Deep Learning-Based Classification of Social Anxiety Disorder Using Continuous Self-Reported Anxiety in Virtual Reality

Social Anxiety Disorder (SAD) is a mental disorder characterized by excessive fear and avoidance of social situations. Traditional assessment methods rely on retrospective self-reports, which may not fully capture moment-to-moment variations in perceived anxiety. To address this, we designed a novel Virtual Reality (VR) scenario to simulate a real-life social situation, specifically a waiting room that gradually fills with virtual characters. A continuous measure of self-reported anxiety was collected via joystick throughout the VR experience, allowing for real-time monitoring of subjective social anxiety. A one-dimensional convolutional neural network (1D-CNN) was trained to classify individuals with SAD based on their reported anxiety trajectories. The model was evaluated using a Leave-One-Subject-Out (LOSO) cross-validation strategy, achieving an F1-score of 0.82, recall of 0.89, and precision of 0.77, demonstrating strong classification performance. These findings suggest that self-reported anxiety alone is a viable signal for distinguishing individuals with SAD, paving the way for more accessible, sensor-free assessment tools in virtual environments. Future work will explore advanced feature extraction from the anxiety signal, integrate physiological markers, and investigate adaptive VR scenarios that dynamically respond to user-reported distress.