Abstract
Objectives: Virtual reality (VR) has emerged as a promising tool for assessing anxiety-related disorders through immersive exposure and physiological monitoring. This study aimed to evaluate whether multimodal data, including heart rate variability (HRV), skin conductance response (SCR), and self-reported anxiety, collected during VR exposure could classify patients with panic disorder and agoraphobia using machine learning models. Methods: Seventy-six participants (38 patients with panic disorder and agoraphobia, 38 healthy controls) completed 295 total VR exposure sessions. Each session involved two road and two supermarket scenarios designed to induce anxiety. Inside the sessions, self-reported anxiety was measured along with physiological signals recorded by photoplethysmography and SCR sensors. HRV measures of heart rate, standard deviation of normal-to-normal intervals, and low-frequency to high-frequency ratio were extracted along with SCR peak frequency and average amplitude. These features were analyzed using Gaussian Naïve Bayes (GNB), k-Nearest Neighbors (k-NN), Logistic Ridge Regression (LRR), C-Support Vector Machine (SVC), Random Forest (RF), and Stochastic Gradient Boosting (SGB) classifiers. Results: The best model achieved an accuracy of 0.83. Most models showed specificity and precision ≥0.80, while sensitivity varied across models, with several reaching ≥0.82. Performance was stable across major hyperparameters, VR-stimulus settings, and medication status. The patients reported higher subjective anxiety but exhibited blunted physiological responses, particularly in SCR amplitude. Self-reported anxiety demonstrated higher feature importance scores compared to other physiological properties. Conclusion: VR exposure with self-reported anxiety and physiological measures may serve as a feasible diagnostic aid for panic disorder and agoraphobia. Further refinement is needed to improve sensitivity and clinical applicability.