business
Misfortune telling: Folding together data to forecast your patients' health futures
■ As health plans try to control costs, they use predictive modeling to target those who aren't yet sick but who might be soon.
By Emily Berry — Posted May 26, 2008
- WITH THIS STORY:
- » What goes in to the calculations
- » Related content
Using health risk assessment surveys, demographic data, and past medical, pharmacy and hospital claims, and then applying some heavy math, health plans are trying to gauge their members' future health problems. It's called predictive modeling, and plans are using it more and more to forecast medical costs.
Predictive modeling is a step above typical insurance underwriting, which assesses risk based on an insured's age, gender, location and sometimes past history.
In predictive modeling, all that information plus an in-depth look at past experience -- such as the heart attack rate for a particular population with certain chronic characteristics -- is used to discover, assess and mitigate risk not picked up by traditional underwriting.
Health plans are developing predictive modeling capabilities in-house or, like WellPoint did recently, by acquiring predictive modeling companies. Health plans and their employer customers figure that by anticipating and trying to prevent illness, they can avoid paying even more for care.
But predictive modeling is sometimes used to tier physicians, profile practices and hike premiums for sicker groups of patients.
Physicians generally have no idea this is being used on them or their patients. "It's definitely increasing as health plans are trying to become more aware of the people they're insuring, whether for identification for disease management or for sophisticated underwriting," said Russell D. Robbins, MD, a Norwalk, Conn.-based principal and senior clinical consultant for Mercer, the large human resources consulting firm.
Some experts say the biggest problem with predictive modeling is that it is being oversold as a surefire way to gauge the future. "They are getting better, but they're not yet very good," said Thomas Wilson, PhD, DrPH, an epidemiologist from Loveland, Ohio, who has both a consulting firm and nonprofit devoted to measuring the quality of quality measurements.
"It doesn't mean they're not valuable," Dr. Wilson said. "My big concern with predictive modeling, although I like it, is that its limitations are not well-communicated."
But that isn't stopping some from touting predictive modeling not only to health plans but also to physicians themselves. Companies are selling doctors predictive modeling software to tell them which patients are likely (or not) to pay their bills, and which patients are in danger of getting sick.
For now, the cost of the software puts it out of reach of smaller groups, who also might not have enough of a sample size for effective modeling. The technology is being used mostly by large independent practice associations such as Hill Physicians Group in California, which uses it to identify patients who need more intensive case management to prevent them from developing major health problems and help Hill's bottom line under its capitated contracts.
"The joke was, do we just pick up the phone and say, 'My computer says you're going to be sick next year, can we talk?' But that's exactly what the program is," said Michael van Duren, MD, Hill's vice president of clinical services.
Defining predictive modeling
Put simply, predictive modeling is examining the past to gauge the likelihood of a certain event -- a hospital stay, a car crash, a nervous breakdown. The algorithms can be used to asses groups or individuals.
"There's a broad spectrum of things that could be predicted," said Earl Steinberg, MD president and CEO of Resolution Health, a predictive modeling company based in Columbia, Md. WellPoint acquired the company in April.
"When most people use the term, they're talking about predicting expenses. They're talking about identifying who is going to be high cost in the future," Dr. Steinberg said. "That information has historically been used for underwriting, case mix adjustment, in profiling doctors to say who was going to be more expensive, and it has also been used in care management."
There is no one formula or method used universally in predictive modeling. Health plans sometimes have their own or use multiple types of software, but there also are independent consulting groups and academics that develop and sell predictive modeling services.
As the programs evolve, there are different forms of predictive modeling: rules-based models, standard regression models and artificial intelligence models, also known as neural networks, said Joel Brill, MD, an internist who is chief medical officer of Predictive Health, a Phoenix-based firm that does predictive modeling for employers, union trusts and health plans.
The artificial intelligence programs attempt to replicate the way the human brain works to tie together disparate pieces of information, factors or events, such as connecting blood pressure readings, a past history of a heart attack and a prescription for diabetes medication.
The regression models, by contrast, are based on probability mathematics, and rules-based programs, not surprisingly, are based on rules, such as if a patient has both high blood pressure and a past heart attack, the additive risk for another heart attack goes up by a certain percentage.
Who is using it and why
Use of predictive modeling in health care has grown quickly since it was first introduced in the 1980s, said Jo Anne Lutz, RN, director of channel partners for Boston-based predictive modeling firm Urix. In her post, she is in charge of relationships between her company and its clients.
"I would say any skepticism about ... predictive models and whether they were applicable to the health care payer-provider uses -- that skepticism was pretty much gone by the late '90s," she said.
But predictive modeling, even by its proponents, is not pitched as a 100% effective look at future costs and care needs.
Lutz said health plans are limited to using claims from prior years from their own members. As an example, she said Cigna would not have access to claims data from a newly contracted group that was enrolled with Aetna the year before. An employer will pay a prior insurer for that information and give it to the new health plan, but those cases are unusual, she said.
Dr. Steinberg said the science of predictive modeling is constantly improving and carries the potential for lifesaving positive health intervention. But formulas are limited, he said, and are much more reliable for underwriting than they are for care management.
Asking patients to rate their health from "excellent" to "poor" is almost as good a predictor of who will be a high-cost patient, Dr. Steinberg said. But the best predictive models can do is identify what Dr. Steinberg calls "actionable opportunities" -- cases where clinical intervention is possible and where it will save money, not just where cost is likely to be high.
David Ferriss, MD, MPH, medical officer for clinical programs development at Cigna, said that by using predictive modeling algorithms developed at the University of Michigan in Ann Arbor, Cigna can pinpoint members who could benefit from disease management or intervention.
In April, Cigna unveiled a new health risk scoring system, linked to a health risk assessment questionnaire. The Trend Management System, as Cigna's system is known, defines 34 dangerous combinations that prompt online coaching and advice for patients to talk to their physicians with their risk assessment report in hand.
Some risk combinations also can prompt enrollment in Cigna's disease management programs if an employer has opted to include those in the member's benefit, Dr. Ferriss said.
Experts differ on whether health plans are using predictive modeling to profile, tier and rank doctors. The answer depends on how one defines predictive modeling.
"Some people may be using it to see how well or how poorly a doctor does in adhering to guidelines," Dr. Brill said. "They can use it for profiling, which is not a good thing."
But rather than looking ahead, most tiering, ranking and physician designations use past data, examining how sick patients are already, Dr. Robbins said. So by a stricter definition, it isn't predictive modeling, he said.
Predictive modeling actually could end up being a physician's best friend, because the more accurate the risk adjustment, the more accurate the tiering, Lutz said.
"There will always be a drive in the marketplace to compare providers," she said, "But I think the provider community practices much more rational medicine than they're given credit for -- in most cases the variation in their cost per patient is explained by these predictive models."
Although most companies contact a patient first, as the use of predictive modeling becomes more common, doctors may find themselves receiving phone calls from health plans or third parties, alerting them to future medical risks among patients, said Swati Abbott, president of MEDai, a predictive modeling company based in Orlando, Fla.