph.D., co-founder and CEO of the company and a professor at MIT, recently spoke with Healthcare Innovation about the company’s origins and focus, as well as partnerships with the American Cancer Society and Froedtert and Wisconsin Medical College.
Healthcare Innovation: David, tell us a little about your background and the formation of Layer Health.
Sontag: I am an AI researcher, and I co-founded Layer Health about two years ago, together with several of my former students from MIT.
We’re tackling what I consider to be one of the biggest sources of friction in healthcare, which is chart review — something that really affects patients, providers, and the life science industry in a number of different ways, and something that we had experienced a lot in our academic research.
My co-founder, Steven Horng, M.D., is an emergency medicine physician, and he and I have been working together since 2011 thinking about how we can bring AI into directly into clinical settings. My students and I were all unpaid employees at Beth Israel Deaconess Medical Center in Boston, where he's a physician. We would go to the second floor of the hospital, push our machine learning algorithms into production, go to the first floor and watch as physicians and nurses used our machine learning algorithms, figure out what was working, what wasn't working, went back upstairs, changed the algorithms and iterated. In this process, we started to learn about where the real impact was going to be from using machine learning in a healthcare setting.
Beth Israel happened to be one of the few places that didn't have a commercial EHR, so we could actually make changes really easily. They've since gone to Epic. We deployed one of the first sepsis early detection algorithms based on our machine learning work, and then we discovered that that wasn't actually going to be very impactful, because although you could get a little bit better predictive performance using machine learning, the much bigger impact was going to come from the workflows that happen as a function of the prediction. How can you actually get clinicians to act on patient being high-risk? What we discovered through that work was that getting the right information to clinicians’ fingertips at the right time was where the real impact was going to be. We tackled that from a couple of different perspectives, all which led to our founding Layer Health.
One of the perspectives was by building what we called a “layer” of clinical variables that summarized what's happened to the patient in the past, what's going on with them now, and what we think is going to happen to them in the future. You could think of that as becoming a new information layer, all output by AI. Then the next generation of electronic medical software could change the way that clinicians and patients interact with the data as a function of that information. So, for example, if a clinician is going to an order screen, you can surface just the right order set for patients who have cardiac etiology, or if you have a patient being moved from point A to point B in the hospital, if that patient’s at fall risk, then you could suggest fall precautions be taken as they're being transported so that they don't break a hip. All of that information is typically information you don't have structured in the EHR in any way, but which our machine learning algorithms were able to make available in a very easy-to-use way downstream.
HCI: So that information was sitting in the unstructured clinical notes of the providers. And the trick is how to get that out and summarized in a way that's valuable to clinicians?
Sontag: It is not just about being valuable; it’s also about being operationalized. You could say we're going to use a machine learning algorithm to tell a clinician what treatment to do or what is the diagnosis. But clinicians are often very good at figuring out what the right diagnosis is if they have the right information for the patient, and clinicians are quite good at figuring out the right treatment plan if they have the relevant information and they have some idea of what the right treatment guidelines are. So it's really about surfacing that information at the right time. That is where we thought the biggest impact was going to be.
HCI: Your work also focuses on clinical research, right?
Sontag: At the same time that we were doing this work on transforming EHRs to become AI-driven, we were also doing work with the life sciences world on real-world evidence. There it wasn't a question of how you change patient care today, but how you take the data that's collected as part of the practice of medicine over the past decade and see which treatments work best for which patients. We found in that work as well that the slowest part of the process is chart review — getting the data that you could then use to drive the downstream causal analyses that are needed to come to those conclusions. So we saw this on both sides — the more retrospective work and the more prospective work. Both cases really needed this ability to do that chart review much faster, in a more scalable way, much more cheaply, and that's what we decided to build at Layer Health. So we've built a platform to do AI chart review, and we're deploying it both to drive discovery of what treatments work best for which patients, and we're using it to drive action at the point of care.
HCI: So the work with the American Cancer Society is an example of the retrospective type of work, correct?
Sontag: Right. When I think of life sciences, it is not just about pharmaceutical companies, it's also about other types of research. How do you accelerate cancer research in the U.S.?
HCI: Well, let's go into that in detail in a minute. But going back to what you said about driving change in care proactively, who are you working on that with?
Sontag: Our first customer in that space has been Froedtert and the Medical College of Wisconsin. We've been in production with them for well over a year now, really working towards that vision that I just outlined at the point of care.
HCI: So does that involve working with their chief medical information officer to implement that in the inpatient setting?
Sontag: One of the learnings that we've had as researchers is that as we think about bringing AI closer to that point of care, what's really important is how you validate that AI, how you build the checks and balances from an implementation perspective to ensure that you have a safe and trustworthy deployment.
Because of that, we actually chose an interesting route for deployment in health systems of working toward that vision I outlined to you, but actually taking a step back and first deploying them for quality improvement purposes. So what health systems today are using our product for actually very closely mirrors what we're doing with, for example, the American Cancer Society. Health systems participate today in a huge amount of chart review for assessing quality in everything ranging from cancer to cardiology and stroke surgery. Froedtert and other health systems are using our platform for what's called clinical registry abstraction to drive all those quality improvement efforts, including oncological registries where the work is basically identical to what we're doing with the American Cancer Society, except the ACS work is driving downstream research and it's working with patient data coming from all across the United States, as opposed to one or two health systems.
