Recommendation Algorithms + Humans: How Health Insurance Exchanges Should Work

The media often like to take certain experiences in health care and demonstrate how they’d play out in other industries. It not only makes for comical and outlandish scenarios, but it provides accurate context as to how outlandish these scenarios actually are in the real world of health care.

A popular video created by whatstherealcost.org shows what it would be like to go food shopping by health care standards. Shoppers wouldn’t know the prices of the items in their basket, and when the cashier would ring them up, they could only cross their fingers and hope that they were within their budget. Then they’d have to keep their fingers crossed until they’d find out for sure upon receiving the bill 30 days later.

There aren’t many aspects of health care transactions and services that consumers wish applied to the way that they handle other types of transactions. On the other hand, health care technology startups constantly adopt approaches they see from tech companies in other industries. ZocDoc has been called the OpenTable for doctors’ appointments and Keas originally set out to imitate some of the features of Mint.

During a presentation Wednesday at Strata 2013 Eric Colson inspired the question: What if health insurance exchanges operated like Netflix? This would mean that, just like Netflix uses algorithms to provide users with the top 10 movie and television show picks for them, health insurance exchanges could use algorithms to present individual consumers with the the best health plans for them. Better yet, the exchanges could go one step further and function like Stitch Fix.

Colson has worked for both companies, first as the vice president of data science and engineering at Netflix, and now he serves as Stitch Fix’s chief analytics officer. Stitch Fix is a new startup with its service still in beta mode. The company’s system involves using computer algorithms in addition to human professionals to make customized apparel selections for a user.

This first step involves plugging lots of data into the computer starting with the product description. Though in this case it’s a simple garment, more than 50 attributes are captured here. These include characteristics like brand, size, color and price in addition to finer details, for instance the price relative to similar items. As much qualitative information as possible is encoded so that it’s machine readable. Next Stitch Fix uses an online assessment to capture nuances about users including height, weight, sizes, brands they prefer and lifestyle characteristics, like whether or not they have children.

“We can find the attributes of the merchandise, or the attributes of the customer, that matter. Not just matter but precisely quantify how much value they provide. We can also come up with measures of compatibility between the merchandise and the customer,” Colson said.

And lastly, before the user gets to see what the algorithms turned out, a human intervenes. In this case it’s a stylist, a professional with lots of related experience in this department. The stylist considers the kinds of information that the customer provided that wasn’t encodable. He or she also rejects the computer’s results when they don’t make sense. “When their overrides are correct, we can incorporate that learning back into the algorithm,” Colson said. “Other times it’s the algorithm that exposes biases and prejudices on the part of the humans.”

Imagine this approach applied to shopping for health plans. The system would involve the same steps. It would first capture all of the information about health plan, which is certainly much more than 50 pieces of data. Next, it would do the same for the customer. Through an online assessment users would log attributes like their age, an affordable monthly premium range, whether they want a high detectible, their employment status, their prescriptions, medical conditions and more. And after the algorithms are run on this data, a professional would narrow the results down to the best health plan picks for that individual.

Last year, a phase of the health reform law required that health insurers create a summary of benefits and coverage (SBC) in order to help consumers understand their options. But come 2014 when millions go to buy insurance, for many with average health care literacy as well as for many who have never gone through the process before ― and even for those who have ― SBCs won’t be enough. As Colson emphasized, getting the Stitch Fix system in working order isn’t just about convenience. It’s about using data correctly to help the consumer find things they’d never have found on their own.