Mo Zhou and Yonatan Mintz at track field with blurred runner in background.

Mo Zhou and Yonatan Mintz, graduate students in the lab of Prof. Anil Aswani, conducted a study that showed how machine learning can help people exercise more. (Photo by Daniel McGlynn)

New exercise app uses machine learning to keep goals within reach

Finding the motivation to exercise can be tough, as anyone who’s trying to be more active can tell you. But now, Berkeley researchers have developed an exercise app that boosts performance by automatically tweaking goals based on the user’s physical activity level, and keeping goals within reach.

Interface of new exercise app developed by Berkeley researchers.A new exercise app developed by Berkeley researchers uses machine learning for adaptive goals. Instead of offering pep talks from coaches or pushing the same goal whether it is met or not, the app uses a machine learning algorithm to monitor performance and adjust goals as needed to maintain motivation.

“Some exercise apps automate exercise goals — say 10,000 steps a day — but they can’t adapt to an individual’s success in reaching the goal,” says Anil Aswani, assistant professor in the Department of Industrial Engineering and Operations Research, who oversaw the research. “As a result, the goals can get out of sync with actual performance, whether the user is exceeding the goal or falling behind.”

In a controlled experiment, people receiving adaptive goals upped their number of steps per day by about 1,000 — roughly an extra half-mile of walking — compared to those who received constant step goals of 10,000 steps per day.

For the user, the app works in a straightforward way, says Mo Zhou, a graduate student in Aswani’s lab, who led the research.

The machine-learning capability monitors how many steps a person walks or runs, and uses that to set a goal that is challenging but attainable, she says. If someone slacks off, the goal will decrease so that it stays within a realistic range and the person isn’t as likely to lose motivation.

This is the first study to use machine learning to automatically yield personalized and adaptive goals, she says.

Personalized apps tend to rely on meetings with coaches who may offer encouragement and revise the goals if needed. But if this feedback can be automated, the cost of using the app goes down, Zhou and Aswani said.

The app was tested in a ten-week randomized, controlled experiment with 64 participants, aged 24 to 65. Through the app, the intervention group received fully automated, adaptively personalized daily step goals, and the control group received constant step goals of 10,000 steps per day. Both groups were able to track their steps in real-time on the app.

After ten weeks, both groups had slipped a bit in their daily step count. But on average, the group with the adaptive goals computed by the machine learning algorithm app logged about ten percent more steps per day than the control group receiving constant goals.

“Right now, about 50 percent of the U.S. population is not active enough,” Zhou says. “We think this type of personalized and adaptive app can be more effective than other approaches and increase levels of physical activity.”

The researchers are now investigating the potential of using a range of push notifications to increase motivation — from stern warnings to messages that essentially say “keep it up,” or those that suggest breaking up activity into small chunks to boost performance.

Yonatan Mintz, another graduate student in Aswani's lab, also co-authored this study. The team plans a larger study, incorporating the current algorithm along with a new one they have developed to assess what type of push notifications best increase motivation.

The research results were published recently in JMIR Mhealth and Uhealth.


Topics: Industrial engineering, Health, Research, Students, Faculty