A woman sits in her office chair staring fixedly at the computer screen before her. Five months earlier she had a heart attack, but her recovery has been going well; in the rehab program, she exercised regularly and made real progress. Now she’s graduated from the program and is back at work. It’s easy to fall into old habits.
The woman shifts but stays seated. In fact, she’s been working on the document now for two hours without budging. And then her phone pings with a message:
Is now a good time for a stroll through the office? Maybe say hello to your colleague at the far end that you rarely see anymore?
She isn’t sure whether to sigh or smile. But she closes the file and heads to the hallway thinking, “Two hundred steps, here I come.”
This particular scenario is imaginary. But it’s a realistic example of a new approach that may fundamentally change how people are supported in coping with a range of medical and behavioral problems, from not exercising to smoking to drug abuse to mental illness.
The method—just-in-time adaptive interventions, or JITAI—borrows part of its name from the as-needed inventory system developed by Japanese car makers in the late 1970s. Except now what’s being provided just in time is a medical or behavioral intervention. “Now we can deliver a treatment in your life at any time you might need it,” explains Susan Murphy, a research professor at the University of Michigan Institute for Social Research (ISR). “At 2 a.m. in a stairwell, we can try to help you.”
An average smartphone already collects a surprising range of information about its user: accelerometers detect motion and walking, GPS systems pinpoint exactly where users are at all times, calendars indicate busyness, and even the number of incoming and outgoing calls can be indicative of a good or bad mental state. Pair that with a wristband and special apps, and all kinds of new interventions become possible.
“This is one!” Murphy says, gesturing to the watch-like wristband on her arm. “Because we are collecting information using sensors, we can try to provide interventions even before you’re aware that you need an intervention.”
Researchers at the University of Michigan (U-M), known for its interdisciplinary prowess, have thrown themselves into this emerging area of research. JITAI projects being tested or planned at U-M include the following:
- Cardiac patients who are transitioning out of the intensive support and supervision of rehab receive messages on their smartphones encouraging them to walk more. Success is measured by the number of steps they take.
- Patients with hypertension who don’t take medications as prescribed receive targeted messages on their smartphones to improve adherence. Special pill bottle caps that record when they’re opened let researchers know whether patients are complying.
- A smoking cessation study, part of a large, eleven university grant, will arm smokers with a wearable wristband, similar to a Fitbit, that senses heightened stress levels that could lead to smoking. The wristband directs smokers to smartphone links with mindfulness exercises and other methods to regulate anxiety.
- A special phone app that analyzes the voices of people with mania or depression as they receive and make calls will learn to spot mood changes so that it can recommend interventions—on the part of the client or a caregiver—before the condition becomes an emergency.
- Patients with alcohol or drug dependencies will carry phones that use GPS to recognize when they’re approaching an area that might present temptation—such as a neighborhood bar. Phone messages and support will help users avoid backsliding.
- Wearable devices will monitor the fluid in the lungs of congestive heart patients so that a smartphone or other device can signal clinicians if the fluid becomes excessive. The goal is to treat the condition before it gets out of hand.
Designing Heart Steps
Creating just-in-time adaptive interventions like these requires a broad cross-disciplinary team. Take Heart Steps, the intervention to encourage cardiac patients to walk more. The team, led by a specialist in human-computer interactions, includes a computer scientist, a behavioral scientist, a cardiologist, and Susan Murphy, who is also the H.E. Robbins Distinguished University Professor of Statistics and a professor of psychiatry at U-M.
Looking at some of the issues this team faced in launching an effective JITAI gives a sense of both the challenges and the promise of the approach. This summer the researchers conducted a 42-day pilot study with participants from the Ann Arbor area interested in walking more. Two more ambitious studies with actual cardiac patients will follow.
Each participant in the pilot received a smartphone with a special app on it and a wristband like Murphy’s. At five set points during the day—the times when research shows people are most likely to exercise—participants either did or did not get a message on their phones encouraging them to walk, stretch, or otherwise get moving. The wrist band then measured the outcome—the number of steps participants took in the hour after getting a message.
Old style behavioral interventions are typically burdensome and quickly discarded. Take a food log for dieting: Tracking every bite may be revelatory at first, but soon becomes drudgery. The drop rate for such aids is “incredibly high,” according to Predrag Klasnja, an assistant professor of information and of health behavior, and the principal investigator of Heart Steps. “They say three-quarters of them get abandoned before the tenth use, and a quarter after the first one.”
By contrast, the Heart Steps team wanted to make its intervention as non-invasive as possible. Participants in the pilot study typically only received two messages a day. That’s because although there were five possible times for messages to arrive, the program basically flipped a coin each time to determine whether or not to send a message. That way, researchers could compare the effectiveness of message vs. no message in different situations. This also had the effect of creating a de facto control group at each time during the study.
