Smartphone Habits Can Detect Stress With Up to 76% Accuracy, Scientists Find

The findings were published in the journal Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies. The research is scheduled to be presented at the ACM UbiComp/ISWC 2026 conference, which is taking place in Shanghai from October 13 to 15.

Rather than tracking which apps a person uses, the system analyzes what they actually do on their smartphone, such as messaging, searching the internet, making payments, or working with email.

On October 15, KAIST announced that a team led by Professor Woochul Lee of the Department of Computing and Artificial Intelligence had developed a technology called SHIRBT-MMF. It identifies recurring activities and takes into account their duration, frequency, and text-entry patterns.

Previous methods of assessing stress using smartphone data relied primarily on screen time, the number of times apps were launched, and activity within specific app categories.

However, when users switched apps or their interfaces were updated, the nature of the collected data could change, reducing the reliability of long-term monitoring.

The new system analyzes recurring behavioral patterns regardless of the specific app being used. For example, switching from KakaoTalk to LINE or WhatsApp does not change the nature of the activity itself — messaging. Using a sequential pattern-mining method developed by the researchers, the system identifies these patterns over several days, while a large language model assigns them descriptive labels such as “messaging,” “web browsing,” “financial transactions,” and “email use.” When tested on its ability to classify activities across different apps, the system achieved an accuracy rate of more than 95%.

To assess the relationship between everyday habits and stress, the researchers analyzed data from 26 university students and office workers aged 20 to 39. In the short-term group, 16 participants provided an average of 30.5 days of observations, while the long-term group of 10 participants was monitored for an average of 105.1 days. Participants assessed their own stress levels daily, and the researchers trained personalized models to distinguish between stressful and non-stressful days.

In total, the system identified 29 types of recurring activities. On average, each participant regularly engaged in about seven of them. The best-performing model detected stress with an accuracy of 75.9% in the short-term group and 73.8% in the long-term group. These results were 5.5 and 3.5 percentage points higher, respectively, than those achieved using the traditional approach based on app usage analysis.