Home-grown Data Fuels Commercial Software for Better Risk Stratification

How Delaware's Christiana Care Health System leveraged a CMMI grant to build out population health across a variety of patient populations.

In 2012, Christiana Care Health System, based in New Castle, Delaware received a $10 million grant from CMS's Center for Medicare and Medicaid Innovation. CMMI made the grant to design a new care model harnessing information technology to bridge gaps in coordinate care for chronic heart disease patients.

Recently I found out how this grant has impacted general knowledge of technology-enabled care coordination. I spoke with Terri Steinberg MD, chief health information officer and vice president of population health analytics at Christiana Care.

"We put another $6 million of our own money in to really stand up three components," Steinberg told me. "One was the technology to manage the population. I like to call that the population health EHR. The second was a data concentrator [and] the development of a data lake that would take data from all sources, both Christiana Care's and outside Christiana Care, so that when we ran our predictive analytics on this clinical data, we could calculate risk scores and present those high-risk patients to the care coordinators in the Medecision software, our population health EHR."

Part of the $6 million Christiana Care chipped in was to build that data lake itself – a home-grown system, Steinberg says. The third and most difficult component was implementing a comprehensive care coordination program for those patients. "That's actually the really hard part," Steinberg says. "That's the part that is really not yet well understood. Whether it's analytics or whatever, but no matter what you do, you've got to have a way to present it to human beings, and human beings have to know how to manage populations of people."

As with other population health programs I've written about, at Christiana Care, risk stratification was the key. That means getting a large enough population of patients in the information systems to tell care coordinators which patients deserved the most urgent follow-up attention after a hospital discharge,  then getting that attention right into the workflow of the care coordinator. At Christiana Care, that particular task is handled by Neuron, a predictive analytics engine sold by vendor PTC.

"When a risk model puts you into the top 15%, a task will be shown on the care coordinator's desktop that will say, Mrs. Jones has the following risk score, and here's why you need to be aware, and here's what you need to do," Steinberg says. Christiana Care's CareLink hub embraces different kinds of care coordinators: doctors, nurses, social workers, pharmacists, and others embedded in the health system's various practices.

Basically, the Christiana Care team leveraged its CMMI grant, using Steinberg's technology insights, to build out population health across a variety of patient populations, covering population health needs ranging from breast cancer screening reminders to HEDIS scores, using a variety of commercial software products and home-grown analytics and rules engines.

Analyzing Clinical Risk

Although care in Delaware is still mainly fee-for-service, Christiana Care does have risk contracts which it voluntarily entered into with some of Delaware's handful of payers. In 2012, "some people understood we needed to change care delivery, but the notion of data integration evolved at the same time the market changed, so we ended up in a place where we were really well positioned to say to a payer, let's share risk," Steinberg says.

ACO arrangements bring together claims-based data from payers and clinical data from providers. In Steinberg's opinion, not surprisingly, clinical data is a superior data source for determining patient risk accurately. Due to its immediacy, unlike claims data, clinical data is available for analysis right away.

"Clinical risk is pattern recognition," Steinberg says. "I look at the clinical components of your medical record, all of your lab results, your imaging studies, your hospital admissions, your blood pressure, your weight, anything I can get my hands on."

Analysts at Christiana Care then write risk models for particular diseases, age ranges, or populations. "The flexibility is tremendous," Steinberg says. "We are in the learning phase, so it's very labor-intensive to write these models. But the important thing is that it's pattern recognition. You look at people who did well, you look at their signals and their data points, and the machine compares."

One key is looking for risk which is rising, instead of focusing too much on patients who are so sick, that palliative care is really the only option, she adds.

Part of what allows Christiana Care's system to get the job done is giving the predictive analytics engine clean data sanitized by the system from its own data lake. "Vendors will say, just give me all your data; there's no work," Steinberg says. "That's not true, so you give them clean, sanitized data."

The ROI on all this remains a work in progress, and certainly since Christiana Care embarked on this project, the number of commercial population health and analytics software offerings has exponentially grown. I saw more than ever at HIMSS at the start of March. But it's encouraging to see pioneers such as Steinberg making what progress they have. It is a harbinger of more progress to follow.