Predictive modeling sharp lens near future

It has long been known that a small percentage of health plan members account for a disproportionately large percentage of medical costs. What has bedeviled health plans for an equally long time is how to identify these members before they generate large claims and how best to keep them from actually becoming so costly.

The old way is to establish cost “triggers.” If a member exceeds a specified dollar amount in claims within a certain time period, he or she automatically becomes a candidate for medical management. But by the time they’re identified, these members are often well on the way to becoming low cost cases again, thanks to some sort of medical intervention or perhaps even the characteristics of the disease itself. The barn doors are closed, but the proverbial horse is long gone.

The latest

The new way is predictive modeling, and it is being implemented at health plans left and right.

“The goal of a clinical strategy is to find the right intervention for the right person at the right time,” says Carol McCall, vice president of Humana’s Center for Health Metrics, “and ultimately, you want to be able to understand enough about people to be able to literally custom-tailor the exact intervention so that you can meet them in a way that they will be the most receptive to it.” McCall is a fellow of the Society of Actuaries and a member of the Academy of Actuaries.

It turns out that remarkably accurate predictions can be made by using sophisticated software to analyze data about plan members. For example, someone who had three emergency department visits in six months and takes five prescription drugs may not exceed any cost triggers, but could be headed for big trouble.

Data needed

Data about the plan member in this highly simplified example come from medical and pharmacy claims. In the real world, these data may be supplemented by demographic data, previous claim history, perhaps by diagnosis codes, procedure codes, ZIP codes (because the incidence of certain conditions varies by geographic location), or even lab results. More data is invariably good in the predictive modeling game, and as ever more data become available, it’s a safe bet that those morsels will become additional fodder for predictive modeling.

Health plans have these data.

“Who better can do this?” asks Val Dean, vice president for care and network management at the TriZetto Group, an IT vendor in Newport Beach, Calif. Who else, he asks, has the provider contracts, the utilization and outcomes data and other databases that are needed?

Hundreds, thousands, millions

Right now, predictive modeling programs can analyze data on thousands, hundreds of thousands, or millions of plan members, connect the dots to identify potentially high cost cases, and provide information about why a particular member was selected — — real actionable information — that can help teams of nurse health coaches and disease management workers hit the ground running to head off a medical disaster.

Risk assessment

In the not-very-distant future, even more sophisticated predictive modeling products will crunch even more member data to reveal not only the likelihood of future medical problems, but also to suggest the best way a health coach or member of the DM or case management team should approach a plan member, based on that member’s behavioral history of interaction with the plan’s web site, interactive voice recognition calls, or nurse triage program.

Still, if you talk with 10 predictive modeling experts, you’re likely to get 10 different definitions of predictive modeling.

It may be helpful to think about predictive modeling as part of what Johns Hopkins Bloomberg School of Public Health Professor and health services researcher Jonathan Weiner, PhD, calls the “risk assessment and adjustment continuum.” As suggested above, the management applications of predictive modeling include case management, disease management, practice resource management, needs assessment, quality improvement, and payment and finance.

“One of the things on my wish list is a shared framework and nomenclature,” says Weiner, who is codeveloper of the Johns Hopkins Adjusted Clinical Groups (ACGs) model. He defines predictive modeling as “a process that applies available data to identify persons who have high medical need and are ‘at risk’ for above-average future medical service utilization.”

Lots of vendors out there are promoting lots of different predictive modeling products based on various algorithms, artificial intelligence, or neural networks, each designed to use certain kinds of data exclusively or in combination, to predict concurrently (using data from a prior period to project medical claim costs for that same period) or prospectively (using data from a prior period to project costs for a future period), and each targeted at a variety of business segments such as Medicaid, Medicare, and commercial.

Transitional population

“It’s pretty easy to find somebody with a chronic illness like CHF who’s very expensive,” says Alan Hinkle, vice president for health care quality, policy, and innovation at Blue Cross Blue Shield of Massachusetts.

“I don’t really need a computer system to help me figure that one out. On the other end, somebody who’s really healthy, like the Olympic athlete I have in my plan, I can pretty well predict will remain healthy. But if you’re going to be proactive, you’re identifying people we call our transitional population.”

