With the onset of value-based care, machine learning is making its mark

In a value-based care world, population health takes center stage.

The healthcare industry is slowly moving away from traditional fee-for-service models, where healthcare providers are reimbursed for the quantity of services rendered, and toward value-based care, which focuses on the quality of care provided. The shift in focus on quality versus quantity also shifts a healthcare organization's focus to more effectively manage high-risk patients.

Making the shift to value-based care and better care management means looking at data sources -- the kind healthcare organizations won't get just from the lab.

In this Q&A, David Nace, chief medical officer for San Francisco-based healthcare technology and data analytics company Innovaccer Inc., talks about how the company is applying AI and machine learning to patient data -- clinical and nonclinical -- to predict a patient's future cost of care.

Doing so enables healthcare organizations to better allocate their resources by focusing their efforts on smaller groups of high-risk patients instead of the patient population as a whole. Indeed, Nace said the company is able to predict the likelihood of an individual experiencing a high-cost episode of care in the upcoming year with 52% accuracy.

What role does data play in Innovaccer's individual future cost of care prediction model?

David Nace

David Nace: You can't do anything at all around understanding a population or an individual without being able to understand the data. We all talk about data being the lifeblood of everything we want to accomplish in healthcare.

What's most important, you've got to take data in from multiple sources -- claims, clinical data, EHRs, pharmacy data, lab data and data that's available through health information exchanges. Then, also [look at] nontraditional, nonclinical forms of data, like social media; or local, geographic data, such as transportation, environment, food, crime, safety. Then, look at things like availability of different community resources. Things like parks, restaurants, what we call food deserts, and bring all that data into one place. But none of that data is standardized.

How does Innovaccer implement and use machine learning algorithms in its prediction model?

Nace: Most of that information I just described -- all the data sources -- there are no standards around. So, you have to bring that data in and then harmonize it. You have to be able to bring it in from all these different sources, in which it's stored in different ways, get it together in one place by transforming it, and then you have to harmonize the data into a common data model.

We've done a lot of work around that. We used machine learning to recognize patterns as to whether we've seen this sort of data before from this kind of source, what do we know about how to transform it, what do we know about bringing it into a common data model.

Lastly, you have to be able to uniquely identify a cohort or an individual within that massive population data. You bring all that data together. You have to have a unique master patient index, and that's been very difficult, because, in this country, we don't have a national patient identifier.

We use machine learning to bring all that data in, transform it, get it into a common data model, and we use some very complex algorithms to identify a unique patient within that core population.

How did you develop a risk model to predict an individual's future cost of care? 

You can't do anything at all around understanding a population or an individual without being able to understand the data.

Nace: There are a couple of different sources of risk. There's clinical risk, [and] there's social, environmental and financial risk. And then there's risk related to behavior. Historically, people have looked at claims data to look at the financial risk in kind of a rearview-mirror approach, and that's been the history of risk detection and risk management.

There are models that the government uses and relies on, like CMS' Hierarchical Condition Category [HCC] scoring, relying heavily on claims data and taking a look at what's happened in the past and some of the information that's available in claims, like diagnosis, eligibility and gender.

One of the things we wanted to do is, with all that data together, how do you identify risk proactively, not rearview mirror. How do you then use all of this new mass of data to predict the likelihood that someone's going to have a future event, mostly cost? When you look at healthcare, everybody is concerned about what is the cost of care going to be. If they go back into the hospital, that's a cost. If they need an operation, that's a cost.

Why is predicting individual risk beneficial to a healthcare organization moving toward value-based care?

Nace: Usually, risk models are used for rearview mirror for large population risk. When the government goes to an accountable care organization or a Medicare Advantage and wants to say how much risk is in here, it uses the HCC model, because it's good at saying what's the risk of populations, but it's terrible when you go down to the level of an individual. We wanted to get it down to the level of an individual, because that's what humans work with.

How do social determinants of health play a role in Innovaccer's future cost of care model?

Nace: We've learned in healthcare that the demographics of where you live, and the socioeconomic environment around you, really impact your outcome of care much more than the actual clinical condition itself.

As a health system, you're starting to understand this, and you don't want people to come back to the hospital. You want people to have good care plans that are highly tailored for them so they're adherent, and you want to have effective strategies for managing care coordinators or managers.

Now, we have this social vulnerability index that we have a similar way of using AI to test against a population, reiterate multiple forms of regression analysis and come up with a highly specific approach to detecting the social vulnerability of that patient down to the level of a ZIP code around their economic and environmental risk. You can pull data off an API from Google Maps that shows food sources, crime rates, down to the level of a ZIP code. All that information, transportation methods, etc., we can integrate that with all that other clinical data in that data model.

We can now take a vaster amount of data that will not only get us that clinical risk, but also the social, environmental and economic risk. Then, as a health system, you can deploy your resources carefully.

Editor's note: Responses have been edited for brevity and clarity.