New York’s HealthFirst Plan Leverages AI to Predict COVID Vulnerability of Plan Members
Exciting things are happening at the New York City-based HealthFirst, New York’s largest not-for-profit health insurer, which, according to its website, offers “high-quality, affordable plans to fit every life stage, including Medicaid plans, Medicare Advantage plans, long-term care plans, qualified health plans, and individual and small group plans. We proudly serve members in New York City and on Long Island, as well as in Westchester, Sullivan, Orange, and Rockland counties.” HealthFirst covers 1.7 million members in New York state, contracts with more than 40,000 physicians and other clinicians, and with more than 80 hospitals, and has 4,500 employees.
Leaders at HealthFirst are using the data science platform from the Austin, Tex.-based ClosedLoop.ai, in order to help predict chronic and preventive care needs (such as COPD -- the leading cause of death in U.S.), prioritize outreach, and personalize their intervention strategies.
Recently, Christer Johnson, HealthFirst’s chief analytics officer, spoke with Healthcare Innovation Editor-in-Chief Mark Hagland about his team’s work in this area. Johnson has spent nearly three decades in the analytics space in healthcare; previously, he was a principal and leader of the healthcare advanced analytics advisory services team at Ernst & Young, where he led advanced analytics to help healthcare companies use predictive modeling and machine learning to gain insights into more personalized and affordable care and services. He was also a partner in the business intelligence and advanced analytics department at IBM. Below are excerpts from that interview.
Tell me a bit about your background?
I’ve been in the data analytics space for, since 1993, 28 years. The first 27 years were in a consulting environment—management analytics, supply chain analytics, customer analytics. Started out at PriceWaterhouse, it became PWC and then IBM. I started collaborating with the math department at IBM Research. I made the move to EY in 2013. Long story short, I joined HealthFirst in October of last year.
And tell me a bit about HealthFirst?
HealthFirst was created in 1993, essentially by most of the hospital systems in New York, in order to manage the Medicaid population. About 70 percent of our business is Medicaid, and 30 percent is Medicare, and the vast bulk of our business is around risk-based arrangements. When we get members from the state of New York through Medicaid, or through MA. When utilization was way down, we didn’t make a huge profit, because our providers were in value-based arrangements. And I think having this reliance of value-based care and risk-based arrangements is what’s needed in order to then leverage AI to make improvements. And our partnership with ClosedLoop is stronger because most of our work.
The ClosedLoop folks helped my team to put together a case study. We built a COVID vulnerability model last year, which we made open-source. If one of our members gets sick, who’s most likely to become hospitalized or become very sick? And we made that model available in an open-source fashion, with the help of ClosedLoop. We put this together in April and May. At that time, the CDC [Centers for Disease Control and Prevention] was stating that being over 65 or having multiple chronic conditions would raise individuals’ risk of illness if they got infected by COVID-19. We used the ClosedLoop tool to help us determine more precisely who might be at higher risk of illness.
And let me talk about some other use cases. Let’s take medication adherence. We’re trying to drive as much meds adherence as we can. And the more of your members you can get onto 90-day prescriptions, as opposed to 30- or 60-day ones, the higher level of adherence you’ll get. And prior to working with ClosedLoop, our activities were based around when someone’s prescription was due, but the level of information wasn’t deep. And the rules around when someone refills a prescription aren’t very sensitive. So we looked at SDOH [social determinants of health]-related variables, geography, past pharmacies they’ve used, and everything around their disease states and comorbidities. And we run that model on a weekly basis, and now we know the probability that every member will stay med-adherent generally and on specific prescriptions. So we know when to intervene on the behalf of members. And anytime a prescription is coming due for a member, we can look at the risk of their non-adherence at the individual level. So it helps us be more efficient in whom we reach out to. But it also helps us to avoid over-saturating members with messages. So if a member is likely to be adherent, we can just send out an email. But for the ones at risk, we can reach out with a phone call; and in some cases, we have to reach out and procure transportation; or work with a local pharmacy around potential delivery.
We have similar models around predicting the risk of readmission. So when we know someone’s at risk of readmission, we can intervene; we primarily focus on people who are under care management already. We obviously run models on those to help us know when to intervene.
How often do you intervene?
It’s very targeted intervention, to date. We’re thinking of ways to expand the number of interventions we do across different digital channels. And as we do that, a key part of our strategy is that the digital channels will allow us to intervene.
What have been the biggest lessons learned so far, as you apply these advanced analytical tools?
The explainability of the models, in the context of explaining everything to doctors, is really a central part of the equation. It’s when you start to have a conversation at the individual member level—if you consider that every week, we’re calculating the risk of a cardiovascular event—when you use a model to create a predictive model for a member, the fact that you can show a trendline for that member to someone on our clinical team or to the doctors we work with, that’s when you can drive change. And that’s so vital to getting buy-in to the use of predictive analytics. A lot of people in healthcare don’t understand what it means to use predictive analytics in an operational way. We have to bring people along, because people don’t get it. A lot of people are used to predictive analytics being a part of, say, clinical trials, where you build models and share PowerPoint presentations. But that’s not where the value is. I regularly hear, Christer, I don’t need to predict the future, I just need to know what’s happening right now. And I don’t need to hear about the past, either, so stop going back and looking at past data. But the thing is, we have to learn from past data and then operationalize the data, and look for patterns that we can use into the future. So I think most people don’t even know what to think when they hear AI or machine learning, and when they hear “predictive analytics,” they’re not understanding that we’re leveraging past data to move processes forward.
What would you say to provider leaders who are just starting to do this?
I would say that the single most important thing is to figure out how to embed predictive insights into their EHR [electronic health record] workflows. They need to focus on use cases that help doctors. A doctor possibly know everything about a member; we need to get those insights into their workflows. And people say, my data isn’t good enough for machine learning/AI. But if you don’t build models, you won’t be able to fix your data to make it good enough.
