(PS2-24) Identifying Virtual Social Contexts Using Socially Anxious Individuals’ Passively Sensed Physiological Responses

Introduction: Many individuals with social anxiety disorder either do not seek or cannot access treatment, facing logistical, financial, and emotional barriers. Just-in-time adaptive interventions (JITAIs), a form of digital mental health intervention (DMHI) that aims to provide targeted interventions that are adapted based on the user’s context and responses, may help reduce these barriers. To effectively deploy JITAIs for socially anxious individuals, it is necessary to determine when an intervention is needed and what types of interventions are feasible and effective given the current social context (e.g., whether they are alone or with others). Passive sensing approaches offer a potentially unobtrusive way to detect social context in daily life. Accordingly, the primary aim of this work was to identify physiological patterns associated with contexts relevant to understanding social anxiety with the goal of identifying the physiological features that are important for social context detection.

Method: Participants were N = 46 undergraduate students with moderate to severe social anxiety symptoms (≥ 34 on the Social Interaction Anxiety Scale) who completed social (dyadic or group conversations) and non-social (watching a video) tasks virtually via Zoom. The tasks were split into different temporal phases (anticipatory, concurrent, and post-event) and involved different numbers of interaction partners and different levels of evaluative threat. Participants' physiological responding was continuously monitored via an Empatica E4 wristband. We used paired difference tests, random forest models, and cluster analyses to explore physiological response differences across contexts (e.g., social vs. non-social; implicit vs. explicit social evaluation).

Results: Although many significant paired differences were identified across comparisons, only the random forest models comparing social to non-social tasks/phases substantially outperformed our random guessing baseline of 50%. Feature importance analysis found that the top two most important features were accelerometer standard deviation (ACC SD) and the count of skin conductance response peaks divided by the total time (SCR Peaksn). The cluster analysis identified many clusters when comparing social and non-social tasks, suggesting substantial heterogeneity in physiological responses at the individual level.

Discussion: Using our passively sensed physiological data, social context (i.e., social vs. non-social) is more reliably distinguishable than more nuanced contextual features within social interactions (e.g., size of interaction group). Features extracted from accelerometer and electrodermal activity sensors may be particularly valuable for detecting when someone is in a social interaction. The clustering revealed variability across participants’ physiological response patterns, but also several potentially meaningful clusters, indicating that physiological response patterns may not be entirely idiographic. These findings suggest that passively sensed physiological data can be leveraged for social context detection, a key step to advance development of context-sensitive JITAIs.