It has long been known that a small percentage of patients account for a disproportionately large share of medical costs. What has stumped insurers for just as long is how to find these individuals before they generate large claims and how best to prevent them from actually needing costly care down the road.
But now many insurers say they've found their proverbial crystal ball in predictive modeling, a sophisticated data-crunching technology that not only can calculate future costs for the currently ill but can predict who will likely become sick in the months and years to come. And health plans are fast embracing the technology for everything from disease management to budgeting, rate-setting and provider profiling.
"Predictive modeling is helping managed care deliver on its promise of providing better care at lower costs," says Randy Hutchinson, director of business development for the Blue Cross and Blue Shield Association.
Indeed, most large insurers, including Aetna, Cigna Corp. and Humana, are either developing the technology in-house or licensing it from outside vendors. And over the past nine months, the Blues association has inked deals with three predictive-modeling firms to make the technology available to its 41 affiliates at preferred rates.
Predictive models analyze members' personal information, including their medical and pharmacy claims, age, gender and ZIP codes, and even lab results. The programs can then generate forecasts in any number of forms, from an estimate of total medical spending for a group to a list of specific members most likely to be hospitalized.
Insurers contend their use of this powerful technology will benefit the entire healthcare industry by enabling them to improve preventive care, cut medical costs and price their contracts with employers and providers more accurately. But critics see predictive modeling as a threat to privacy and fear the technology is ripe for abuse by insurers eager to weed out members who are about to become expensive.
"The market for predictive modeling has heated up," says Marilyn Schlein Kramer, president and chief executive officer of Boston-based DxCG, which has sold predictive-modeling tools for seven years. "For the first five years, the world didn't talk much about predictive modeling. Now it's on everyone's lips."
At its most basic level, predictive modeling can help insurers maximize their resources by ranking their membership according to health risk. If a plan has 10,000 diabetic members but can afford to manage only 500 at one time, the technology can pinpoint the 500 who are most likely to be the costliest patients in the coming year. These members are then shepherded into disease-management programs, where they receive information brochures, nurse monitoring, regular checkups and other services designed to head off a major medical disaster.
Christopher Scheib, manager of health economics modeling at Highmark Blue Cross and Blue Shield, says predictive technology has helped the insurer allocate call-center resources more efficiently. "We don't want to be spending time and money calling everyone, including members who are healthy," Scheib says. "So now, only those who are imminently ill are most likely to get a nurse outbound call from us, while people with `softer risk' will likely receive less-resource-intensive interventions such as mailings."
Insurance executives say it's difficult to put a dollar figure on the savings generated directly from predictive models, which can cost anywhere from $20,000 to several million dollars. But Blue Cross and Blue Shield of Tennessee says it has cut its need for case-management nurses by 30% and has tripled its savings from case management since implementing the technology about 18 months ago.
The Blues plan says its predictive model has proved 85% accurate in forecasting medical costs across a large group and 60% accurate in predicting individual members' costs. That's far higher than the 2% accuracy rate it saw when age and gender were the only data available.
Some vendors and insurers contend there's not much point in predicting members' health risk more than one year out because, given industry trends, there's a good chance those members will switch plans before long anyway. Others, however, expect the biggest financial savings to come from forecasting--and ideally treating--members who may be headed for a serious illness several years from now. "Presumably, the earlier you catch a developing condition, the better the patient will be over the long haul," says Steven Coulter, chief medical officer for the Tennessee Blues, which now flags potentially high-risk members as far as two years out.
And the power of predictive modeling will continue to grow as new types of data become available for analysis.
The technology, for instance, took a big leap forward a few years ago when companies began to incorporate the results of laboratory tests into their programs--creating a far more detailed snapshot of members' health than medical claims alone, says Michael Cousins, executive director of health informatics for vendor Health Management Corp., a unit of insurer Anthem. "It's the difference between knowing that a member has had a HDL cholesterol test and knowing how that member actually scored on the test," he says.
Health Management has designed a program that analyzes specific patterns in insurers' claims data to zero in on the high-risk members who can be most readily helped. For instance, it may alert insurers to a diabetic who hasn't had his blood checked by a specialist in the past year. This member offers a greater opportunity for cost savings than, say, a diabetic who is expected to be high-cost next year but is already receiving optimum care.
One day, experts says, predictive models may even incorporate the results of genetic testing, which can detect a patient's predisposition to specific hereditary diseases, such as breast cancer. But for now, that area remains a legal and ethical minefield, experts say.
Meanwhile, privacy issues abound.
