Researchers from Providence Health & Services and MedStar Health Research Institute are working with Eugene, Ore.-based Ksana Health on a Large Health Behavior Model. The partners describe this as a new class of AI designed to advance the prediction, prevention, and management of mental health and substance use disorders. Staci Wendt, Ph.D., director of Providence’s Health Research Accelerator, recently spoke with Healthcare Innovation about this project, which is being funded by the federal Advanced Research Projects Agency for Health (ARPA-H).

Healthcare Innovation: Before we talk about this project, could you describe the Health Research Accelerator at Providence?

Wendt: We are part of Providence's Institute for Clinical Innovation. We were started about five years ago with the purpose of conducting innovative research and testing new care models, looking at AI technologies and conducting real-world evidence studies. We partner with a variety of sponsors, from industry sponsors to government grants, like this one from ARPA-H. Providence has the largest single instance of Epic. We have about 6 million unique patient visits and about 30 billion data points for patients. We also serve the largest number of patients living in rural areas, with about 800,000 patient lives in rural areas. One of the reasons Providence is such a great place to do this population-focused research is because we have a very large patient population and data that makes this type of research accessible.

HCI: Providence is joining a collaborative to build an AI framework around managing mental health crises. Could you talk a little bit about the problem this is hoping to solve and how using AI models in a new way might be able to help?

Wendt: We have seen breakthroughs or innovations in cardiovascular care and oncology care, but we haven't seen that type of enthusiasm or focus with algorithms in the mental or behavioral health space yet. What we would like to get to long-term is some sort of support tool that could be helpful to clinicians or to individuals to recognize a behavioral health crisis before it occurs. I'll talk about some of the steps that we're going to need to take to get there, but you could imagine a space where we are able to predict this person needs outreach immediately. We see that their physical activity has fallen off, their social interactions are not as high as normal,. This is a person who might otherwise fall through the cracks, but we could have a safety net here as kind of a helper to clinicians or to the individual saying it's time to get support to that individual.

HCI: So what is it going to take to develop something like that? It is a combination of behavioral health data from smartphones and wearables, the electronic health record data Providence has, and then these new foundational models?

Wendt: Yes, to make a predictive model that works, you need a lot of data, so that is the focus of the first part of this project. We're not even building models yet. For the next two years, we're just collecting the data to put into the modeling. Providence’s main role, alongside MedStar Health, will be working to recruit 25,000 patients who want to participate in this type of research, We'll be asking patients to download Ksana Health’s app, which will be gathering information about their daily behaviors. You can think of it as like a diary. We’ll be linking those data, with patient consent, to health records.

Then if we're able to recruit those patients, then we'll move forward into the modeling phase, and we'll be working with our partners at University of Washington and the data science teams from Ksana and Providence and MedStar Health to work on building those algorithms. Then we will deploy in a small pilot and and check if this algorithm is working before deploying any further system-wide.

HCI: So you’d have to decide what's the most valuable way to signal something to either the clinician or the individual that you're recognizing a pattern about the person's behavior or health that would help to either escalate notification to a care team or to the individual….

Wendt: Yes, exactly. You could think about it as a way to connect patients to care before things reach a crisis level. From a population health strategy point of view, it would be reducing risk and getting people connected to the right care at the right time before it's something that's more intensive or higher cost for the patient.

HCI: Are there already clearly defined things about patterns of people's physical activity, sleep, mobility, social connection that when they hit a certain threshold it is bad, or are you going to have to figure some of those things out as you go along?

Wendt: I think we're going to have to figure some of those things out. I think there's going to be a lot of individual variability, so one of the things with the modeling that will be innovative is that hopefully we can build a model that evaluates the person based on what their typical behaviors are, and where those start to fall out of a typical cadence. We know from the existing behavioral health research that things like reductions in physical activity, or not wanting to have social connections can be tied to loneliness, can be tied to depression, but I think what we are missing and hoping to get here is more of that individual variation in prediction, and being able to build that for people uniquely, and understand the pattern of that happening. What are the first signals that we start to see?

HCI: Would all of this work not have been possible before these breakthroughs in foundation model architectures?

Wendt: No, I think that this is totally possible because of these foundation architectures. Another thing is that we are thinking completely outside of the box. I mean we’re shooting for the moon here, and I think that's really made possible by the funding from ARPA-H.

HCI: What are the next steps? Will we hear about incremental steps in this work?

Wendt: This funding came through April 1. One of our first milestones, which we’ve already met, was to create a lived experience and ethics panel that includes patients and community members who have lived experience with receiving mental health services. We recruited individuals across Providence and the MedStar footprints to participate in the panel, which will be in place across the lifetime of our project. They'll be providing feedback on the work that we're doing, and especially when we get into the model development phase. 

Our next milestone will be getting together our ethics review board protocols for patient recruitment, and that's a big undertaking. We want to be transparent with patients and consent patients into the research study portion of this work, so that will be overseen by Providence's Institutional Review Board. We'll be aiming to submit that protocol package in June, and we'll be aiming to recruit our first 200 patients in the next year. This is really like a proof of concept for the project to make sure there are systems in place and that we can recruit these patients and have he data that we're looking for.

After that, our next big milestone will be the recruitment of those 25,000 patients. We expect to be able to do that within two years, and then hopefully we will receive additional funding to do the modeling work.

HCI: And Providence and MedStar have to do this work in parallel within your own health systems, right? 

Wendt: No data is going to cross between the health systems, but we're on the same trajectory. That’s been really nice, because we're able to think about the best practices and strategies that we have. I think having both of these health systems as part of this group, we do have coverage nationally for the patients that will be represented in this research study.