Dr. Richard Merkin, the president and CEO of the Heritage Provider Network in California, faces a dilemma familiar to most health care executives: A relatively small number of his company's patients account for most of its spending. Unlike any other health care executive in America, though, Merkin has put a precise value on solving this problem: $3 million. That's what Heritage is offering to anyone who can build the algorithm that best predicts which patients will be hospitalized and for how many days over the course of a year, based on a given data set. Merkin hopes the award, called the Heritage Health Prize and kicked off today, will help make Americans more healthy and the American health care system more efficient.
Heritage's previous attempt to use "predictive modeling" to reduce costs while aiding patients proved less ambitious: It had its physicians use common sense. "They would know their patients and say, 'Gosh, this patient has a greater likelihood of being high-risk for this condition or hospitalization,' " Merkin explains. Heritage identified patients with complex, chronic medical conditions—such as obesity-related diabetes—as the most "high-risk." It assigned them to physicians with relatively few patients. The individuals received their doctors' cell numbers, and got encouragement to call. They also got frequent checkups.
The changes proved preventive, reducing costs. But Merkin wondered whether there weren't broader and more accurate ways to use data to improve patient care, ways that could bring down even national health care costs. The United States spends an estimated $30 billion a year on unnecessary hospitalizations and $2.5 trillion overall—thus, the stakes are huge. "We did it as physicians," he says. "But we thought that there were other people, the kind of people who put people on the moon, who could do it better with advanced mathematics—predictive modeling, stochastic algorithms, stuff that is not in our tool box." .
Rather than having Heritage contract with a data company for private results, Merkin decided to create a competition open to the public, with a $3 million grand prize and six smaller prizes. Heritage created a data set composed of three years of anonymized, real patient information with things like prescription information and prior claims available. From that data, the predictive modelers who are competing need to predict who will end up hospitalized and for how many days in the year after the data set ends. Because the data set stems from real data, Heritage already knows who gets hospitalized. Thus, it can run a live "leader board," ranking teams as they refine their algorithms.
Answering the question of who is going to get hospitalized with a prize makes sense for Merkin: He sits on the "vision circle" board of the X Prize Foundation, which hosts the grande dame of American "inducement prizes." (It first offered $10 million for advances in commercial spaceflight in 1996. A team backed by Paul Allen won in 2004.) Such prizes have a long pedigree of helping solve seemingly intractable problems. An 18th-century prize spurred an English clockmaker to find a way to measure longitude at sea, and a competition offered by Napoleon led to the development of modern canning. In recent years, such prizes have come into vogue again. A McKinsey report found $2 billion in available prize money, much of it seeded in the past decade. The Obama administration has made prizes a centerpiece of its campaign to gin up innovation.
Such prizes tend to bring alternative forms of expertise to bear on tough questions, breaking down research silos. The world's physicists might not be able to solve a problem, but a chemist working on a related question might be able to. Prizes also tend to be cheaper than traditional R & D spending. The X Prize Foundation found that contestants spend multiples of the prize money offered developing their technologies.
That said, prizes do not always work. Some are not lucrative enough, specific enough, or interesting enough to attract good responses. To ensure that the $30 billion question of unnecessary hospitalization gets answered, Merkin turned to Kaggle, an 11-month-old start-up that has run more than a dozen crowd-sourced, predictive modeling contests. Thus far, Kaggle's competitions have tended to be small-bore—$1,000 for predicting how countries will vote in the Eurovision song contest, $500 for predicting the progression of HIV in certain patients, $10,000 for predicting the outcome of certain chess games. But they have been successful. For instance, the statisticians beat the betting markets in predicting the Eurovision voting.
"We often find the winners come from electrical engineering and physics," explains founder and CEO Anthony Goldbloom. "They're common-sense disciplines, where people are used to problem-solving. Rather than spending time on questions like 'Should I be using this algorithm or this system?' and 'What outcomes are we looking for?' the engineers and physicists just try to answer the question."
The size of the kitty should pull in quality teams—Heritage says it expects scores of competitors. The question surely seems answerable. But there are concerns that remain. For instance, Goldbloom says that Heritage and Kaggle have worked hard to ensure that the people behind the data set remain anonymous. It seems like an outlandish possibility. But de-anonymization has killed prizes before. Netflix, for instance, pulled its second $1 million public competition after computer scientists figured out who some of the users in the data set were. (Goldbloom is working with Canadian researchers who specialize in keeping health information private, as well as with one of the computer scientists who cracked the Netflix prize, to test the data set.)
But on the first day of the two-year competition, everyone was optimistic. "I'm hoping that people will be attracted [to the contest] intellectually and for the betterment of mankind," Merkin says. "The only way families can have affordable health care is if we try to make the system a little more efficient."