Ronald Green, director of the Ethics Institute at Dartmouth College, says close monitoring of members' data could exert a dangerous "chilling effect" on their willingness to seek care. "Just as automobile owners sometimes do not report minor accidents or damage for fear that it will increase their insurance rates, so patients may shy away from medical treatment in the fear that every encounter with the doctor may tarnish a good `medical history,' " Green says. "This problem will be exacerbated if any link is made to coverage or pricing."
But Lonny Reisman, a cardiologist and CEO of vendor ActiveHealth Management, says most patients are glad to share their medical information once they see it leads to better care. "We get the `Big Brother' thing sometimes," he says. "But once they learn how the program works ... they basically say, `Sign me up!' "
ActiveHealth's program mines claims data and lab results, looking for deviations from nationally recognized practices--and alerts members' physicians to potential problems or misdiagnoses. The program, for example, flagged a man who hadn't had a biopsy although his test for prostate cancer indicated the need for one. It also caught a potentially life-threatening drug mix-up in which a man was prescribed the relatively harmless drug quinine for restless legs but was mistakenly given the potentially dangerous heart drug quinidine by his pharmacist.
ActiveHealth typically faxes or mails alerts to physicians on less pressing matters but will phone physicians about immediate dangers, such as potentially fatal drug interactions. According to Aetna, which uses the program, about 70% of the 100,000 alerts it generated last year were acted on by physicians.
"In terms of achieving actual changes in practice, the technology is quite astounding," says James Cowan, medical director at Aetna Integrated Informatics, a unit of Aetna that stores and analyzes members' clinical data. He estimates that the program has saved Aetna an average of $2 to $4 per member per month. Aetna had 1.25 million members enrolled in the program as of March 31.
Reisman argues that such global monitoring is already commonplace in the banking and airline industries and can serve as a powerful safeguard in healthcare as well. "You might have a great pilot--just like you might have a great doctor--but you wouldn't get in the plane if there wasn't a much broader navigational, air-traffic-control system out there to tell him, `There's a mountain in front of you. Pull up,' " he says. "The same should be true in healthcare. Your doctor is still in control. ... What we do is alert them to the mountain."
As with other vendors and insurers, compliance with federal privacy laws is a top priority for ActiveHealth, Reisman says. Employers are never shown members' personal data; all they get is an aggregate report.
Empire Blue Cross and Blue Shield, which has used ActiveHealth's technology for the past five years, tells members about the program upfront and gives them the chance to opt out. "But only a small percentage do, about 1% or 2%," says Alan Sokolow, chief medical officer for the New York-based insurer.
As for doctors, some regard medical advice doled out by third parties as intrusive. But most appreciate being kept up on the latest medical research and are receptive to recommendations designed to improve outcomes, Reisman says. Doctors have also praised the program for helping to protect them against liability suits, he says.
Adding a new dimension to the debate, predictive modeling is fast expanding into the actuarial realm, with an increasing number of insurers using it to set premiums and establish payment rates for providers. The benefits are clear. While traditional actuarial tools estimate future costs based on members' age, gender, occupation and claims history at the group level, predictive models can forecast medical spending by each individual member, giving the insurer a far more detailed cost estimate for the group as a whole. This helps insurers justify their rates to large-group clients and decide if and how to raise or lower their prices at contract renewal time. It also allows health plans to set more appropriate capitation rates, paying more to providers that treat sicker members and less to those treating largely healthy populations.
"It allows us to do precision pricing, which means getting the right rate to the right group," says Coulter of the Tennessee Blues, which has been using predictive modeling for underwriting since 2001.
Coulter uses the example of a small business that racks up huge medical costs in 2004 because two employees suffer heart attacks, one of whom has bypass surgery and the other of whom dies. An insurer using traditional underwriting would likely raise this employer's rates in 2005 because of its recent claims history. But an insurer using predictive models might very well forecast lower costs for the employer in 2005, given that the bypass recipient would likely generate fewer claims than before while the deceased patient wouldn't generate any claims at all.
On the flip side, Coulter adds, an employer that had few medical claims in 2004 might see its rate jump significantly in 2005 because the insurer's predictive model shows that a number of its employees are about to get sick. "It's about transparency--helping our clients understand the reasons behind our decisions," he says.
Medical ethicists, however, take a more cynical view, contending that predictive modeling could be misused by health plans eager to shield themselves from high-risk enrollees. While insurers in most states are prohibited from denying coverage based on customers' health status--let alone their potential health risk--they could use predictive models to price certain high-risk individuals right out of the market, ethicists suggest.
That's precisely what American Medical Security Group was accused of doing when, in 1998, it canceled all its individual policies in order to separate its healthy customers from the sicker ones who posed a bigger cost risk. It then reissued coverage, but jacked up its rates for the sicker customers while charging more modest rates to the healthy. One policyholder saw her premiums soar to $4,860 per month, or more than $58,000 annually, after the company discovered she had diabetes and wore a pacemaker.
