Predictive modeling being tested in data-driven effort to strike out hospital readmissions
Hospitals waiting on Congress to squeeze spending may sympathize with the Oakland Athletics' reversal of fortune more than a decade ago, when the ball club's owners refused to continue profligate spending.
Soon thereafter, as readers of the bestseller Moneyball
know, Oakland A's General Manager Billy Beane relied on fancy math and economic theory to put together a competitive baseball team on a shoestring major league budget. Beane's conviction—that statistical analysis could trump baseball tradition that blinded the sport to valuable players and left October to teams able to afford errors—became the subject of Michael Lewis' 2003 book and a newly released movie.
“We take 50 guys and we celebrate if two of them make it,” Beane fumes about the major league draft in Lewis' book. “In what other business is 2-for-50 a success?”
Now hospital executives, flooded with data from information technology investments and under pressure to curb waste and spending, face the same question that Beane and others before him, including Bill James, an influential author and baseball statistician, sought to answer: What measures matter the most?
In healthcare, the question is critical when it comes to who ends up in the hospital.
Hospitals, which house squadrons of nurses, pricey technology, pharmacies and laboratories, are costly places to care for patients, with hospital care accounting for healthcare's single largest expense. Policymakers see significant potential to curb spending by keeping more patients out of the hospital. In 2005, Medicare paid hospitals $7,200, on average, and $12 billion in total for repeat hospital visits that could have been avoided, according to one estimate by the Medicare Payment Advisory Commission.
Hospitals also can be unsafe. Patients risk infections or other avoidable complications during a hospital stay that, paradoxically, can land them back in the hospital shortly after leaving.
And hospitals now have added incentives to better identify such patients. Medicare, the single largest customer for many hospitals, will penalize those with too many repeat patients starting in 2013. The penalty starts at up to 1% of hospital Medicare revenue, then increases over the next several years to as much as 3%.
Indeed, within hospitals and health plans, some have started to use the same fancy math as the Oakland A's and other predictive models to identify patients at risk for unnecessary hospital stays.
“Nothing good happens when you go to the hospital,” says John Billings, director of the Center for Health and Public Service Research at New York University's Robert F. Wagner Graduate School of Public Service. Billings and colleagues developed an algorithm being tested by Medicaid in New York. “Going to the hospital means something went wrong outside of the hospital,” Billings says.
Similarities between baseball's adherence to sometimes ineffective and costly convention and the inefficiency of U.S. healthcare have not escaped notice.
“America's healthcare system behaves like a hidebound, tradition-based ball club that chases after aging sluggers and plays by the old rules: We pay too much and get too little in return,” wrote Beane and health policymakers Newt Gingrich and Sen. John Kerry (D-Mass.) in a 2008 New York Times
In New York state, where state spending per Medicaid enrollee is among the highest anywhere, state officials have bet that the algorithm developed by Billings and others five years ago may answer which variables can predict which patients will be hospitalized soon after leaving.
Roughly 15% of hospitalized New Yorkers returned to the hospital within a month in 2008 at a cost of $3.7 billion, according to an analysis by Mathematica Policy Research. Repeat hospitals visits from complications and infections cost $1.3 billion that year. Some 2% of Medicaid patients landed back in the hospital 30 days after leaving because of complications or infections (See chart).
Researchers who developed the algorithm teased out 21 common variables from more than 60 possibilities to find the most informative measures. Diabetes, chronic obstructive pulmonary disease and alcohol-related diagnoses were included, as was older age (at least 65) and recent emergency admissions, among others.
Billings and other researchers tried to refine the algorithm by testing additional variables from electronic health records—such as lipid concentrations, blood pressure, glycated hemoglobin levels, accident and emergency data—but found hospitals' differing record systems made access to uniform data difficult.
As healthcare predictive modeling evolves, some research suggests not all high-risk patients will benefit.
One 2008 survey published in the Milbank Quarterly
of 30 organizations that build, use or evaluate predictive healthcare models—including the Veterans Affairs Department, Kaiser Permanente, Humana and UnitedHealth Group—found some care-management companies exclude the highest-risk, and often highly vulnerable, patients as unlikely to respond to support.
But others sought to tailor preventive care to the specific circumstance of patients, developing different approaches based on marketing research for patients with similar diagnoses, the survey found.
For hospital admissions, a review of recent hospital stays won't identify who will return repeatedly the following year, Billings says. “A lot of people who were high-cost last year will be high-cost next year, but not everybody,” he says. And some high-cost patients require high-cost care; not all repeat visits are unnecessary, he says.
Hospitals that use the algorithm can expect to identify two out of three patients who will land back in the hospital but didn't have to, Billings says. He describes the algorithm's predictive power as “strong” and one that does not falsely identify too many patients as chronic users.
The ability to weed out patients who will stop returning to the hospital without intervention from caregivers is important to prevent hospitals from mistakenly directing limited resources where they're not needed, he says.
“An effective case-finding tool is one that identifies as many patients as possible who will have future high costs or hospital resource use without intervention but is not so broad that it includes large numbers of patients who will not incur such costs,” wrote Billings and colleagues in the British Medical Journal.
