Abstract
Social anxiety disorder (SAD) is one of the most prevalent mental health conditions, affecting approximately 13% of the U.S. population during their lifetime. Left untreated, SAD can lead to serious functional impairment and reduced quality of life. Yet, many individuals with SAD remain untreated due to systemic barriers in traditional care delivery. This gap highlights the need for scalable, accessible interventions beyond conventional treatment models. Just-in-Time Adaptive Interventions (JITAIs), delivered via smartphones, offer a promising approach. JITAIs aim to provide individuals with the right type of treatment, at the right time, and only as much as is needed. To achieve this for SAD, two technical challenges must be addressed: (1) the ability to passively infer social context, enabling interventions to be tailored to the moment, and (2) the ability to predict availability and receptivity, ensuring interventions are delivered when individuals are open to receiving them. This dissertation advances both capabilities. First, we examine how passively sensed physiological signals vary with social context. Second, we develop state-of-the-art models for predicting intervention availability and interest using the contextual bandit problem formulation, a framework well-suited for real-time decision-making and adaptive learning. Finally, we introduce a novel online learning algorithm inspired by episodic memory and reinforcement learning, demonstrating its ability to improve contextual bandit performance. Together, these contributions establish a passive sensing and online learning framework for social context inference and receptivity prediction, advance the state of the art in JITAI personalization, and move us closer to the deployment of real-world JITAIs for SAD.
