(PS2-16) Detecting State Anxiety in Virtual Social Interactions Using Passively Sensed Biobehavioral Data

Social anxiety disorder (SAD) is highly prevalent, with some estimates indicating 13% of adults receive this diagnosis in their lifetime. Effective treatments exist, but only a small proportion of individuals with SAD access care. This is attributable to many factors, including mental health stigma, high cost, and limited provider availability. Additionally, our central care delivery model typically involves meeting in person with a provider, which can be a deterrent to individuals who feel anxious about interacting with others. Given the many barriers to receiving care, it is important to develop alternative care options for individuals with SAD. Just-in-time-adaptive interventions (JITAIs) are a promising avenue to increase access to care, especially since they can often be integrated into individuals’ daily lives, without therapist input. To successfully deploy JITAIs, we need to first identify when individuals may benefit from an intervention (e.g., when feeling particularly anxious). One approach is to ask individuals to self-report when they are especially anxious and may benefit from support. However, this can quickly become burdensome. A viable alternative is to leverage passively sensed information – gathered via smartphones and smartwatches—to detect state elevations in anxiety. Work to date indicates it is feasible to use biobehavioral information (e.g., accelerometer, heart rate, location, call and text logs) to unobtrusively detect variations in self-reported anxiety over the course of weeks or days. Given anxiety can shift at a rapid pace (e.g., over the course of minutes), it is important to examine whether we can detect fluctuations in anxiety at more granular timescales. To this end, in the present study, we used machine learning to predict fluctuations in anxiety over 2-6 minute interactions using biobehavioral features. We also explored the benefits of incorporating information about the social context (i.e., the number of interaction partners and the degree of social evaluation) on detection accuracy. We examined these questions among N = 46 undergraduate students high in social anxiety symptoms who completed four social experiences (two dyadic conversations and two group conversations) via Zoom in groups of up to six people. Throughout these experiences, participants wore wristband sensors that captured their heart rate variability, skin conductance, wrist movements, and skin temperature. Using machine learning, we detected whether or not individuals were experiencing an elevation in self-reported, subjective anxiety. Our models correctly identified whether individuals were experiencing elevated anxiety 60% of the time, indicating we can detect this outcome at above-chance levels, though there is clearly room for improvement. Model performance was even better if the model incorporated information about the social context (i.e., if it was a dyadic or group interaction, and the level of social evaluation), such that accuracy rose to 67%. These findings indicate the feasibility of using biobehavioral markers to detect anxiety fluctuations over the course of minutes. Discussion will focus on implications for the development of JITAIs for social anxiety.