Improving Context-Aware Habit-Support Interventions Using Egocentric Visual Contexts

摘要

Habits are automatic actions in stable contexts and can help sustain behavior change. State-of-the-art context-aware habit-support interventions are, however, in predominantly non-visual contexts (time, geolocation, and physical activity), and not in egocentric visual contexts, e.g., eating, meeting people, etc. Using a survey, N=51 participants, we identified the user-desired visual contexts and interventions mediums for habit support. Based on our survey, we created a wearable system, named PAL, with privacy-preserving on-device deep learning, to deliver real-world habit-support interventions in personalized visual contexts. In our 4-week study, N=10 participants, interventions using PAL's personalized visual contexts led to ＞75% increase in habit formation, compared to <40% increase in habit formation using interventions in only non-visual contexts. The habits also persisted in the post-study evaluations 1 and 10 weeks later. Thus, PAL，s interventions using personalized visual contexts improve real-world habit formation for sustainable behavior change.