The team also made sure the messages—pulled from 180 different categories or buckets— were context appropriate and thus different for each participant. If someone were at home, for example, they wouldn’t get a suggestion to take the stairs at work. Or if they were driving or already walking, they wouldn’t get a message at all. “The idea is to have messages that are highly actionable in the current situation, so it gives people ideas of what they can do right now,” says Klasnja.
Creating the right kind of messages was also critical: If participants felt bossed, they might just turn the app off. So most of the recommendations were couched as friendly suggestions to—for example—pick up some trash at a nearby park, take the stairs, return a stapler, stroll with a child, or just stand up and stretch. You know what’s awesome? asks one message. The sunset! How about looking up tonight’s sunset time & heading out for a walk to check it out?
With the pilot study complete, the team is analyzing the results and creating preliminary decision rules to make the two future studies more effective—what researchers are calling “the warm start.” For example, they may learn that when people’s calendars are really busy, an encouraging message to walk is a waste of time. Still, Klasnja says, “Twenty percent of the time we’ll flip a coin and send a message regardless. And that allows us to figure out, are there some people for whom the current decision rules are not working very well.”
All of this discovery will be brought to bear on the final year-long study. This is the one researchers are most excited about, because the algorithm will actually learn and adapt on the fly, personalizing support based on how people respond. For example, if it sees that one person walks when the weather is bad and another doesn’t, it will tailor the messages they receive to account for that. “By the end of that last study,” Murphy says, “we’ll be able to look at you and me and we’ll see that our decision rules will be very different. That’s the personalization. That’s the learning from you.”
Challenges and promise
Of course, people aren’t easy to learn from. In the past, reinforcement learning systems like this one were used to create programs to play chess or Jeopardy, or to control robots. By contrast, people are “noisy,” says Satinder Baveja, a professor of electrical engineering and computer science who designed the machine learning part of the pill bottle cap medication adherence study at U-M. “Who knows why they some days take their medicine and some days don’t,” he says. “People are busy, they have family over, they’re traveling, it’s a holiday, there’s stress at work. There are all these factors that we have no way of sensing that make it variable.” He adds: “We can’t measure things perfectly. We’re not dealing with buildings or automobiles or things we can instrument to our heart’s desire.”
People being people, their interest in and adherence to the program also is not guaranteed. For example, while messages may work great at first, they can lose potency, particularly if there are a lot of them. “That’s the whole game here,” Murphy says. “You don’t want to be pinging someone all the time, giving suggestions, because then they’re not going to pay attention anymore. So you really want to provide that support just in time. And only enough.”
Inbal Nahum-Shani, an ISR research assistant professor and behavioral scientist working on the multi-university smoking cessation study, says there will even be people who can’t abide getting very many messages at all. Like her. “I’ve tried so many apps in the last couple of years that my tolerance for interrupting messages is gone!” she says, with a wry smile. Still, she believes wholeheartedly in the promise of the approach. “There’s a lot of potential for us to actually help and not interrupt people. They have this phone all the time. They don’t have access to a care provider or a therapist all the time.”
Finally, there’s privacy. Heart Steps gathered GPS data on its participants every hour in order to be able to send location-appropriate messages. (The data, stripped of identifying characteristics, is stored on a secure server.) Gathering that kind of information may not be a problem when the goal is encouraging people to walk, but tracking the movements of alcoholics or drug addicts will likely prove far more sensitive.
“My own sense is that people are willing to share more if two conditions are met: They trust the system and they get value out of it,” says Ambuj Tewari, an assistant professor of statistics and electrical engineering and computer science who is designing the algorithm that allows Heart Steps to learn from participant responses. The trick, he says, will be to hit the right tradeoff between privacy and efficacy.
Overcoming the obstacles will be worth it. Researchers think JITAI will change the nature of health interventions and give patients a degree of control without devaluing human caregivers. According to Tewari, it’s the distinction between AI—artificial intelligence—and IA—Intelligence Amplification. “As a society, do we want machines that will just go and replace more and more humans, or do we want machines that will augment our capacities to do whatever we like?” he asks. “In this case, machines and algorithms will augment the capacity of mental health professionals and behavioral scientists to help more people.”
This could be especially critical in parts of the country and the world where caregivers are in short supply. Tewari, who is from India, says there is an “order of magnitude gap” between the doctors available and the number needed. Yet with the help of smartphones, which are becoming common at all levels of Indian society, that could change. “Right now maybe this mental health professional can only see 10 patients, but augmented with mobile devices and smart algorithms they can deal with 100 patients,” Tewari says. “Suddenly you’ve really increased their reach.”
And even small interventions can be profound. As one Heart Steps message asks.
Wouldn’t the sun feel nice on your face? Taking a walk now would brighten the rest of your day.