That transitional population is getting very intensive scrutiny from health plans these days.

Earlier enrollment

At Blue Cross Blue Shield of Massachusetts, appropriate intervention could include earlier enrollment in a disease management program or in its new nurse-staffed Blue Health Coach program, or a primary care office visit. According to Hinkle, one predictive modeling tool he is using can identify up to 40 percent of the actual patients who are relatively healthy today, but will be high cost next year.

Relatively new “neural net” PM tools look for patterns in large batches of data, such as an emergency department visit followed by a cardiac stress test.

By combining data from a variety of sources, the PM tool is able to identify only those members who are most likely to have heart problems. After all, some stress tests will have ruled out a diagnosis of coronary artery disease.

Pattern recognition has been used in other industries for decades. Chances are that consumer catalogs you’ve never received before come to you now because a PM tool searched a credit card database for another kind of pattern.

For example, if you live in New Hampshire, drive a truck, and shop at L.L. Bean, that pattern has a certain predictive value and may make you a good candidate to receive a fly fishing catalog. The same principle is at work when you get a call from your credit card company to make sure it was you using your credit card.

PM can also optimize disease management resources. If a plan has 35,000 diabetics in its population, but can manage only 500, by programming in the appropriate parameters, the PM tool can identify the 500 that are likely to be the most costly next year.

Three different products

Hinkle will say only that he’s been using three different PM products since June 2002. He doesn’t expect to have determined his return on investment until sometime in 2004, but says patient satisfaction and physician satisfaction have been gratifying. Members appreciate being offered preventive services they would not otherwise have considered, and physicians are pleased to have made a positive contribution to their patients’ health.

“The hard work of managing the patient begins after the predictive modeling is done,” says Hinkle, “but it’s clearly a superior tool to better identify members for appropriate health management programs.”

Michael Cousins, PhD, is in the business of developing PM tools for disease management and case management. He’s manager of health informatics at Health Management Corp., a disease management company headquartered in Richmond, Va. HMC is a wholly owned subsidiary of Anthem, but also serves other health plans and employers. Like most DM companies, HMC develops its own PM products. (“DM companies consider predictive modeling to be a core competency,” says Al Lewis, executive director of the Disease Management Purchasing Consortium.)

Two advances

Cousins expects two major advances to revolutionize PM in disease management within the next five years. First, the actual results from an increasing variety of laboratory tests will be incorporated into predictive models. Currently, most claims indicate only that a lipid panel was performed on a member, but do not include the actual values.

Because recently published evidence suggests that C-reactive protein is strongly predictive of future cardiac events, actual values from that test will probably be among the first to be included in data for PM. The ultimate frontier in the realm of testing, says Cousins, is genetic information, but that involves a host of controversial issues that may not be resolved for a long time.

Second, more detailed information about the clinical progression of conditions such as diabetes, heart disease, CHF, asthma, and COPD will improve the ability of PM to recognize patterns and make better predictions about the progress of the disease.

For example, a lipid panel followed nine months later by a stress test may be a pattern that is highly predictive of an MI.

An even better predictor may be a lipid panel followed by a stress test performed at a teaching hospital by a specialist instead of in a primary care physician’s office. In reality, of course, these events are combined with dozens of other variables. Cousins says HMC is developing two types of PM models to analyze a series of such events.

The technique is called a time series analysis. When asked whether the actual specialist, or teaching hospital, or PCP in the preceding example would ever be identified, Cousins responded that the question addresses a politically sensitive area.

“I was talking about the type of specialist, not a particular physician,” he says. “However, I can see a day where we incorporate information about particular providers. Health plans profile physicians, and that information can be employed to make a better prediction about the future health of the member.”

Improving value

“The gold standard would be to get to the individual practitioner level, but we’re years away from that,” says Dave Kelleher, president of Healthcare Options, a consultancy in Indianapolis, Ind. “We’ve at least got it down to a care system level.”

