Summary: As providers increasingly enter into contracts that reward them for improving patients’ health while controlling costs, they are making greater efforts to understand who they’re serving and engage those who may need help. This issue looks at health systems’ efforts to go beyond risk prediction to gain a more complete picture of patients’ needs and ensure no one is overlooked.
By Martha Hostetter and Sarah Klein
As health systems take on financial risk for the health outcomes and costs of designated patient groups, they often turn to segmentation approaches that allow them to identify patients who might be able to avoid health problems if given greater support.
Many hone in on patients with multiple chronic conditions, functional impairments, behavioral health conditions, and/or advanced illnesses who account for a high proportion of total health care spending. To do so, many use risk-prediction algorithms, only to find that in many cases, they capture only a small fraction of the patients who could benefit from greater oversight or help in managing their conditions.1 One reason is that many risk-prediction tools were developed for payment purposes (e.g., to predict anticipated costs) and do not take into account the socioeconomic characteristics and behaviors that have a bearing on health outcomes.
Another problem with grouping patients solely by their utilization of services or health conditions is that it assumes homogeneity within these groups when heterogeneity abounds. Two patients with uncontrolled diabetes may look the same based on lab results and hospital use, but differ dramatically in ways that determine how they will fare.
This issue of Quality Matters looks at the efforts of health care organizations to tap into other sources of information—including clinicians’ observations, data on consumer habits, and social factors—to create more nuanced segmentation approaches. Some providers are using these segments to tailor interventions for narrow subgroups that often include high-need, high-cost patients, while some are segmenting all the patients that use their systems.2 Ultimately, the goals of these efforts are to make it easier to identify and engage patients who may need help, ensure no patients are overlooked, and align resources and strategic plans with population needs.
Moving Beyond Risk Prediction
In 2012, Denver Health, an integrated health system and the largest safety-net provider in Colorado, received a $19.8 million grant from the Center for Medicare and Medicaid Innovation to support its efforts to tailor services to low-income children and adults’ particular needs. As part of this work, it began segmenting all patients the traditional way—using risk-prediction software that examined utilization patterns from claims data.3 While this approach produced a list of high-cost patients, it didn’t distinguish those who may be able to avoid health problems if given the right kinds of help.
To dig deeper, the health system developed a set of rules to supplement the predictive modeling approach. The rules—which were debated and finessed over time through a collaborative process involving clinicians, executives, health care researchers, quality experts, and IT professionals—take into account clinical indicators, medication data, and in some cases functional health status. The rules were applied to nine previously established categories of patients, ranging from healthy people to those with catastrophic conditions that include long-term dependence on medical technology (e.g., dialysis machines or respirators) or conditions that necessitate ongoing care (e.g., cystic fibrosis, AIDS, or heart transplant).
Patients in each of the nine categories are put into one of four risk tiers (see exhibit). This information is used to identify patients who may need the help of patient navigators, clinical pharmacists, pediatric nurse care managers, social workers, and behavioral health consultants. Even within the “healthy” category, some patients are considered higher risk, such as an otherwise healthy mother who has had an adverse birth event that predisposes her to another and may merit closer monitoring.
Sidebar: Attending to the Social Determinants of Health
In recognition that people’s financial and housing stability, stress levels, use of drugs and alcohol, and other aspects of their personal lives have an enormous impact on their health and ability to follow medical advice, health systems are starting to look for ways to incorporate such information in their segmentation approaches. A first step is to systematically gather and record it, which is a challenge today because many electronic health record (EHR) systems only include free-text fields for social determinants, making it hard to analyze the data. A recent report from the Institute of Medicine suggested a standard set of domains and measures of social and behavioral determinants to be included in EHRs as part of the meaningful use program—a step that could make this information much more actionable.
Another promising trend is the sharing of information across communities. The Parkland Center for Clinical Innovation (PCCI), a nonprofit that was spun off from the safety-net Parkland Health & Hospital System, is building the Dallas Information Exchange Portal, a platform that will allow hospitals, clinics, homeless shelters, food aid groups, and other service organizations to share information about the social and economic needs of vulnerable patients (who must grant permission for their information to be included).
