• Risk Stratification for Better Population Health Management

    Karen Wagner Jul 28, 2016

    At a time when managing patients with chronic conditions has become increasingly vital, organizations can take various approaches to better understand their patient populations and manage their resources more effectively.

    In 1996, when Montefiore Health System first began stratifying patients according to risk of utilization, the approach was fairly simple.

    "We used very basic stratification models based on claims data that we received from payers," says Urvashi Patel, PhD, senior director and chief data scientist with Montefiore's care management organization.

    That process in fact predated Patel's tenure at the New York City-based health system, which comprises 11 hospitals, a medical school, a school of nursing, and various primary and specialty care clinics.

    That initial approach involved aggregating patient data from a variety of sources into several databases, then stratifying those patients into different categories based on their use of services and calculating a risk score, Patel says. For example, a patient visiting the emergency department (ED) multiple times for chronic conditions would be given a higher risk score than someone visiting the ED for appropriate care, such as a sprained ankle.

    Today the health system takes a multipronged approach to risk stratification that includes rigorous analysis using statistical modeling, Patel says. But that basic method still plays a role.

    "As we gained more experience, we started to bring in additional data elements, such as electronic medical record (EMR) data, when they became available," she says. The organization also takes into account nonclinical data, including health risk assessment information gathered internally, such as demographic data on financial and housing status, Patel says.

    Managing patients with chronic conditions has become a central strategy in population health endeavors and other efforts to optimize the quality of care. Risk stratification enables providers to better understand their patient populations and manage their resources more effectively.

    Montefiore uses risk stratification for several of its patient populations, including patients with heart failure, end-stage renal disease, chronic kidney disease, asthma, and chronic obstructive pulmonary disease (COPD). The tools help identify the right programs for subgroups of patients. With its diabetic population, for instance, Montefiore has different care management approaches for patients who manage their diabetes well and those who do not.

    "You want to correctly identify and assist patients based upon their specific needs," Patel says.

    Begin at the End

    For those just entering the field, risk stratification can be more than a bit daunting. (For a breakdown of the most prominent risk stratification models, see the sidebar below.)

    Sidebar: A Closer Look at Risk Stratification Models

    Patel says it is important to first get a feel for the data by understanding what information it can provide and how that information can support the organization's population health goals. Regardless, providers should not feel overwhelmed by all the choices, she says.

    "I would start with the basics, get an understanding of the data, and organize it in a way that will allow you to extract it easily and understand it when you're giving it to the person receiving the data," Patel says. "You don't have to conduct fancy, rigorous, statistical models to start."

    A good way to begin the process is by identifying the intended outcomes, Patel says. If the goal is to reduce readmission rates, for example, the risk stratification model should incorporate data that can impact that goal. She cautions against incorporating too many variables of risk, which can result in excessive data. If the goal is to reduce A1C levels in a diabetic population, the model should include a sufficient number of variables (typically four or five, but possibly more depending on the model) related to the risk level of a person with diabetes, such as measures of how often A1C levels are tested and the number of ED visits within a certain time frame.

    "You don't want to put in extraneous variables that you may not be able to measure the impact of," Patel says.

    Simple Methods Can Suffice

    The tools used in risk stratification do not have to be overly complex or highly technical. In fact, in the beginning simple methods may be more effective.

    The initial risk identification process is a lot like picking fruit, says Ladd Udy, director of population health and accountable care organizations for Mercyhealth, Rockford, Ill. The first step is picking the fruit that has already fallen to the ground—it is easy to reach, Udy says. The next step is picking the low-hanging fruit, which is a little harder to reach. Finally, there is the fruit on the hard-to-reach branches, which requires extra effort.

    "If you're just getting started, do you need to climb to the top branch to shake it to get the fruit that's all the way up there? You can work up to that point," he says.

    At the start, one required resource may be an analyst who is well-versed in working with databases, importing claims data, and then exporting and formatting the data into reports for final users, Udy says. "You can do a fair amount with fairly basic tools," he says.

    Provider organizations also have an in-house risk stratification source that is easy to apply and should not be overlooked: physicians. Paul Takahashi, MD, professor of medicine at Mayo Clinic College of Medicine, Rochester, Minn., says one of the most common ways to classify patients is by asking physicians which patients use a lot of services and who they consider to be high-risk. Incorporating physician input is important because the risk models are not perfect. "There should always be room for both," Takahashi says.

    What the models often miss that physicians may recognize are the socioeconomic factors that can significantly impact the patient's risk category, Takahashi says. The lack of a caregiver or of financial resources to buy medications may cause patients to seek care in the ED or their conditions to needlessly worsen, for example.

    Nadya Doll and Kristen Goelzer, Mercyhealth

    Nadya Doll, left, RN care coordinator, and Kristen Goelzer, MD, internal medicine physician, Mercyhealth, review information about upcoming appointments. Mercyhealth stratifies patients based on risk of readmission. (Photo: Lucas Blahunka/Mercyhealth)

    No Single Answer

    In all likelihood, no single model will be used to risk-stratify the various patient populations of a provider organization.

    "We're finding that a single approach doesn't quite do it, because different models work better for different situations, and you might have different sources of data for different populations," Udy says. Hypothetically, he says, an organization could use an EMR-generated risk score to risk-stratify patients using a "cutoff" score. If that resulted in too many patients to manage, a next step could entail comparing those patients with claims data to see what the Hierarchical Condition Category (HCC) scores were for those patients, or where the highest spend was, and whittle the list to a manageable number that way. "They could also do that in the reverse order," he says.