According to class-action lawsuits filed in Alabama, Florida and Georgia in 2000 and 2001, American Medical violated state insurance laws by creating a "death spiral" in which rates for the sicker policyholders continued to climb, forcing them to either leave the plan or pay exorbitant premiums. The Howard, Wis.-based insurer argued that, as an out-of-state association health plan, it was exempt from most of those states' laws.
American Medical lost the Florida suit in April 2002 and settled the other two for $9 million in March 2004, without admitting wrong-doing. In April 2003, though, a Florida appeals court blocked an attempt by the state insurance department to suspend the company's license.
"To the extent that they seek reimbursement, patients realize that their condition and care are known by the insurance company. However, (predictive modeling) converts every disclosure into decision-making that can affect both their coverage and their care," says Green of the Ethics Institute. "This strikes me as highly intrusive."
Insurers bristle at such accusations, arguing that state and federal laws exist to prevent discrimination by health plans. What's more, picking and choosing between members just doesn't make good business sense, Coulter says.
"We can't go around saying, `We don't want you anymore, or you or you,' because that defeats the purpose of insurance, which involves pooling risk," Coulter says. "If we did that, we wouldn't have clients for long. So not only would it be unethical, it would bad business."
Randy Hutchinson of the Blues association adds that many insurers have been pushed to adopt predictive modeling by their clients, which have applied the technology successfully to their own operations. Retailers, for instance, use predictive modeling to determine exactly where specific merchandise should be placed in their stores to best attract customers and boost sales, he says. Banks have used it for years to determine the risk of a borrower defaulting on a loan, while meteorologists use it to forecast possible storms.
Good for the gander
To be sure, predictive modeling can also be used against insurers--by employers eager to combat double-digit premium increases or providers that want to ensure they're getting reimbursed enough for their services.
Take the Massachusetts Group Insurance Commission, which uses DxCG's software to adjust payment rates to the nine health plans with which it contracts.
The commission, which covers 260,000 state workers and their families, first estimates the relative cost risk of enrollees in each plan and then uses its findings to "reality check" insurers' bids at renewal time, says Dolores Mitchell, the commission's executive director.
"If a health plan comes to us and says, `You have to pay us more because your members are older and sicker,' we can say, `Hogwash. No they're not, and here's proof,' " Mitchell says. "It makes for a much more sophisticated discussion on what our rates should be or how much they really need to receive to cover their projected costs."
Meanwhile, Boston-based CareGroup Healthcare System, a network of five hospitals and 2,300 physicians, uses DxCG's predictive-modeling tools to evaluate the adequacy of insurers' capitated payment rates. For example, if its program shows rising utilization among a health plan's members, "We can ask for larger reimbursements from that payer," says Kathryn Burke, vice president of contracting for CareGroup's Provider Service Network.
But she admits that it's sort of a "one-sided argument" that hasn't always worked. That's because insurers have data on their entire membership, while CareGroup has data only on the members it treats. "We don't know what we're dealing with when they negotiate with other provider groups," Burke says. "We can show changes in our population over time, but we can't show how sick our patients are relative to (the insurer's) overall population."
But that could change, because a number of the insurers that CareGroup contracts with, including Tufts Health Plan and Harvard Pilgrim Healthcare, recently adopted the same predictive-modeling program, allowing for the possibility of greater data-sharing. "It could put us more on the same page," Burke says.
In the meantime, CareGroup is using predictive modeling to distribute payments more fairly among its network providers, paying more to those that treat sicker populations. It also uses the technology in physician profiling: Before comparing doctors based on their financial performance, it adjusts for the relative cost risk of their patients. That way a doctor isn't penalized for, say, treating high-cost AIDS patients while others aren't.
Burke says the risk-adjusted profiles have met with greater physician acceptance. Before CareGroup started using predictive modeling, doctors could explain away the results of utilization reports by claiming their patients were sicker. "They'd just shrug it off" as invalid, she says. "Now we can give them a full list of their patients who are high risk and show them exactly how sick their (patient population) is compared with their peers."
Similarly, insurers and employers are incorporating predictive modeling into their pay-for-performance programs, to reward physicians on a risk-adjusted basis and to set performance targets for future incentives. And further down the chain, individual medical groups are starting to use the technology to ensure they are getting enough money from insurers to treat their patients.
"People talk about predictive modeling because it sounds neat and `Star Trek'-like. But the bottom line is that it's a very powerful tool with any number of uses," Coulter says. "The possibilities are almost endless."
What do you think?
Write us with your comments. Via e-mail, it's [email protected]; by fax, 312-280-3183.