More Data in Health Care Will Enable Predictive Modeling Advances

Rich data streams from electronic health records, when used appropriately, can transform population-based care management

David Bodycombe

Predictive modeling (PM) has grown to be a linchpin of care management. Health plans, integrated delivery systems, and other health care organizations (HCOs) increasingly channel their patients to interventions based in part on what they deduce from predictive models that have traditionally been run against databases of administrative claims. In this arena, the Affordable Care Act (ACA) is likely to exert a profound effect. ACA-influenced trends that will affect how HCOs use predictive modeling in the coming year include:

  • Increased adoption and use of electronic health records (EHRs), including enhanced access to EHR data, such as family history, lab results, prescription utilization, prior successful/unsuccessful treatments, and treatment compliance
  • Enhanced emphasis on population-based patient care management
  • An intense focus on clinical, operational, and financial performance and cost rather than cost alone

Let’s look at the implications of these trends for predictive modeling and population health management.

With the rapid expansion of EHR implementation, which extends through accountable care organizations (ACOs), medical homes, and individual practitioners’ offices, the ability to gather data that can enhance predictive modeling performance has improved dramatically. Growing numbers of HCOs, including physician practices, are routinely utilizing EHRs. In fact, close to 40 percent of physician practices use some form of an EHR system, according to an annual survey from the Center for Health Statistics of the Centers for Disease Control and Prevention.

Other HCOs have experienced similar results in their physicians’ utilization of EHRs. Hospitals and health systems are more than halfway along in their adoption nationwide, although many face challenges associated with “meaningful use” requirements for federal funding. Still lagging larger physician groups and systems are small and rural health care providers. However, a growing number of health care experts, including the Care Continuum Alliance, see predictive modeling as an opportunity to prevent complications, control readmissions, generate more precise diagnoses and treatments, predict risk, and control costs for a more diverse array of population segments than previously attempted.

PM in real-time and on the fly

The process of risk assessment and prediction will become increasingly flexible, fluid, and dynamic. As HCOs implement systems that are capable of adding deeper, more compelling data to EHRs, clinicians and public health officials will be better positioned to develop and implement more focused predictions affecting care interaction. For example, real-time access to drug databases could reduce the number of prescriptions issued for commonly misused drugs, according to a 2012 study in the Canadian Medical Association Journal (link is external). HCOs will increasingly rely on predictive modeling to review and re-engineer care processes, possibly leading the Food and Drug Administration to become more involved in the oversight of predictive models.

Increased availability

New data streams will become available to providers, payers, and government as EHRs draw from a broader array of data to create more complete insight into patients and the care delivery process. Current data used in quality reporting may be inaccurate or incomplete, according to a study published in the Journal of the American Medical Informatics Association (JAMIA). As HCOs gain access to data from more varied sources, such as health risk assessments, behavioral assessments, laboratory results, and pharmacy prescriptions (filled and unfilled), the impact of predictive modeling will increase.

Payers will ultimately organize populations into subgroups that are highly amenable to specific ­actions or interventions, predicts the author, David Bodycombe.

Further, the integration of this richer, more encompassing patient data may have a powerful effect on the quality of patient care, including the resolution of potentially confounding treatment directives and instances of over- and under-treatment. The performance of predictive models will also become increasingly more credible with R-squared values exceeding 0.30.

Progress has already been made in various health care environments. Kaiser Permanente has begun a project to link patients’ genetic data with EHRs to better understand how genes, health behavior, and environmental factors interact to cause disease.

Clinicians will soon use EHR data to track pain medication usage. This will allow for a more comprehensive approach to limiting the misuse of such medication and will enable people and organizations with access to the data to identify patients who may be “seekers” of such medications, according to a study in the Archives of Internal Medicine.

Plans and providers will tap EHRs to facilitate communication with large groups of patients — potentially in the millions, as already documented by Stanford University researchers in JAMIA.

Predictive models will expand to include a much broader array of risks, including outcomes, procedures, compliance, and safety. As an example, Columbia University and Massachusetts Institute of Technology (MIT) researchers have already used EHR data on gender, ethnicity, prescriptions, and medical history to build an algorithm that helps to predict an individual’s health. Each is placed on a risk trajectory and tracked. Advances in predictive modeling may make it possible to predict and therefore influence inflection points in this trajectory that may result in more positive population health outcomes.

New types of data will allow for a more sophisticated understanding of varied risk pools. For example, access to lab data could provide HCOs with more accurate profiles of disease severity. Meanwhile, more sophisticated health risk assessment data will give HCOs a better sense of the depth of a patient’s social network and financial resources, which often affect treatment outcomes. We will be able to develop population-based interventions that focus on commonly occurring disease patterns or phenotypes.

HCOs will ultimately organize populations into subgroups that are highly amenable to specific actions or interventions, including nutrition therapy, physical activity, weight management, health literacy tutoring, and self-management education.

This increases our ability to provide the appropriate care at the appropriate time. It also affects costs by identifying populations at the highest risk for commonly managed diseases and providing target interventions that will affect patient outcomes and quality of care. The bottom line will be much more specific and therefore effective intervention strategies.

Improved accuracy of critical data

Even if EHRs do not achieve their full potential within the next several years, HCOs can still use their nascent EHR data to enhance what is already being gathered through administrative claims. EHRs could be used to improve information capture within administrative claims, providing improved diagnostic accuracy, while systems that can take advantage of the new data streams evolve.

The growth of ACOs and patient-centered medical homes points to industrywide interest in long­itudinal, coordinated, and integrated care. Risk sharing arrangements, including contracts with ACOs, medical homes, and health insurance exchanges will demand population health management and facilitate even more rapid adoption of predictive modeling.

HCOs must be ready to quickly identify subpopulations that are at greatest risk and are open to the kinds of interventions that will control or mitigate these risks. The increased availability of EHR-derived data will help HCOs segment populations by level of risk and overcome the limitations of claims databases.

Health care reform has introduced many incentives to improve outcomes and patient safety. When the federal government launched its accountable care program, it did so with a lengthy set of measures it will use to make payments based on the quality of visits, rather than the quantity. The new norm will be to evaluate care based on value — units of quality care delivered per unit of cost.

Challenges remain. Outcomes are difficult to measure and often require expert opinion. Until recently, we have not had enough data to measure performance. This is changing. Collection of performance data is now mandated in Medicare and Medicaid and is being widely adopted in managed care. The rest of the commercial payer sector will probably follow soon, and the result will be a growing domain for predictive modeling. Access to new data sources will improve our ability to model performance.

Predictive modeling will help enhance HCOs’ clinical, financial, and operational performance, although an HCO’s ability to attain top performance goals will vary according to levels of compliance within managed populations. New sources of data will make performance modeling more accurate, timely, and realistic.

HCOs must realize that they will only succeed in population health management with shared ownership of goals with other sectors. These sectors include physicians, hospitals, payers, employers, social service organizations, and public health agencies that together will:

  • Define the issue and target population
  • Recognize external and internal baseline strengths
  • Set measurable consensus goals
  • Develop a plan
  • Implement initiatives
  • Analyze progress

The promise of predictive modeling for population health management can be achieved as long as HCOs understand its inherent shortfalls and focus on maximizing benefits such as enhanced clinical decision support and highly personalized health plans. Collaboration is key to successful predictive modeling.

David Bodycombe is assistant scientist and managing director of the ACG System at the Johns Hopkins Bloomberg School of Public Health. According to its Web site, the ACG system “measures the morbidity burden of patient populations based on disease patterns, age, and gender.”