Stress can show up in everyday habits: messaging, searching for information and paying for purchases via smartphone. A team at the Korea Advanced Institute of Science and Technology (KAIST) led by Professor Uichin Lee developed the SHIRBT-MMF technology, which automatically finds repeated actions in apps and uses them to assess daily stress levels.
Earlier work relied on screen time, the number of app launches or typing patterns. But when apps and their interfaces change, such data change too, and models become less robust. The new system looks not at the app's name but at the goal of the action: switching from one messenger to another remains "messaging".
The authors called the repeated tasks SHIRBT - Smartphone Interaction Routine Behavioral Tasks. To identify them, the team applied its own method of multi-level sequential pattern mining (ML-SPM). It finds repeated transitions between screens over several days, and a large language model (LLM - artificial intelligence trained on texts) assigns them understandable labels: "messaging", "web browsing", "financial payment", "email work".
When testing whether the same action is classified identically across different email, browser and messenger apps, the system showed accuracy above 95%.
To protect privacy, the developers did not store the message text itself. Only the fact of input, its frequency and transitions between screens were recorded. Data about recipients and interlocutors was processed by hashing - a transformation that does not allow the original information to be directly reconstructed.
The study involved 26 students and office employees. Among those for whom both stress self-reports and smartphone records were available, 16 people provided an average of 30.5 days of data, and 10 - an average of 105.1 days. Based on daily self-reports, personalized models were built that distinguish days with and without stress.
In total, 29 types of repeated actions were identified, with each person regularly performing on average about seven of them. The best model achieved an average accuracy of 75.9% for the short-term group and 73.8% for the long-term group - roughly three correct answers out of four. This is 5.5 and 3.5 percentage points higher than the traditional approach that analyzes app usage with the same AI method. A percentage point is a unit of difference between two percentage figures.
Accuracy held even when the number of actions was reduced. If only "messaging" and "financial payment" were left, or even a single "messaging" was analyzed, the model still classified better than the standard method. On long-term data, metrics based on repeated actions changed less over time and remained more stable.
The authors emphasize: the work revealed a statistical association between behavior and stress but did not prove a causal relationship. The data were obtained from Android users, so the scientists plan to extend the study to other age and occupational groups and to test the method's robustness in different smartphone environments. According to Professor Lee, a smartphone is not just a device for counting time but a space of everyday life where communication, information search and financial operations are repeated.