Can big data analytics predict individualized risk of metabolic syndrome in patients and help them create personalized treatment plans? According to results from a recent study published in The American Journal of Managed Care, conducted by Aetna and a data analytics firm, it’s doable.
The study, done by Aetna Innovation Labs— where Aetna tests initiatives, such as those related to disease prediction and intervention—and GNS Healthcare, a Cambridge, Mass.-based provider of big data analytics products and services in healthcare, analyzed data from nearly 37,000 members of one of Aetna’s employer customers (who has asked not to be named) who had voluntarily participated in screening for metabolic syndrome. The data analyzed included medical claims records, demographics, pharmacy claims, lab tests and biometric screening results over a two-year period.
Metabolic syndrome is characterized by five factors: a large waist size, high blood pressure, high triglycerides, high blood sugar, and low high-density lipoprotein (HDL), considered the "good" cholesterol. Those factors can lead to chronic heart disease, stroke and diabetes. Combined, the conditions account for nearly 20 percent of all healthcare costs in the U.S., Aetna says. “All of these issues have significant medical implications, as well as utilization and cost implications,” says Greg Steinberg, M.D., head of clinical innovation at Aetna Innovation Labs. “Anything we can do identify these people and predict who might be going down these paths can be useful from a whole host of perspectives.”
Greg Steinberg, M.D.
For this study, the Aetna and GNS teams utilized two distinct analytical models: a claims-based-only model to predict the probability of each of the five metabolic syndrome factors occurring for each study subject; and a second model based on both claims and biometric data to predict whether each study subject was likely to get worse, improve or stay the same for each metabolic syndrome factor.
Like most payers in the U.S., Aetna has large amounts of data on its members, including diagnostic claims data, procedure data, medication data, and lab data—in addition to standard demographic data. What’s more, Aetna has access to biometric data for a number of its customers, including the one used for this analysis. “So we’re talking about weight measurements, blood pressure measurements, while the rest of the risk factor data comes from lab data anyway,” says Steinberg. We had all of this data for two consecutive years on the entire membership.”
Steinberg says that two models were constructed mainly because going in, the thought was that a model that utilized biometric data would likely be more accurate than a model that didn’t. And in fact, that was the case, he says. “We also were cognizant of the fact that it is difficult to obtain biometric data on all of the folks that we would like to get it on. There is reluctance for people to provide the data or employers to request it, or both. So the pragmatic reality was that we felt we needed to do it both ways,” he says.
LESSONS LEARNED
Both analytical models predicted future risk of metabolic syndrome on both a population and an individual level, with ROC/AUC (receiver operating characteristic/area under the curve) varying from 0.80 and 0.88. The researchers were able to develop detailed risk profiles for individual participants, enabling a deep understanding of exactly which combination of the five metabolic syndrome factors each of the study subjects exhibit and were at risk for developing. For every Aetna member whose data was used in the study, the researchers used a scale that measured the percentage risk that individuals had of exhibiting each of the five metabolic syndrome factors. For example, in an individual patient who exhibited two of the five risk factors, researchers could predict which third factor is the most likely to develop.
Creating those individualized predictions was the most basic and significant lesson learned, says Carol McCall, chief product officer for GNS. “For each model, you can create an individualized prediction of who is at high risk of having metabolic syndrome and developing it in the next year, and which specific factors will be the driving force behind that,” she says.
“We have been able to deliver reports at an individual level that say, for example, ‘Ms. Jones, based on your available data today, this is your metabolic syndrome risk profile based on the five factors being in range or out of range.’ And this is what we predict you will look like in one year’s time for each of these five factors,” adds Steinberg. “And in general, since things usually worsen over time, we also offer steps to take to mitigate this predicted risk, based on their own individualized plan and care management program.”
While Aetna has done this sort of predictive analytics before, the thing that was difficult about this was the degree of precision that was being identified down to the individual risk factors for each person, admits Steinberg. “Their change over time was only made possible with the way this particular model was constructed,” he says.
Another key challenge for anyone in the population health management business is poor client engagement, Steinberg adds. “Not that there is any dearth of available programs out there, but no one is using them much,” he says. “So anything that improves engagement is a good thing. Our thesis is [based on] individualizing the information. This isn’t general data like ‘smoking is bad for you and you shouldn’t smoke.’ This is what will happen to YOU, so it’s a whole different level of conversation.”
And according to McCall, this type of “hyper-individualization” becomes more possible as the data becomes “big.” There is a lot of data and many different types of variables, so while you can see where the opportunities will come from, the analytics actually get more difficult. But this model has transformed that into something that is deeply actionable and very personalized, she says.
One of the ancillary objectives of the exercise was to look at the factors individually and determine which would be most significant in terms of predicting overall future metabolic syndrome risk, and future utilization and cost, says Steinberg. “Not surprisingly, the two at the top of the list, in order, were waist circumference and blood sugar. Bearing that in mind, we now have a number of focused interventions that are aimed to helping people lose weight, based on the fact that they have at least one other risk factor that is out of range. So we target those risk factors and we are using the results to [determine] how we develop and deploy care management programs,” he says.
The analytical models also helped identify individual variable impact on risk associated with adherence to prescribed medications, as well as adherence to routine, scheduled outpatient doctor visits. A scheduled, outpatient visit with a primary care physician lowers the one-year probability of having metabolic syndrome in nearly 90 percent of individuals. “This is a $300 billion problem annually in the U.S, and those costs are incurred because of non-adherence to chronic medication,” says McCall.
Thus, this is where big data comes in, she says. And while metabolic syndrome is an issue that affects one-third of U.S. adults, this type of predictive analytics activity can really lead to useful information and development of care management programs for other conditions as well, adds Steinberg. “But it’s the personalization part that is so critical to improving engagement, which is right now the Achilles heel to care management.”
