Columbia: More than 20 percent of Americans will have symptoms of depression or anxiety this year, but only one in five of them will get adequate treatment. Increasingly, people are turning to smart phone apps for help. There’s only one problem: none of the thousands of mental health apps available have so far been adequately tested to see if they provide relief.
To bridge this gap, a group of researchers at Northwestern University developed a collection of a dozen apps that harness techniques from clinical psychology and artificial intelligence with the goal of improving users’ moods. As part of an ongoing National Institute of Mental Health-funded study, Ken Cheung, professor of Biostatistics at the Mailman School, is providing his expertise to test whether these apps work the way they’re intended.
“There is huge potential for smart phones to help people with depression and anxiety,” says Cheung. “But so far we have no idea what works and what doesn’t. Just because an app uses a technique that’s effective in a doctor’s office doesn’t mean it will translate in the context of a mobile app.”
The app suite, called IntelliCare, is available free for Android users and targets common causes of depression and anxiety like sleep problems, social isolation, lack of activity, and obsessive thinking. Worry Knot, for example, gives users techniques to identify and untangle their worries. Social Force helps people reconnect with friends and build support systems. All of the apps track user progress and offer encouragement along the way.
IntelliCare uses a state-of-the-art method developed through machine learning, a branch of artificial intelligence, to “learn” from the way people interact with the software. Anonymous user data trains the system to find techniques that best encourage individual users to meet their mental health goals. The system also provides recommendations for other apps in the suite that could meet their needs much in the way Amazon uses shopping and browsing history to recommend the perfect pair of sneakers.
This recommendation process is also central to how Cheung is able to make comparisons between the apps. An evaluation methodology he pioneered called open-ended adaptive randomization is more flexible than a traditional randomized controlled trial, which has a defined start and finish. “Adaptive design also allows us to evaluate all the apps simultaneously,” he says, “even as people are continuously downloading, using, and discarding any number of them.”
Cheung recently completed a test run using data from 366 users, focusing on 90 days of usage, including every second of user interaction—a huge accumulation of data. To assess the merits of the apps, he explains, it’s crucial to first understand how they’re used. And every app is used in a different way: You shouldn’t expect someone to interact with Worry Knot the same way they would with Social Force. If you spend more time with the former, it doesn’t necessarily mean it’s the more effective app.
“One app might ask you to identify what is making you anxious and analyzes it with you over a course of several minutes,” he says. “Another might make a quick suggestion that you call a friend or take a walk outside.”
To overcome the problem, Cheung developed a toolkit of statistical techniques to dig into the masses of data, tease out five or six signature usage patterns, and subsequently associate these patterns with data from the study’s clinical phase. His Northwestern colleagues have started following people with anxiety or depression who agreed to use the IntelliCare apps for up to eight weeks. They will undergo an initial clinical assessment as well as a series of telephone and online questionnaires about their mood. At the end of this assessment period, the team, working with collaborators at IBM’s T.J. Watson Research Center, will look to see if any of the patterns Cheung identified line up with clinical improvements.
Going forward, Cheung anticipates that the IntelliCare system will continue to adapt to encourage users to use the apps in ways that help alleviate their symptoms. He says this aspect of implementation science is not unlike how commercial app makers refine their apps to make them more engaging and popular, only with clinical outcomes, and not profit, as the ultimate goal.
“Our job isn’t done once we prove one or more of the apps works in a clinical sense,” says Cheung. “They can’t just sit on a virtual shelf. To be successful, they have to be something people use.”