Stress prediction from smartphone data using continual learning

What is it about?

Smartphones capture many aspects of daily life, such as movement, conversations, and app usage, which can signal changes in stress. Most AI models, however, are trained once and cannot adapt when a person’s habits shift. This paper examines how continual learning can address this limitation by allowing a model to adapt to new users while retaining knowledge from previous ones. We compare a memory-based strategy with a prompt-based method designed to improve efficiency and stability during adaptation.

Featured Image

Photo by Gift Habeshaw on Unsplash

Photo by Gift Habeshaw on Unsplash

Why is it important?

This work highlights the importance of personalization in mobile mental-health technologies. Behavioral data collected passively through smartphones is inherently dynamic, and models that rely on static training fail to capture this evolution. By showing that continual learning strategies maintain predictive accuracy across shifting user behaviors, our study underscores the need for adaptive modeling techniques in real-world stress assessment. These contributions advance the development of reliable, scalable, and clinically relevant stress-monitoring systems.