Exploring EEG-Guided Virtual Reality-Based Attention Training for Stress Detection and Reduction: A Machine Learning Approach

We investigate the potential of technology-based attention training (AT), particularly virtual reality (VR), as a stress-management tool. Mental stress is rising globally, and researchers increasingly use immersive technologies, wearable sensors, and machine learning (ML) for its detection and control. This feasibility study examines the impact of fully immersive VR-based AT on mental stress using electroencephalogram (EEG) signals and automated classification. We designed virtual exercises targeting different attention types and analyzed EEG responses with an ML framework; the resulting dataset, collected at the Arab Academy for Science and Technology (Alexandria, Egypt), is publicly available. For an unbiased estimate, we adopt a leakage-free evaluation in which the train/test split is performed by time, before segmentation into overlapping windows, so neighboring windows cannot appear in both sets. Subject-specific tree-based classifiers detected stress with a mean accuracy of about 97%, whereas leave-one-subject-out (LOSO) validation yielded about 67%, indicating strongly individual stress signatures and motivating a subject-specific strategy. Using these models, we compared the number of classifier-predicted stress segments before and after AT and visualized the feature space with T-distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP). Under subject-specific models the number of stress-predicted segments decreased after AT (Wilcoxon signed-rank p<0.05); however, because this reduction was corroborated neither by a non-circular (LOSO) detector nor by a centroid-separation measure, we interpret it as an exploratory, classifier-predicted effect rather than independently validated stress reduction. The results highlight the promise—and the current limits—of integrating immersive VR with EEG-guided analytics for mental-health support.