George Oliver, M.D., Ph.D., vice president of clinical informatics at PCCI, says they hope to be able to exchange information about people’s medical care, medications, use of medical equipment, social needs, and details such as whether patients lack transportation and whether they have good support networks. The portal, which is funded by the W.W. Caruth Jr. Foundation at the Communities Foundation of Texas, may also include demographic information needed to determine eligibility for benefits.
Denver Health is now working to develop rules that take into account patients’ social and economic status (see sidebar for other examples of such work). “Doctors told us, for example, if you are a tier 2 patient and you are an immigrant you ought to be moved up to tier 3 because of the challenges these patients may face in navigating the health system,” says Tracy Johnson, Ph.D., director of health care reform initiatives at Denver Health.
Kaiser Permanente has followed a similar path to move beyond risk prediction. It began when a geriatrician, Warren Wong, M.D., realized that Kaiser’s prior methods of assigning risk—using likelihood of hospitalization scores, diagnoses, and utilization—didn’t take advantage of what physicians know about their patients’ health based on empirical observation. He asked several primary care physicians to sort their patients into four categories: those who had no chronic conditions, those with one or more chronic condition, those with advanced illness, and those who were extremely frail or likely near the end of life. Analytical experts then devised a segmentation algorithm, using EHR data to reproduce the clinical classifications as closely as possible. Rules were added to the algorithm based on whether they improved concordance with clinical judgment. For example, a rule was added to flag elderly patients who use home oxygen, which is strongly associated with COPD.
The segmentation algorithm is now used across Kaiser Permanente to inform care plans, most commonly to help determine which patients might benefit from specific team-based care and case management services, and which cases should be reviewed to determine the appropriateness of palliative care or hospice services.
Ensuring No Patients Are Left Out
In addition to focusing on high-risk patients, including those dubbed “frequent fliers” in the emergency department or those in need of support after hospital discharge, some health systems use segmentation to make sure they’re not missing patients who don’t seek out care. Bellin Health of Green Bay, Wis., looks for patients who are “missing” from the system (because they don’t visit the doctor) but may nonetheless be at high risk for problems. “We want to understand what are the issues they face, what are the barriers that they face, who are they, where are they going, what’s happening to them. And we can use that information to drive strategy,” says Pete Knox, executive vice president at Bellin.
"We want to understand what are the issues [relatively healthy patients] face, what are the barriers that they face, who are they, where are they going, what’s happening to them. And we can use that information to drive strategy." —Pete Knox, executive vice president, Bellin Health
Bellin also uses information such as where patients live, their insurance status, and whether their medical bills have been sent to a collections agency as ways to flag potential problems for all of its patients—such as poverty or housing instability—that may imperil their health. Using this information combined with data from electronic health records, Bellin has developed four groups of patients: 1) high-need, high-cost ones who make use of the health system and may or may not benefit from additional oversight; 2) those at very high risk who are not actively engaged; 3) patients at low risk who nonetheless have high spending, often because of difficulty navigating the system; and 4) patients who are relatively healthy and have little interaction with the system. “While the latter seem to be spending an appropriate amount for their risk level, we still need to reach out to them, engage them, and keep them engaged to maintain their health,” Knox says.
Similarly, Bronx, N.Y.–based Montefiore Health System uses an internally developed data analytics tool, Clinical Looking Glass, to identify patients who may need extra help ( see exhibit). It can be used by individual clinics to find out not just which diabetes patients don’t have their condition under control, but also which ones failed to return for follow-up visits. “To find people who have not come back in, not achieved their care goals, you need a cohort analytic tool that allows you to follow patients over time,” says Eran Bellin, M.D., one of the developers of Clinical Looking Glass and vice president of clinical IT research and development at Montefiore Information Technology.4
The tool has also helped Montefiore discover that some indicators are not as predictive of readmission risk as you might think, such as having H.I.V. “These patients could have compromised immune systems, but when you put the H.I.V. variable into our model, we saw it actually protects against readmission because of the AIDS clinics and other supports we have in place,” Bellin says.