    Sometimes it takes several iterations of the data to glean actionable information. For example, in trying to identify high-risk Medicare beneficiaries, Udy and his team created a registry of all patients with two or more chronic conditions who had a physician visit in the prior three years. As it turned out, this list included the majority of the 11,000 patients who received care in Mercyhealth's Medicare Shared Savings Program (MSSP) accountable care organization (ACO)—not helpful as a means of directing care management efforts.

    Udy and his analysts honed the number of patients in the cohort by including only those with specific conditions, such as heart failure, diabetes, asthma, and COPD. They narrowed the list further by soliciting provider input as to which of the highest-risk patients they thought required more care management.

    "There's so much data available that to get to what you actually need, you have to go through several rounds of whittling it down—unless you have unlimited care coordination resources, which is unlikely for most organizations," Udy says.

    Health Plan Data: Invaluable

    The ability to obtain useful data is vital to risk stratification. Provider organizations by themselves generally do not have access to all the data that can be instrumental in assessing high-risk patients. Health plans can supply additional data that may prove crucial. Recently, the Centers for Medicare & Medicaid Services (CMS) decided to provide HCC scores of Medicare beneficiaries to MSSP ACOs, Udy says. The scores are used by CMS to risk-adjust the spending benchmark for program participants, who in turn can use the tool as part of their risk stratification for Medicare beneficiaries, Udy says.

    Mercyhealth uses data from its EMR to identify high-risk patients. However, the EMR includes data only for the care provided within Mercyhealth's system, not by outside provider organizations. Because the HCC score from CMS includes more comprehensive data, it can provide a truer depiction of a patient's level of risk, Udy says, although he notes that the timeliness of the CMS data may be an issue.

    The comprehensive data that a health plan can provide is critical in driving population health management, says David Jeans, vice president of member management analytics in the healthcare analytics division of Anthem, which provides commercial, Medicaid, Medicare Advantage, and individual health plans.

    A Sample Population Health Management Dashboard
    A sample population health management dashboard

    "To achieve the Triple Aim, you're going to have to do a lot more to invest in preventive care," he says. "Many of these provider groups have no way of seeing all the care a patient is receiving, or for that matter they may not even see all of the diagnoses."

    One type of data that Anthem supplies to provider groups helps identify gaps in care—which may arise, for example, if members of a diabetes population do not receive appropriate nephrology or eye exams. "If the provider did not do an exam himself, he has no perfect way of knowing whether that exam occurred," Jeans says. "Whereas, from our standpoint, we can easily identify those that have occurred and those that haven't, and we can share that with the provider.

    "It's one of the advantages that health plans have always had. We have that longitudinal view of all the care a patient has received, oftentimes over a period of years, whereas individual physicians can't see more than what's in their own electronic medical record."

    Leveraging the Data

    Compiling risk data represents only part of the task of achieving better population health management. The data must be coupled with appropriate care interventions to improve quality or reduce unnecessary utilization and costs.

    Mayo Clinic, for example, has been able to reduce 30-day readmission rates by 32 percent, from a 19 percent rate down to 13 percent, by implementing a care transitions model, Takahashi says.

    Patients at high risk for readmission are identified using the Elder Risk Assessment model, which is used with older patients (see page 35). Nurse practitioners then visit these patients in their homes within a few days of discharge to provide services such as follow-up care and medication reconciliation.

    On the other hand, Takahashi says an application of risk stratification followed up with an intervention of home telemonitoring was not successful because the intervention did not include the appropriate clinical infrastructure to monitor the information supplied by the telemonitoring program. The trick is to develop appropriate care models after high-risk patients are identified, Takahashi says. "I think it has to go hand in hand."

    Making sure the risk stratification model performs as expected is also critical. Is the patient really diabetic or asthmatic? Have all patients with diabetes and asthma been identified? "That's the other part of this: You don't want to miss people who could benefit from care management services," Takahashi says. "We have to continually refine the systems to be as accurate as possible so we're taking care of the right people the right way."

    Continuous Learning

    The better a provider understands its patient populations, the better it will understand the types of data required and which risk stratification model may be useful. In that context, organizations should recognize that risk stratification is a continually evolving process.

    Anthem is expanding its consumer health profiles to provide a greater view of members, including best ways to contact patients according to their preferences. Generally, older populations prefer direct contact, for instance, while those with more education prefer mobile contact versus mail or a phone call, Jeans says.

    "I think this could be a place where we could potentially partner in the future with providers," he says. "That's an area that has a lot of excitement around it."

    Montefiore is considering using consumer data such as information from loyalty cards and similar sources for risk stratification. Insights into buying patterns, including where patients shop and what they purchase, can help inform care approaches, Patel says. For example, a person who has diabetes and purchases unhealthy foods may require additional education regarding lifestyle and diet.

    "We're still all learning, and no one has found the perfect model," Patel says. "And I don't think there is one perfect model that will meet the needs of everyone. We're trying to utilize some of these informational pieces in our stratification model to help guide our programs. That's essentially what this is about—helping support our programs and resources to work better and smarter with our patients."

    Karen Wagner is a freelance healthcare writer based in Forest Lake, Ill., and a member of HFMA's First Illinois Chapter.

    Interviewed for this article: Urvashi Patel, PhD, senior director and chief data scientist, care management, Montefiore Health System, New York City.

    Ladd Udy, director, population health and accountable care organizations, Mercyhealth, Rockford, Ill.

    Paul Takahashi, MD, professor of medicine, Mayo Clinic College of Medicine, Rochester, Minn.

    David Jeans, vice president of member management analytics, healthcare analytics, Anthem, Indianapolis.