Locomotion in virtual environments predicts cardiovascular responsiveness to subsequent stressful challenges

The study addresses whether high-dimensional, objective behavioral data can predict physiological vulnerability to stress without exposing individuals to strong stressors. Stress responses, mediated by the ANS and HPA axis, are linked to the development of psychopathologies and cardiovascular disease, but individuals differ in susceptibility. HRV is a robust biomarker of stress resilience/vulnerability; low HRV is linked to psychopathology and CVD, and cardiovascular reactivity can outperform resting measures for predicting disease development. However, existing approaches are often limited by short, non-ecological stressors and few variables. The authors propose leveraging immersive virtual reality (VR) to capture rich locomotor behavior during mildly arousing novelty exploration, inspired by rodent phenotyping tasks, to predict HRV responses during a subsequent stressful VR challenge. They compute a respiration-robust integrated HRV index (iHRV) and use machine learning to test whether exploratory locomotion predicts HRV under stress, with validation and generalization analyses.

Prior work links stress exposure to increased risk for mental and cardiovascular diseases. HRV has emerged as a top biomarker for stress resilience; low tonic HRV predicts psychopathology, CVD, and mortality. While cardiovascular reactivity in laboratory settings can predict disease better than resting measures and shows trait-like stability, traditional stress reactivity paradigms have modest long-term prognostic value due to limited ecological validity and few variables. Digital phenotyping using passive locomotor data from sensors is promising but has not been shown to predict objectively measured stress vulnerability. VR reliably elicits behavioral and physiological responses and provides standardized, controllable environments. Rodent studies show that locomotor patterns in novelty exploration predict later stress susceptibility; human adaptations of open field and elevated plus maze have identified anxiety-related behavior but typically rely on a limited set of features. This work builds on these literatures by integrating high-dimensional VR locomotion features with machine learning to predict HRV reactivity to stress.

Participants: 135 healthy male adults (18–38 years; mean age 20.58 ± 2.05) participated; none reported psychiatric diagnoses or psychotropic medication use. Ethics approval and informed consent were obtained. For modeling, participants were split into training/discovery (N=66; age 21.00 ± 1.93; enriched for low STAI-T <35 and moderately high STAI-T >45) and testing/replication (N=69; age 20.19 ± 2.10; moderate 35<STAI-T<45) sets. A generalization cohort comprised 107 of the 135 participants (age 20.48 ± 2.19) who completed additional stress/anxiety tasks.

VR scenarios and procedure: Participants explored three VR scenarios for 90 s each, starting from a standardized position. Scenario 1 (empty room; human open-field analog) and Scenario 2 (elevated alley with narrowing width; analog to open arms of elevated plus maze/successive alleys) were designed to elicit variance in exploratory locomotion under mild arousal. Scenario 3 (dark maze with flashlight; persistent threat) delivered three white-noise startles (t=20, 40, 60 s) to elicit sustained parasympathetic withdrawal (low HRV). Participants returned to the start position after each scenario.

Generalization tasks: A 10-min VR stress test combining mental arithmetic, social evaluation, negative feedback, and challenging navigation with performance titration and occasional false negative feedback; baseline was recorded in a low-stress VR environment. A sustained anticipatory anxiety paradigm (habituation phase of differential delayed fear conditioning; 2 min, no shocks delivered after warning participants that shocks could occur) was conducted on a separate day; pulse and skin conductance were recorded.

Apparatus and recording: VR used HTC Vive HMD with room-scale tracking; scenes developed in Unity3D; participants moved freely within a 3.5 m × 6 m × 3.5 m room. Lower-body motion capture (MVN XSENS Awinda) captured gait and center of mass movement. Physiology was recorded wirelessly (Biopac Bionomadix, 1000 Hz; ECG and respiration), with ECG decimated to 500 Hz and respiration to 100 Hz; pulse recorded at 250 Hz in the anticipatory anxiety task.

Physiological measures and iHRV: From ECG, HR and HRV metrics (RMSSD, SDNN, HRV triangular index) were computed after artifact correction (Pan–Tompkins R-peak detection, filtering, interpolation). Respiration rate (RR) derived from filtered respiration signal using peak detection. To create a respiration-robust parasympathetic metric, principal component analysis was applied to HR, RMSSD, SDNN, HRVTi, and RR across the VR scenarios; PC1 (iHRV) loaded positively on HRV metrics and negatively on HR, while RR loaded on another component, indicating minimal respiratory contamination.

Behavioral features: High-dimensional locomotor features from scenarios 1–2 were extracted from headset tracking (positioning features such as distances to walls/corners/center, time in zones, velocity/acceleration, trajectory patterns) and motion capture (gait parameters). Movement bursts and immobility features were also computed. An initial set of 172 features was assembled (see Supplementary Tables 1–5).

