ealth insurers are scooping up huge quantities of personal information in a bid to figure out when you’re likely to get sick — and to design interventions to keep you healthy.
Insurance companies have always had access to your medical records, and in some cases your genetic data, too. Now, they’re paying data miners to sift through information on everything from what model car you drive to how many hours you sleep, from which magazines you read to where you shop and what you buy.
The goal: To decipher patterns that will allow them to steer you away from health emergencies. And to save themselves a whole lot of money in the process.
article continues after advertisement
“I think I could better predict someone’s risk of a heart attack based upon their Visa bill than their genome,” said Dr. Harry Greenspun, a director at Deloitte who leads a team that mines data for health insurers and other clients.
Shopping at home-improvement stores, for instance, turns out to be a great predictor of mental health. If you suddenly stop shopping at Lowe’s, your insurance company may suspect that you’re depressed, Greenspun said.
And if you drive a foreign-made car, you’re more likely to lose your eligibility for Medicaid in the coming year, according to Chris Coloian, president of Predilytics, a Massachusetts health-care data analytics company recently acquired by the firm Welltok.
Not all of this information is useful in crafting interventions to keep patients healthy. But with the help of data miners, insurers are finding that some patterns can make for powerful tools.
The intervention can be as simple as determining patients’ ethnicities to make sure they’re receiving information in the right language. Or it can be as aggressive as sending a patient a free digital scale — unprompted — if, say, she has congestive heart failure; unexpected weight gain from pooling fluids can be a sign the condition is worsening.
An insurer might target a chain-smoking motorcycle buff with an action-packed video game designed to help him quit — while appealing to his profile as an adrenaline junkie.
A smoker who subscribes to Better Homes and Gardens, by contrast, might receive an invite to join an online social group to chat with fellow gardeners who are also trying to break the tobacco habit, Greenspun said.
Or the algorithm may tell the insurer it’s not worth trying to intervene at all.
Take a patient who routinely forgets to take his medication. The insurer could keep on him with texts and emails, even send nurses to his home. But if the data mining predicts he’s unlikely to respond to these nudges and is at low risk of a costly hospitalization in the next year, the company may not bother. It may not even remind the patient when it’s time to refill his prescription.
Privacy advocates worry that insurers are using all this highly personal, often sensitive information without informed consent and with little transparency or accountability.
Long before Big Data, of course, health insurers made decisions about which patients to prioritize. But using an algorithm to determine how and when to intervene raises troubling risks, said Kirsten Martin, an assistant professor at George Washington University who studies business ethics and Big Data.
If an insurer writes off groups of patients as “not worth the time of a nurse … certain people are going to become sicker and certain people are going to get well,” Martin said. “You have a problem if all of a sudden you’re only using the nurses on certain groups.”
Health insurers say they don’t deny care to anyone based on algorithms; they just use the data to customize the approach to each patient.
In one popular intervention, for instance, three health insurers in the Northeast designed outreach around neighborhood demographics. They staged outdoor health fairs for customers in walkable neighborhoods. Those more likely to drive everywhere were targeted instead with messages about healthy behaviors on social media, Coloian said.
Leading the Big Data revolution
Entrepreneurs like Colin Hill are leading this Big Data revolution.
Fifteen years ago, as a young Cornell physics graduate student, Hill got swept up in the excitement around the Human Genome Project. He and a fellow physics graduate student started a company that aimed to harness the promise of all that new data.
For nearly a decade, GNS Healthcare worked mostly with pharmaceutical companies, mining genomic and lab test data to help discover drugs and evaluate them. But the firm, based in Cambridge, Mass., has since expanded its focus.
Now, for a given project, GNS culls through millions of medical records, trillions of data points of genetic information, and a huge amount of consumer spending data, using artificial intelligence and machine learning to help companies, hospitals, and insurers figure out how to handle specific patients.
GNS helped the insurance giant Aetna predict which of about 37,000 policyholders were most at risk of developing metabolic syndrome or seeing their condition worsen within a year, using an analysis that took into account demographic variables including ethnicity, cigarette usage, and nightly hours of sleep. Those results helped inform Aetna’s experiment with personalized corporate wellness programs, which included sending members customized vitamin supplements and staging weight-loss contests, among other interventions.
GNS will also rank patients by how much return on investment the insurer can expect if it targets them with particular interventions, such as sending a text message reminding them to refill a prescription or sending a nurse to their home for a checkup. For example, the firm helped a group that manages pharmacy benefits for Regence Blue Cross Blue Shield’s policyholders in the Northwest make decisions about how to target patients who skip their pills.
All patients, of course, should take the medication prescribed to them, “but as a health plan with precious finite resources, where do you focus your energy?” asked Hill, the chief executive of GNS. The algorithm, he said, can tell the insurer not to waste time and money trying to get certain patients to take their pills — but to spend resources on other patients instead.
“Where these algorithms become really powerful is where you can start to match that right intervention to who’s going to be more responsive,” said Iya Khalil, GNS’s co-founder and executive vice president.
This type of work is attracting interest: This month, new investors, including biotech giant Celgene, pumped $10 million into GNS.
A 6-year-old smoker?
Of course, such analyses are only as good as the underlying data and the algorithm used to mine it. And even advocates of the practice have concerns about the accuracy of some of the personal information they use.
“If you’ve got unreliable data, you’re going to make unreliable decisions,” said Dr. Doug Fridsma, president of the American Medical Informatics Association.
This past summer, for instance, Lauren Kiakona’s 6-year-old daughter got a letter from the family’s insurer advising her to quit smoking to better manage her chronic obstructive pulmonary disease. She does not have that disease. Nor, of course, does she smoke.
“HOW DO THEY KNOW? #sarcasm #dontcallCPS,” Kiakona jokingly tweeted, referring to Child Protective Services.
“The fact that [the letter] came to a 6-year-old in the first place just made me feel like they’re not paying attention really to what they’re doing,” Kiakona said.
Kiakona suspects her insurer — the Hawaii Medical Service Association, an independent licensee of the Blue Cross Blue Shield Association — erroneously targeted her daughter because she got a prescription for an asthma medication during a bout of the flu. (Robyn Kuraoka, an HMSA spokeswoman, said the insurer suspects the diagnosis may have been coded incorrectly by the doctor’s office or a billing processor.)
HMSA wasn’t running a sophisticated data analytics operation, but such episodes aren’t unusual as health insurers try to bolster their analytics efforts with new data sources.
For health insurers, Greenspun said, “the real struggle right now is trying to figure out what’s real, what’s not real, where can they get [data], and how much they can trust this.”