HCI: We have heard health systems describe manual chart abstraction as a time-consuming and expensive task. Are there other AI-based companies taking this on? Is there a race on to be the best at his?
Sontag: There have been companies for over a decade that have been building natural language processing algorithms of some kind for understanding clinical notes. What's unique about the way that Layer is tackling chart review is that we don't believe there's such a thing as a one-size-fits-all solution for processing medical data, whereas many other companies tackle this from the perspective of ‘we will structure everything in your EHR.’ We think every specific clinical area and every use case is going to need a slightly different set of things. So we've built our platform with that goal in mind.
And actually, our clinical registry work and our work with the American Cancer Society are both really good examples of that. We think of our platform as a horizontal platform. It enables tackling chart review for many different use cases. Even though we have first deployed modules for clinical registries, it is much, much broader. For each of the use cases, one can use our platform to give a set of tasks or definitions — like these are the 300 questions I need to answer for this use case for each patient. And those questions can be very complex. For each question a clinician would be using the guidelines to read through the patient's medical record, to answer the questions in a very subtle way. We import those guidelines into our platform, and then we answer those specific questions. What that means is you can really go the last mile. Each of the clinical registries in areas ranging from oncology to neurology to cardiology all have a different set of questions, and we can ingest those into our platform, which then builds models to answer those specific questions, and then it makes it really fast for domain experts to validate that the models are doing the right thing.
HCI: Can you talk in a little more detail about the work with the American Cancer Society?
Sontag: The American Cancer Society has been running their cancer prevention studies for many, many years now. It's a really interesting, really inspiring study where they follow hundreds of thousands of Americans over decades. They send out surveys to them every couple of years to find out who's developed cancer. When they find out that someone has developed cancer, with the patient's consent they pull their medical records, and they go get biopsies of the underlying tumor.
They then do genomic sequencing of those tumors, and they do data curation of the patient's medical records. What they get out is an incredibly rich data set of underlying biology around the patient's cancer and the patient's cancer outcomes over time. That is then used to drive cancer research in the U.S.
They've been doing this process of the curation manually for the last several years, and it's been very expensive and very slow. Some of those questions are completely analogous to the questions you would have to answer for a cancer registry or tumor registry on the health system side, while some are very specific to the research needs of both the American Cancer Society and all the researchers whose research they support. Even for questions that are unique to the American Cancer Society's research needs, our platforms has been able to do the AI-guided chart review and provide good answers to them.
HCI: I read that you did a pilot with them and once they determined that you hit the mark on what they were looking for, now you're expanding it out to their other research projects. Is that right?
Sontag: Correct. We had done a pilot in for a limited number of patients, and it was highly successful. We were able to completely re-envision the work flows from the bottom up, and we were able to get the full set of results back to them within hours of them sending us the initial patient data.
Now we’ve expanded the collaboration to all of the cancers that they have to do data curation for and it's a multi-year partnership where we will be doing all the data curation for these cancer prevention studies, and then potentially other studies as well.
HCI: How do you work with partners like the American Cancer Society to validate the results?
Sontag: In our pilot, we sent them back our results of doing chart review for a number of different questions that they had to answer. This was a setting where they had previously manually done the chart review so they already knew, or thought they knew, what the right answers should be. We didn't know that on the Layer side, but we sent them back the results, and they were able to compare our algorithm’s output to what they had previously manually abstracted. Now AI is not perfect, and we did make mistakes, but so did the human abstractors. When the American Cancer Society adjudicated the differences, they found we were at or exceeding human-level performance.
We're also doing that kind of validation in our work with health systems’ clinical registries. We think that this validation process is really important to build that foundation of rigorous performance measures that will enable trust in this AI.
HCI: Your company also says it can help streamline healthcare operations and financial performance. What are some ways that could help in that realm?
Sontag: As I think about the friction that appears in healthcare, it occurs in everything from the back office to front office. You see it on the revenue cycle management side with things like coding and prior auth. At Layer we are using our platform, which is very much a horizontal platform that goes across these use cases, to get rid of each of these different sources of friction. Examples could be in care optimization: in discharge planning, which skilled nursing facility should a patient be discharged to? In hospital-at-home programs, which patients are good candidates for it? Or which patients could get surgeries in ambulatory surgical centers vs. in big hospitals? In each of these settings, there are questions that are like inclusion/exclusion criteria in a clinical trial. Our platform makes it really easy to surface the right information to make those site-of-care decisions optimally.
HCI: Did you get venture funding to spur all this work you’re doing?
Sontag: We raised a seed round and announced it in 2023 from Google Ventures and from General Catalyst, and also from Froedtert, which has an investment arm called Inception Health. We are very interested in health systems becoming part owners of Layer Health because we have a very long-term goal of really transforming healthcare and improving patient outcomes, and we think that can only be done in very close collaboration with health systems. So that's what we've been using to kick-start our work. We will be looking for additional funding very soon.
