By Allison Proffitt
September 4, 2019 | During his last year of medical school, in an Emergency Department rotation at Massachusetts General Hospital, Andrew Le met a diabetes patient who came in with a foot ulcer and a ream of printouts from the internet explaining why he had been right to wait several days before seeking help and why amputation was unnecessary.
That was when Le realized it: the front line of healthcare is not the emergency room; the front line of healthcare is the internet. Regardless of what the printouts said, the patient lost his leg. And Le gained a new perspective on what it means to meet patients at the starting point of their care.
Le took a sabbatical from medical school (he finished later) and started Buoy Health. That was six years ago, and now Buoy’s AI-powered health assistant boasts seven million users across the country, seeing a new user every 13 seconds. Earlier this month the company closed a $15 million Series B round with investments from Bill Hambrecht, Humana, F-Prime (Fidelity Investments venture capital firm), and Optum Ventures.
Fixing the First Step
Buoy is “obsessed” with the idea of fixing the first step in healthcare, Le tells Diagnostics World News. “We define that first step as the moment you feel that you are sick or injured. You have to answer these pivotal two questions: ‘What do I have?’ and ‘How do I fix it?’”
The company built the artificial intelligence over the course of four years from scratch. “We read thousands of clinical papers by hand to teach the program the statistics that underlie medicine,” Le explains. “We went back to the primary literature and essentially ‘mapped’ medicine.” The AI asks foundational questions—gender, age, most bothersome symptom—and then asks the most appropriate next question based on the things that are statistically impactful.
This isn’t a nurse call line, Le says. “Nurses are doing triage largely driven by decision trees written in a book that the nurse is flipping through while on the phone,” he says. “The problem with that is that people are so different from one another when they come in with different diagnoses. It becomes very difficult to understand which decision tree to use when people come in with more than one symptom or they have some complicating factor.”
In Buoy’s AI, “thousands of questions get re-ranked, thousands of diagnoses get re-ranked, and the next best question gets asked,” Le says—in real time. As Buoy sees new patients, the program is constantly readjusting and getting better. It’s a process that mirrors the way doctors think.
But the problem is we don’t start with doctors. Nearly three quarters of us will turn to the internet first to find answers to health problems.
So Buoy meets us where we are, creating a chatbot, optimized for a mobile browser, that interviews you like a clinician would, asking progressive questions and narrowing potential diagnostic matches down to three possibilities, giving reasons for or against each one, and explaining the type of care you should seek.
Should you go to the ER? Should you make an appointment with your primary care physician later in the week? Should you watch the issue, and maybe administer some self-care?
Users “go from, ‘Oh! I’m so scared, I’m going to the emergency room!’ to ‘Oh, I can wait at home for a few days and see how it goes’, Le says. “Removing that fear and uncertainty leads to more efficient and ultimately more safe healthcare decisions that are better for everybody.”
There are about 75 clinical scenarios that immediately exit the program and ask the users to call 911, go to an ER, or call the suicide prevention hotline. “We think there are many situations where a computer isn’t appropriate to be interacting with someone. We try to minimize—at the top of the funnel—those scenarios that are totally inappropriate for a computer program like Buoy to be working with them,” Le says.
But for many other end points, the system suggests self-care, Le explains, including recommendations like icing and elevating a sore joint, doing some stretches, and rest. The system may recommend Ibuprofen or Tylenol. And it may suggest that you see your primary care physician within a day or so.
Buoy will not refer users to specialists. “Buoy isn’t interested—for now especially—in sending someone directly to a cardiologist or dermatologist or orthopedic surgeon,” Le says. He believes the technology will likely be capable of that sort of diagnosis with enough users and feedback, but that is not the intent of the company.
“A lot of people look at us and think we’re trying to replace doctors. That is far from what we’re doing. We’re trying to replace the act of searching online to help you figure out where to go when you’re sick. We feel that today the realm of specialist referrals is appropriately owned by clinicians who are using their sound clinical judgement to say whether to send to a specialist or not.”
Cost Savings and Model Finding
Buoy is a free service—accessible at BuoyHealth.com—and anyone can access the site anonymously with no account and no login.
Buoy’s first clients were health systems; the value proposition was to help people navigate to the right physician or the right service. Le reports that Buoy saved the health system about $174 per use of Buoy Health, but perhaps worked a little too well.
“When we started measuring how people were changing their behavior based on what Buoy was helping them understand… what we saw was significant de-escalations of care. People who originally wanted to go to the emergency room, for instance, after using Buoy something like 40% of them ended up de-escalating themselves to a lower level of care.”
It wasn’t a model that “jived” with a fee-for-service health system, Le explains, but it aligned well with insurance providers, the government, and large employers who were paying for employee care.
“We shifted our client base to self-ensured employers and health plans who are ultimately thinking about how you get people to the right care at the right time,” he says.
Now when you use the Buoy Health system, the chatbot asks if you have insurance through your employer. “If you get health insurance through your job or a family member's job, I may be able to show you customized suggestions based on your insurance. Do you receive health insurance through your job or a family member's job?” the system asks. If you share that information, the system can make suggestions specific to your benefits, for example, suggesting a telemedicine provider that is already part of your benefits plan, or showing doctors in your network. If you skip the question, the system returns the three diagnostic matches anyway.
Either way, Le insists that the Buoy is “keeping the AI pure”. There is no difference in the diagnostic matches in the free versus enterprise version.
But how does Buoy know if the AI is doing a good job?
“We think of figuring out what ended up happening to the patient as three different pieces of information: what they said happened, what the claims said happened, and what the EMR said happened,” Le says. For the past two years the company has focused on gathering patient reported outcomes from its large user base. About 15% of users report a doctor’s diagnosis back to the system. In 2020 to 2021, Buoy plans to bring in EMR and claims data, “to further refine and understand what ultimately happened to people.”