Researchers have developed an artificial intelligence system that identifies daily stress by analysing repeated behavioural patterns on smartphones, rather than tracking which apps a person uses.
A team at the Korea Advanced Institute of Science and Technology (KAIST) developed the technology, called SHIRBT-MMF. Previous smartphone-based stress research has primarily analysed factors such as screen time, app launches and typing behaviour. However, when apps or interfaces change, these data also change, reducing the long-term stability of stress-tracking models.
The KAIST team focused instead on what a person does rather than which app they use. Whether someone switches from one messaging platform to another, the underlying activity remains “messenger chatting.” The researchers defined these repeated behaviours as Smartphone Human Interaction Routine Behaviour Tasks (SHIRBTs) and used a large language model to assign understandable labels such as “messenger chatting,” “web browsing” and “financial payment.”
The system achieved more than 95% accuracy in consistently classifying the same behaviour across different apps. To protect privacy, it recorded only behavioural information such as whether text was entered and how often, without storing actual message content.
The team analysed real-life data from 26 university students and office workers aged in their 20s and 30s. Using personalised AI models trained on these behavioural patterns, the best-performing model achieved 75.9% accuracy for short-term participants and 73.8% for long-term participants in distinguishing between days with and without stress.
The researchers noted that the study identified associations between behaviour and stress but did not demonstrate causality. Further research will expand to broader age groups and smartphone environments.
Source: Lee H et al. Leveraging smartphone human interaction routine behavior task mining and modeling for daily stress monitoring. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (2025). DOI: 10.1145/3770644