Kelleher has been working since 2001 with the Employer Forum, a group of large Indianapolis area employers and health plans, to enable individual employees to make choices based on value. Measuring health plan performance doesn’t really drill down far enough, argues Kelleher, because employees get their health care from systems like medical groups, IPAs, and PHOs that contract with HMOs. By aggregating individual risk scores in their patient panels, PM makes it possible to compare the performance of care systems.

“First, we asked the health plans to provide quality information at the care system level,” Kelleher explains. “Second, we’ve asked them to give us an acuity-adjusted premium for each care system. Third, we’ve asked employers to straighten out their contribution policies.”

Essentially, that’s what the Pacific Business Group on Health (PBGH), an employer purchasing group in San Francisco, has been doing since 1999, but still at the health plan level. PBGH uses PM to determine which health plans offer better value or manage populations with greater medical needs more efficiently, and adjusts premiums accordingly. The health care plans, in turn, will use the same approach in evaluating their providers.

“There’s a huge opportunity to work more proactively in identifying and recruiting members into care management programs,” says Emma Hoo, PBGH’s director of value-based purchasing. “It’s also an opportunity to use these tools to identify people and engage them in self-care so that they’re more active in managing their own conditions, because the returns are measured not only in the reduced medical costs, but greater productivity and reduced absenteeism.”

In the trenches

You don’t have to sell Jack Mahoney, MD, on the benefits of proactive intervention or predictive modeling. He’s medical director at Pitney Bowes and, using at least two PM products, he has been able to meet his self-insured population’s preventive health care needs much more effectively over the last year.

Again, the question was, “Which otherwise healthy employees would incur substantially higher future costs in medical, worker’s compensation, and disability claims — and why?” This time, however, figuring out “why” was more important than “whom.”

“Other entities use predictive modeling to identify individuals that they can reach out to proactively,” Mahoney explains. “We were looking for something that we can use on a population basis, because going directly to individuals would mean major problems with HIPAA confidentiality and our own ethics guidelines.”

One PM product identified diabetes, asthma, depression, and cardiovascular disease as the four most costly conditions in Pitney Bowes’ self-insured population. But further analysis revealed that these conditions themselves were less powerful predictors of future cost than compliance with treatment. With this breakthrough information, Mahoney was able to focus on maximizing compliance by increasing employee awareness of the four disease states and increasing access to appropriate treatment.

First, he moved the drugs most commonly prescribed for the four diseases into the lowest tier of the company’s coinsurance pharmacy benefit plan.

Second, he switched his nurse call line from one that was designed to help with acute episodes to one that provides coaching for employees with chronic conditions. Third, he focused the company’s nationwide Health Care University wellness program on strategies to maximize compliance. In conjunction with a web site, the program emphasized education, self-care, exercise, and nutrition.

“I think what we accomplished with predictive modeling was to refocus our efforts on where we could contribute the most as a company to improving the health of the individual, and where we could synergize with the care delivery system so that we were augmenting the efforts of their primary treaters,” says Mahoney.

Analysis of claims later in 2003 will show whether these initiatives have lowered costs. Mahoney says it’s too soon for definitive results, but indicators such as prescription refills, appropriate check-ups, and participation in the nurse-coaching program suggest positive changes.

Seeing the future

At Humana, PM is being used to identify healthy members whose health status is likely to decline, to model individual member interaction with Humana programs and services (e.g., how likely is this member to buy the drug that just became available over the counter if we offer her discount coupons?), and to prepare for consumer-directed health care. The goal, argues McCall, is to “unleash the power of consumer engagement” by offering consumers choices that are meaningful, transparent in terms of cost and consequence, and based on independence and autonomy.

“It’s not about telling you what to do as a consumer,” she says. “It’s basically saying that there is a range of options out there and it’s important for you to understand what they are, what the consequences are emotionally, physically, financially, for you and for your family.”

Whether it’s choosing health benefits or deciding to undergo surgery, much of the information consumers will want and need is exactly the information health plans already have. Along with sophisticated PM tools to project this data into the future for decision-making, health plans may be ideally positioned to be the information brokers in the new consumer-driven health care marketplace.