The formula, first developed for the National Health Service in England, relied on statistical analysis embraced by Beane and Sandy Alderson, his predecessor in the Oakland A's front office.
Alderson put his faith in statistical analysis that challenged baseball convention and showed that some strategies (such as stealing base) could be pointless, Lewis wrote. “I couldn't do a regression analysis,” Alderson says. “But I knew what one was. And the results of them made sense to me.”
(Beane hired Harvard graduate Paul DePodesta, Lewis wrote, a numbers guy obsessed with walks. DePodesta concluded “that foot speed, fielding ability, even raw power tended to be dramatically overpriced,” Lewis wrote, and that “the ability to control the strike zone was the greatest indicator of future success. That the number of walks a hitter drew was the best indicator of whether he understood how to control the strike zone.”)
Across New York, six locations are testing a modified version of Billings' algorithm with state funding to identify and manage care for patients with the most costly and avoidable medical bills.
One small prior pilot at Bellevue Hospital Center in New York paired patients identified by algorithm with case managers and a housing agency and saw hospital visits drop by one-third within one year. New York Medicaid saved $5,080 per patient after deducting expenses for the small-scale pilot—and after outpatient Medicaid costs increased an average of $474 per patient, according to Dr. Maria Raven, an emergency room physician who oversaw the pilot.
Raven, now an assistant professor at the University of California at San Francisco, continues as director of the algorithm pilot for the New York City Health and Hospitals Corp., which includes Bellevue. She says HHC will continue to use the algorithm as New York adopts medical homes under the Patient Protection and Affordable Care Act.
The algorithm will help identify which medical home patients could benefit from intensive case management, she says.
At Henry Ford Hospital in Detroit, an analysis of frequent emergency room patients underscores the demand placed on providers by those who repeatedly return. Repeat patients more than accounted for the 9% increase in overall emergency room visits between 1999 and 2009 at the hospital, according to research by Dr. Stephanie Stokes-Buzzelli and colleagues.
ER visits from patients who turned up at least five times in any given year increased 83% during the same period, according to the analysis. Results also found many patients who made multiple visits each year to the hospital's ER.
Roughly 350 patients were logged among chronic ER visitors in more than one year. The research measured visits during five-year intervals for repeat patients, with some who arrived at the ER at least 20 times in a year. Forty-one patients were regular ER visitors in 1999, 2004 and 2009.
Stokes-Buzzelli says the findings raise questions about whom providers should target with prevention and care-management measures.
But the hospital's more general approach has shown some success.
Henry Ford Hospital seven years ago began a voluntary effort to better manage care for ER patients well known to doctors after frequent visits, including one patient who returned 127 times in one year. The hospital relied on physicians to find high-use patients.
Stokes-Buzzelli helps to draft care-management plans for regular emergency room visitors. Plans are flagged in EHRs, but serve only as a guide for other doctors or nurses. The effort has reduced the time patients spend in the ER, the number of visits and costs, she says.
SelectHealth, the health plan owned by Intermountain Healthcare, started to use predictive models earlier this year to identify patients at highest risk of being high-cost. Intermountain, a Salt Lake City-based health system, includes 23 hospitals in two states, and its health plan covers 500,000 people, according to SelectHealth.
SelectHealth refers potentially high-cost patients to Intermountain care managers or one of 35 nurses the health plan employs to coordinate patient care.
Dr. Mark Briesacher, senior administrative medical director for Intermountain Medical Group, says nurses at 10 Intermountain clinics with newly created medical homes contact patients identified by SelectHealth. Six months into the experiment it was too soon to measure any possible differences, Briesacher says.
Meanwhile, nurses contact every patient who returns home from the hospital in a separate pilot at five medical home clinics. Nurses call with information about medication and clinic appointments. Briesacher says not all patients may need the assistance, but it's worthwhile for nurses to contact every patient.
And nurses use medical records and judgment—not predictive modeling—to identify who might best benefit from additional assistance, he says. Nonetheless, Intermountain has developed its own scoring system for high-risk patients based on registries of chronically ill patients (with the exception of cardiology), laboratory results and prescription claims data. Primary-care doctors tested the scoring last year, he says, adding that the medical group has focused on data development and has not conducted any research on the scoring method.
Dr. Stephen Barlow, medical director for SelectHealth, says it's too soon to project potential savings from the modeling effort, which added statistical analysis of medical claims to existing identification of high-risk patients based on information gleaned from EHRs.
Barlow says SelectHealth contracts with Thomson Reuters and OptumInsight for its attempt at predictive modeling, which focuses on the top 2% of potentially high-cost patients. Typically, the top 1% of patients account for 30% of spending and the top 5% accounts for half of costs, he says.
But Barlow acknowledges that predictive models have limits for hospital and clinic officials in search of potential opportunities to reduce waste, errors and costs.
Patients identified as high-cost include some, roughly 60%, who drive up spending for one year but not subsequent years, he says. Nor are commercial predictive models that Barlow surveyed highly accurate; SelectHealth can expect to explain 30% of its costs with the model.
“It just means you have to be careful how you use them,” he says of the models. “We're still doing better than if we were just throwing darts.”