- System detects daily stress from repetitive phone behavior with up to 75.9% accuracy

- Only behavioral patterns analyzed, not message content, enabling long-term mental health monitoring

From bottom left: Professor Lee Eui-jin, Dr. Lee Han-su, Professor Lee Jae-gil, undergraduate Park Tae-hyun and postdoctoral researcher Ko Young-ji. [KAIST]
From bottom left: Professor Lee Eui-jin, Dr. Lee Han-su, Professor Lee Jae-gil, undergraduate Park Tae-hyun and postdoctoral researcher Ko Young-ji. [KAIST]

South Korean researchers have developed an AI technology capable of gauging a person's stress levels from everyday smartphone activities such as chatting, browsing the web and making payments.

KAIST announced Thursday that a research team led by Professor Lee Eui-jin of the School of Computing had developed an AI technology called SHIRBT-MMF, which automatically identifies repetitive behaviors on a smartphone and analyzes the duration, frequency and text-input patterns of each behavior to determine a user's daily stress state.

The research team focused not on which apps a person used, but on what they actually did with their phone. The key insight was that switching from KakaoTalk to Line or WhatsApp does not change the underlying activity — the user is still "chatting."

The team defined such recurring activities as "smartphone interaction-based routine behavior tasks," or SHIRBT.

Using a proprietary analytical method, the researchers identified screen-transition patterns that repeated across multiple days, then used an LLM to classify behaviors into categories such as messenger chatting, web browsing, financial payments and email composition.

To protect user privacy, the system does not store the actual text typed or the content of messages. Instead, it records only behavioral data such as keystroke counts and screen transitions, and applies hash processing to information about messaging and call contacts.

The team validated the technology by analyzing real smartphone usage data from 26 university students and office workers in their 20s and 30s.

Data from 16 short-term participants were collected over an average of 30.5 days, while data from 10 long-term participants spanned an average of 105.1 days. Personalized AI models were trained against daily stress self-assessments recorded by each participant.

The analysis identified 29 distinct repetitive behaviors across participants' smartphone use. On average, each individual routinely performed seven of those behaviors.

The AI model achieved stress-detection accuracy of 75.9 percent on average for short-term participants and 73.8 percent for long-term participants — outperforming conventional app-usage-based analysis methods by 5.5 percentage points and 3.5 percentage points, respectively.

A schematic diagram of the research findings. [KAIST]
A schematic diagram of the research findings. [KAIST]

Notably, the system maintained its performance even when the number of behavioral categories used in the analysis was sharply reduced.

Using only two behaviors — messenger chatting and financial payments — or even chatting alone still yielded higher average detection accuracy than conventional methods.

When applied to long-term data, the repetitive-behavior-based indicators also proved more stable over time than existing approaches, showing less variation as the observation period lengthened.

This means stress changes can be tracked over an extended period by analyzing only a person's key behaviors, without collecting a full log of all smartphone activity. The technology is expected to support the development of personalized mental health monitoring services that reduce the burden of data collection.

"We viewed the smartphone not merely as a device for measuring screen time, but as a lifestyle space where an individual's communication, information-seeking and financial activities recur," Professor Lee said. "We have laid a foundation for more stably analyzing long-term changes in mental health by converting complex smartphone records into behavioral units that people can understand."

The researchers noted, however, that the study examined the association between smartphone behavior and stress — it does not establish that any specific behavior causes stress, nor is it a tool for diagnosing mental illnesses such as depression or anxiety disorders.

The team plans to expand the research to participants of diverse ages and occupations and to verify the long-term analytical stability of the system.


nbgkoo@heraldcorp.com