"MAVISE: Multimodal dataset for analyzing Anxiety experience in a Virtual Immersive Social Environment"

"We present the MAVISE Dataset, the first, large-scale multimodal resource for studying social anxiety in immersive virtual environments. It contains synchronized behavioral and physiological recordings from a large sample of N=108 participants, which capture goal-directed navigation in socially stressful VR scenarios where autonomous agents systematically modulated social pressure. During this goal-finding task, participants exhibited environment-dependent realistic behavioral responses such as changes in locomotion patterns, interpersonal distance keeping, and attentional gaze shifts. After the task, participants provided continuous, frame-level anxiety annotations of their first-person videos. All behavioral, physiological, anxiety, and event-marker annotation streams were synchronized at 120 Hz, including, among others, EDA, PPG, RSP, pupil tracking, and agent\u2013player spatial relations. Event-based analyses revealed clear markers of social anxiety: agent intrusions into personal space increased anxiety, altered locomotion, and elevated sympathetic arousal, whereas goal completion produced sharp reductions across behavioral and physiological indicators. Across-participant deep-learning baselines further show that behavioral features are the dominant predictors of momentary anxiety, outperforming physiology alone and slightly exceeding full early-fusion models. Within physiology, EDA and pupil diameter contributed most. Locomotion-related features were uniquely indispensable, and high-variance target-guided masking improved anxiety prediction performance by emphasizing windows containing meaningful affective transitions. By integrating large-scale multimodal signals, continuous self-reports, and high temporal fidelity, MAVISE provides a unique, comprehensive dataset for research into anxiety in social immersive environments and offers first benchmark baselines for multimodal, predictive models."