Project Ekhi: what healthy volunteers can and cannot tell you about VR-based digital phenotyping of anxiety and somatoform dissociation

Abstract

The integration of digital phenotyping, virtual reality (VR), and physiological biosensing into a single reproducible framework remains underdeveloped. Most VR-based neuropsychological instruments stop at the feasibility stage, and few provide normative data from healthy samples. Practical analytical challenges arise when biosignal-driven models are built on non-clinical data. This study evaluates the technical feasibility of Ekhi, a modular VR research platform, by building a proof-of-concept protocol for exploratory classification modeling in healthy volunteers and documenting data characteristics as a methodological reference for researchers planning comparable psychiatric or neuropsychological protocols. Forty healthy adults (community volunteers without a declared psychiatric diagnosis; 21 female, 19 male; mean age 28.6 years) completed a 14-min VR session on a Meta Quest 2 headset comprising five sequential scenarios: office familiarization, forest relaxation, intensive-care-unit stress, a Stroop color-word test with social-evaluative avatar-induced context, and forest recovery. The ECG and EDA were recorded for the purpose of platform validation. Self-report instruments (Beck Anxiety Inventory, Somatoform Dissociation Questionnaire, Simulator Sickness Questionnaire) were administered before and after VR exposure. Physiological reactivity was assessed across conditions using Friedman tests with post-hoc Wilcoxon signed-rank comparisons. The exploratory logistic regression classifiers were evaluated by AUC and LOOCV. All participants completed the protocol with no significant increase in simulator sickness. Self-report scores exhibited pronounced floor effects and non-Gaussian distributions. Friedman tests revealed significant condition differences for all biosignal features (all p < 0.001), with the Stroop task producing the most consistent autonomic shifts. The ICU scenario did not differ significantly from either of the forest conditions. Exploratory classifiers targeting anxiety (BAI) and somatoform dissociation (SDQ-20) reached AUCs of 0.762 and 0.831 in fitted models (n=35), but validated performance was substantially lower (leave-one-out 0.594 and 0.727; nested out-of-fold 0.594 and 0.655), and the classifiers discriminated the non-anxious majority more reliably than symptomatic individuals. Together, these results indicate healthy-sample models characterize a normative reference range rather than flag clinical cases. The Ekhi platform is shown as technically viable, well-tolerated, and capable of capturing physiologically distinguishable VR-induced states. The major observations for future studies – what a healthy sample can and cannot tell researchers – include distributional constraints, feature redundancy, and the limits of transferability of these classifiers to clinical settings.