Taking Patients’ Values, Attitudes into Account
Some health systems are also starting to look beyond traditional data sources in an effort to understand patients’ preferences, attitudes, and behaviors related to their health and care. Segmenting consumers by such “psychographic characteristics” is common in retail, consumer product, and financial services industries, and has been used in some public health campaigns. For health systems, such data can help explain why some patients don’t take their medications nor pursue preventive screenings or continue to smoke or remain sedentary while others follow doctors’ orders to the letter.
Using findings from its annual survey of U.S. health care consumers, researchers at the Deloitte Center for Health Solutions have divided the U.S. population into six groups based on their views on health and how they navigate the care system (see exhibit). The segmentation approach found differences related to factors such as willingness to trust a doctor vs. desire to consult a variety of sources, satisfaction with health plan and providers, interest in price and quality information, and use of digital health technology.
“Typically segmentation relies on data that are easily obtainable—demographics, insurance status, health status,” says Laura Eselius, researcher, Deloitte Center for Health Solutions. “We found each of these segments appears to some degree within each cohort. For example, not all young people are online and onboard, and not all seniors are content and compliant.” And while the segments may not show up in the same way in every health system or practice, their profiles can serve as guidelines as to differences in consumers’ preferred approaches and the likely success of outreach programs, Eselius says.
Similarly, the consulting firm c2b Solutions has developed a 12-question survey that asks people about their views on health and health care—for example whether they believe they can control their health, regardless of family history; or they believe there are benefits to alternative medicine—and then segments them into one of five groups based on what their responses reveal about how they make decisions and engage with health care providers. (See the Q&A with c2b cofounder Brent Walker.)
In a pilot effort, c2b is partnering with the TriHealth system of Greater Cincinnati to use psychographic segmentation to guide the work of its health coaches. Coaches invited participants in two of its disease management programs for TriHealth employees—for diabetes and musculoskeletal conditions—to take the survey. They’re using the resulting segmentation to help build connections with patients and motivate them to achieve their health goals. For example, one coach realized that her tendency to ask a lot of questions to guide people toward their own realizations was less effective for a segment known as “Direction Takers,” who want clear guidance and specific advice. For “Balance Seekers,” who like having lots of information and choice, coaches may spend more time talking through different treatment options.
“We, as health care providers, are used to working with patients in ways that are comfortable for us,” says Terri Hanlon, COO of TriHealth’s Corporate Health program. “We all want to do right thing, but to provide patient-centered care we need to make an effort to understand what motivates patients.”
Chris Coloian, CEO of Predilytics, a firm that uses machine learning and analytical analyses to understand consumers’ health care behaviors, says when doing this work it’s important to realize that people’s views and behaviors can and do change—particularly in response to major changes in their health. “We all need to be careful not to overgeneralize,” he says.
Indeed, the best way to use psychographic segmentation “is to have a specific target audience and a specific purpose in mind,” says Dimple Vyas, M.D., an anesthesiologist with the British National Health Service who has studied this approach. “You can then combine the information from psychographic segmentation with utilization, functional, and clinical data and some sort of profiling tool that gives you an understanding of demographics and other social factors. This gives you a bespoke segmentation model for that particular purpose.” As an example of this approach, she cites the U.K. Department of Health’s anti-obesity campaign, which used psychographic segmentation to tailor outreach efforts to the different types of families with children at risk for becoming obese.
Using Geographic Information to Understand Patients’ Needs
Some health systems are also starting to use geographic information systems (GIS)—software that presents geographic data—to better understand patients’ needs. The firm Esri markets a GIS tool that can map some 10,000 data points about consumer behaviors, shedding light on communities’ per capita spending on health care as well as things like spending on sugary drinks, caffeine, and cars. Such information can be combined with demographic information, vital records data, disease prevalence information, and health services use to create pictures of the health of communities.
Children’s National Health System of Washington, D.C., for example, used Esri’s mapping tool in combination with data from its electronic health record system to understand more about scald burn victims who arrived in its emergency department (ED). This revealed that the majority of cases originated in a number of Hispanic communities and were the result of high water heater settings. The health system then partnered with the District of Columbia’s Office of Latino Affairs to promote prevention tips—significantly reducing the number of thermal burns coming to the ED.
