Roanoke, Va.-based Carilion Clinic has been investing in population health management since 2006, when the then Carilion Health System transitioned to a clinic model and began to invest in developing capabilities needed to operate as an integrated delivery system, including installing an electronic medical record system.
In the eight years since, EMRs have become part of clinicians' daily workflows, and the system has gathered robust electronic records on its patients. However, this electronic data wasn't being used to its full potential as it wasn't yet being analyzed against predictive models or for real-time decision-support.
Two years ago, system leadership was ready to invest in data warehousing as a foundation for better data analysis. "As we began to move into population health, transitioning our primary care sites into patient-centered medical homes and learning more about [accountable care organizations] through our partnership with Aetna and our Medicare Shared Savings ACO, that really got us focused on thinking about what type of data we would need, and what type of analysis we would need, as we approached population health in our community," says Stephen Morgan, chief medical information officer at Carilion Clinic.
Carilion selected IBM as a partner and sought to use its software to better understand the system's patients, and better direct their care. To start, Clinic leaders identified the management of congestive heart failure as ripe for improvement. CHF patients accounted for 25 percent of the MSSP ACO's spending and impacted the system's hospitals' readmissions rates. While Carilion providers had programs they followed for managing CHF patients, there were a handful of different programs being carried out, and it was unclear which ones led to better outcomes.
Working in partnership with IBM and Epic, predictive modeling was performed on three years of patient data to identify patients at risk for developing CHF. In addition to analyzing structured data — data entered into EMR fields — unstructured data, including physician notes and discharge documents, were also analyzed, providing a more complete picture of a patient's history. The unstructured data was analyzed using natural language processing technology — the same technology used in the IBM Watson cognitive system.
Although meaningful use has led to the increase in the structured data providers are entering into EMRs, three years ago, there was much less structured data than there is today, making unstructured data even more necessary to understand the complete patient history.
Enabling access to more comprehensive data is key to unlocking the potential of predictive analytics to truly improve care processes. "The more we can liberate the information, the more exciting the possibilities are," says Sean Hogan, vice president of IBM's global healthcare division.
Additionally, healthcare providers don't have to wait until their EMRs are fully optimized to begin to analyze data to gain valuable insights about their patients' care. "You don't have to wait to derive value from [EMR data]," says Mr. Hogan. In fact, using the data quickly is important so that clinicians can see the value of an EMR install.
Specifically, the predictive model for CHF looked at physiological data such as maximum systolic blood pressure; prescription drug use of alpha blockers, beta blockers, beta agonists, and others; previous diagnoses such as chronic obstructive pulmonary disease; obesity; and lifestyle and environmental factors, such as occupation, marital and smoking status.
The analysis of the data was completed in just six weeks, and 8,500 patients at risk for CHF were identified. Carilion plans to study these patients to uncover any similarities or differences among patients at risk who eventually were diagnosed with CHF, and those that were not. "We know there was a group that had went on to develop heart failure, and a group that did not, even though they had all of those predictive indicators," says Dr. Morgan. "Why didn't they? What is unique about that group? Were they treated more aggressively?" By exploring these questions, providers can begin to understand what interventions are most effective at preventing CHF and begin to implement them throughout the organization.
Eventually, as the model is refined, providers will be alerted if patients meet the risk model and specific care protocols will be recommended, perhaps through provider alerts within the EMR.
In the future, Dr. Morgan hopes to use IBM's predictive technology to understand other disease states. "I think we'd like to continue to learn with IBM, particularly around prescriptive analytics to understand what treatments are best for what patients, based on their histories," he explains. "With genetics, some of that may come along, but for now, there's a huge benefit to using technology that can compile more information than our minds are able to, and then presenting that to physicians so they can respond. It's going to be a gamechanger."