American Marketing Association of Las Vegas

Bringing Predictive Modeling to Health Care

As the amount of data within the health care realm increases by the minute, it can be difficult for health care systems to identify the insights that will best benefit the patients. Carolinas HealthCare System, based in Charlotte, N.C., is putting this vast amount of patient data to use with its in-house team of data analysts and health researchers. The team analyzes patient data and uses predictive modeling to assess patients’ health risks, such as likelihood of hospital readmission or chances of contracting a disease. Marketing Health Services caught up with the health system’s ;lassistant vice president of advanced analytics, John Carew, to hear more about how his team is using data to support the organization’s clinical work.

Q: How do you use the predictive models that you and your team have developed?

A: We use [them] in a number of different ways. One area, in particular, is care-management information: trying to predict risks for patients who are likely to get worse, and understand what intervention and what treatments could impact that risk for the patient.

Some examples of that are looking at patient’s likelihood to be hospitalized within the next six months, or looking at the likelihood of a patient [who might] return to the hospital within 30 days. That allows us to be more effective and more efficient with our resources because we’re prioritizing the patients who have the greatest need to help them get better outcomes.

Q: With so much data available within health care, how do you sift through it all to find what you need in order to develop your team’s predictive models?

A: A lot of it is enabled by taking a different view of data. Traditionally in health care, most data have been organized around the facility, the clinic or the physician—not around the patient as an individual. As an organization, we’ve started to make great strides in that area by building a centralized data warehouse where we aggregate all data at the level of the patient. Instead of aggregating data at the hospital side, we can look across all of our hospitals, all of our physician practices and other areas where we’ve provided care for the patient. …

What we’re trying to do is build data assets that we like to call “panorama.” It’s a patient’s 360-degree view, and really, it’s everything that we know about the patient. We use that data to predict risk. [For example], we could look at past utilization patterns: How many times has a patient been in our emergency department in recent months? We could look at gaps in preventative care for the patients—has a patient had a colonoscopy, or has a patient had a breast exam?—and make sure those gaps are filled.

We also can look at geographical data, census data and community data to look at where an individual may live and what additional risks there may be. If a patient is asthmatic, we could look and see if [he] lives in an area of higher pollution where worsening asthma may be an issue.

The other area that we’ve been exploring over the past year or so has been third-party consumer data. How can that data help us better understand the risks of a patient, as well as actually understand more about how a patient likes to interact? We know that a lot of being successful and [effecting] change for a patient involves engaging with them in a way that really resonates.

Q: How helpful is non-health-care-specific consumer data? What conclusions can you draw from it?

A: In some ways it’s very helpful. As a health care provider, we only know about the patient based on the care we’ve provided. If a patient is new to our health care system, we don’t know a lot about the patient or other utilization patterns that the patient may have. By looking at consumer data, we’re able to match that up with patients who we know well internally, and we can use the consumer data as a proxy for what some of those health care needs may be for patients who we don’t know very well. A lot of lifestyle factors and interests are probably more predictive of health outcomes than just clinical data alone, particularly in settings of chronic disease.

Q: Are some models more accurate than others?

A: There’s a broad range of accuracy in models, and it depends on a number of factors,  what variables you have available, or on the event that you’re trying to predict. Some things are very rare events, so the accuracy might be lower, and it also depends on what type of data is most relevant. So if there’s an outcome that’s heavily influenced by something that happens outside of the hospital or might be driven by things other than clinical measurements, then we probably can’t do as well at predicting the outcome. But if it’s something that relies on clinical lab values, we can generally do a better job of predicting an outcome.

We really have to take a population health perspective. A predictive model can get you somewhere. It can get you a guess as to what may be the likely outcome for a patient, but it’s not, in some cases, so precise. [There is] other information that we don’t know, but a physician may know from interacting with a patient. The physician can use that as a way to start asking additional questions, but it’s not always that we would rely fully on the predictive information. There’s no substitute for clinical judgment and talking to the patient.

This article was originally published in the January 2015 issue of the Marketing Health Services​ e-newsletter.

Author bio: Julie Davis is a staff writer for the AMA’s magazines and e-newsletters. She can be reached at

Categories: Digital/Technology,Healthcare,Q&A,Strategies.