Artificial Intelligence-Integrated Virtual Reality in Mental Health Care: A Scoping Review of Current Evidence, Clinical Applications, and Future Directions

Results

Eighteen studies met the eligibility criteria. The evidence base was heterogeneous, generally small, and primarily focused on feasibility and predictive modeling, with limited outcome-focused randomized trials. Geographically, studies were concentrated in South Korea (6/18), followed by China (2/18), with single studies from Romania, Spain, Poland, Canada, the United Kingdom, Italy, and Israel, and multinational studies from Ireland and Romania, Saudi Arabia and Pakistan, and France and the United States.

Clinical applications fell into four domains: assessment and screening, prediction and monitoring, treatment augmentation via adaptive or biofeedback-supported VR, and usability and implementation evidence. Anxiety spectrum conditions were the most commonly studied, including social anxiety disorder, public-speaking anxiety, acrophobia, and panic disorder with agoraphobia. PTSD-related applications focused on military personnel and post-disaster populations. Additional applications addressed stress quantification, biophilic affective response, adolescent depression screening, neurocognitive rehabilitation, and AI-enhanced VR self-talk for psychological counseling.

Most studies used immersive head-mounted display VR combined with multimodal physiological sensing (EEG, heart rate, electrodermal activity, heart rate variability, motion) and machine-learning analysis. AI methods included convolutional neural networks, ensemble tree-based models (random forest, XGBoost, LightGBM, CatBoost), Gaussian process classifiers, regression models, reinforcement learning, and large language model-enabled virtual agents. AI roles spanned offline post hoc classification or prediction, real-time embedded classification or biofeedback, and adaptive personalization.

Reported outcomes were predominantly technical performance rather than clinical effectiveness. Anxiety classification during VR exposure reached high accuracy using combined heart rate and electrodermal features. EEG-based stress classification achieved up to approximately 96 percent accuracy across three stress levels. Multimodal prediction models achieved F1 scores up to 0.842 for social anxiety subdomain prediction and 0.706 for VR sickness. A CNN-based real-time anxiety classifier outperformed comparator models, reaching approximately 75 percent accuracy in a VR therapy context. A Gaussian process classifier with explainability analysis reached around 81 percent accuracy for PTSD classification using physiological habituation features. Other studies reported successful discrimination of panic disorder and agoraphobia from controls, prediction of early treatment response in panic disorder using a CatBoost classifier, and preliminary symptom improvement and reduced session completion time in adaptive AI-assisted VR exposure for specific phobia.

Feasibility and usability findings were generally favorable. Off-the-shelf headsets, short session durations, and low cybersickness rates were reported. Adaptive biofeedback pipelines combining EEG and heart rate variability were implementable in VR exposure workflows. Qualitative feedback on EEG-based intelligent agents and reinforcement learning VR tasks indicated good acceptability among older adults with cognitive decline and their caregivers. AI-enhanced VR self-talk was acceptable for guided self-reflection.

However, key limitations were consistent across studies: small samples, single-site designs, sensor dropouts, limited external validation, heterogeneous outcome measures, and minimal direct comparison between AI-integrated VR and standard VR or usual care. Adverse-effect and safety reporting was sparse. Most evidence remained at the proof-of-concept stage, with insufficient data on long-term outcomes, real-world clinical workflow integration, and equity or subgroup performance.