Southern California’s Loma Linda University Health used Esri’s software to help explain a surge in the number of people who were delivered to the hospital for 72-hour holds because of their psychological state. Mapping revealed that many such patients were coming from certain neighborhoods, and further investigation revealed that police in those communities needed additional training on how to appropriately identify people in need of hospitalization.
With no one proven model, health systems are developing their own approaches to segmenting patients. In doing so, they face significant challenges, including privacy concerns and problems accessing data that is proprietary, locked away in siloes (e.g., in patient registries), or not linked to individual patients.
For segmentation approaches to be useful, they will need to produce an enduring, valid, and parsimonious set of segments and complement the clinical diagnoses that clinicians use every day, which are themselves an empirically demonstrated form of segmentation.
"Health systems need to hone the provision of scarce resources down to the segment of the population that can best be supported. Population health management today has to be careful not to try to do everything for all. It’s not affordable and not efficient." —Chris Coloian, CEO of Predilytics
It will also be important to pay attention to the signals that patients themselves are sending. “As health systems collect more data directly from patients about their outcomes and functional status, there will be opportunities to include this in segmentation efforts,” says Clemens Hong, M.D., M.P.H., an internist and health services researcher at Massachusetts General Hospital. He notes that baseline health risk assessments provide important information, but the “holy grail” will be finding ways to track patients over time and intervene when there’s a crisis—when someone loses a loved one and stops taking their medication, for example. Real-time data from remote monitoring devices may help in this regard, he says.
Even better tracking of how and when patients interact with the health system may point to opportunities to intervene. The data analytics firm Predilytics has monitored the timing, volume, and content of calls to call centers to help health systems predict and reduce hospital readmissions. Similarly, Hong found through research that frequent “no-shows” to appointments were an independent predictor of future utilization and poor outcomes in primary care.
It also will be important not just to focus on those at greatest risk, but to find opportunities to support those who are healthy, or who may need only the right nudge to get them there. Predilytics worked with one 300,000-member health plan to hone in on a subset who were not compliant with their medication regimen, and appeared to be missing some support mechanism—say, timely reminders from their pharmacy—to comply.
“Health systems need to hone the provision of scarce resources down to the segment of the population that can best be supported,” Coloian says. “Population health management today has to be careful not to try to do everything for all. It’s not affordable and not efficient.”
1 D. Kansagara, H. Englander, A. Salanitro et al., “Risk Prediction Models for Hospital Readmission: A Systematic Review,” Journal of the American Medical Association, Oct. 19, 2011 306(15):1688–98, http://jama.jamanetwork.com/article.aspx?articleid=1104511.
2 Segmentation approaches owe a lot to the well-known example of the Camden Coalition’s Jeffrey Brenner, M.D., who used claims data for three Camden, N.J., hospitals to identify “super-utilizer” patients in need of support. See A. Gawande, “The Hot Spotters,” The New Yorker, Jan. 24, 2011.
3 Denver Health segments all of its patients, but the CMMI project is focused on patients covered by Medicaid, the Children’s Health Insurance Program, and Medicare, as well as uninsured patients. For a detailed explanation of Denver Health’s experience, see T. Johnson, R. Estacio, T. Vlasimsky et al., “Augmenting Predictive Modeling Tools with Clinical Insights for Care Coordination Program Design and Implementation,” eGEMs (Generating Evidence & Methods to improve patient outcomes), under review. Denver Health’s 21st Century Care project is supported by the Department of Health and Human Services, Centers for Medicare & Medicaid Services, Contract Number 1C1CMS331064. The contents of this publication are solely the responsibility of the authors and do not necessarily represent the official views of the U.S. Department of Health and Human Services or any of its agencies. The analysis presented was conducted by the awardee. Findings may or may not be consistent with or confirmed by the independent evaluation contractor.
4 Streamline Health of Atlanta, Georgia, now markets a commercial version of the tool, known as Looking Glass.