Feature selection: Conducted on the training set. Removed zero-variance features and those with Spearman correlation |r| < 0.1 with iHRV during the persistent threat scenario. Remaining features were grouped (gait, movement bursts, position tracking). For each group, cross-validated Lasso GLM (MATLAB lassoGLM) predicted iHRV; unshrunk variables retained across groups were combined, yielding 18 features for model training.

Modeling: An extreme gradient boosting regression model (XGBoost, Python XGBRegressor) was trained to predict iHRV during the persistent threat scenario (Scenario 3) from the 18 selected behavioral features from scenarios 1–2. Hyperparameters (eta, max_depth, min_child_weight, nrounds) were tuned via Bayesian optimization (hyperopt). Evaluation used a holdout test set; 10-fold CV on the training data assessed internal performance. Model interpretability employed SHAP values to estimate feature contributions.

Statistical analysis: Used MATLAB, R, and JASP. Normality assessed with Shapiro–Wilk; Spearman correlations used when non-normal. Repeated-measures ANOVAs assessed HR/HRV changes across blocks; Holm correction applied for multiple comparisons. Correlation confidence intervals reported. Physiological measures for VR exploration were computed over each 90 s scenario; for the stress test (10 min) and anticipatory anxiety (2 min), measures covered the full task duration.

• The persistent threat scenario (dark maze) elicited escalating cardiovascular reactivity across three 30 s blocks: HR increased and HRV (RMSSD) decreased (p<0.001), indicating sustained parasympathetic withdrawal. In contrast, HR and RMSSD habituated during the two exploratory scenarios.
• An integrated HRV index (iHRV; PCA PC1) loaded positively on HRV (RMSSD, SDNN, HRVTi) and negatively on HR, while respiration rate loaded on a different component, yielding a respiration-robust parasympathetic index.
• Prediction performance of the XGBoost model using 18 locomotor features from scenarios 1–2 to predict iHRV during the threat scenario: • Training set: r=0.91, p<0.001 (10-fold CV shown in supplement). • Test/replication set: r=0.54, p<0.001. • Combined (train + test): r=0.72, p<0.001.
• SHAP analysis identified top contributors: minimum distance to corners (empty room), vertical acceleration (elevated alley), ratio of time center/periphery (empty room), time on the narrowest ledge (elevated alley), movement focus (empty room), stride speed variability (empty room), number of head scans while walking (empty room), time on starting board (elevated alley), and maximal longitudinal distance (elevated alley). Closer corner proximity and higher vertical acceleration were associated with lower iHRV (stronger parasympathetic withdrawal).
• Model predictions correlated with individual physiological measures: RMSSD r=0.60, SDNN r=0.48, HRVTi r=0.34 (all p<0.001), and HR r=-0.52 (p<0.001); no correlation with respiration rate (r=-0.11, p=0.199).
• The model outperformed other potential predictors: trait anxiety (STAI-T), state anxiety (STAI-S), and self-reported threat anxiety showed no correlation with iHRV (all |r|<0.1, p>0.386). Time in center (empty room) showed a small correlation (r=0.23, p=0.008) far below the model’s prediction–iHRV correlation (r=0.72).
• Generalization to other stressors: In a separate 10-min VR stress test (N≈107), predictions correlated with HRV (RMSSD) r=0.38, p<0.001 and HR r=-0.40, p<0.001. In a sustained anticipatory anxiety paradigm (N≈104), predictions correlated with pulse rate variability r=0.24, p=0.016 and pulse rate r=-0.32, p<0.001.
• Overall, rich locomotor behavior during novelty exploration robustly predicts interindividual differences in vagally mediated HRV reactivity to subsequent stress, outperforming anxiety questionnaires and single behavioral metrics.

The findings demonstrate that high-dimensional locomotor behavior during mildly arousing novelty exploration in VR predicts individual differences in parasympathetic cardiac control (iHRV) during a subsequent stressful challenge. This reverse-translational approach from rodent open-field and elevated-plus-maze paradigms reveals that behavioral patterns often construed as anxiety-like (e.g., thigmotaxis, avoidance of open/narrow spaces, reduced exploration) and decision/locomotor features (movement focus, stride speed variability) meaningfully contribute to predicting physiological stress responses. The model’s predictions align with HRV measures and oppose HR, are robust to respiration confounds, and outperform self-reported trait/state anxiety and prototypical single behavioral readouts, highlighting the value of integrating multiple behavioral features. Validation in an independent cohort and generalization to distinct stressful and anticipatory anxiety tasks support prognostic utility beyond the original context. VR-based standardized, reproducible environments permit efficient, ecologically relevant elicitation and measurement of behavior and physiology, offering advantages over traditional brief laboratory stressors. The approach suggests pathways for digital phenotyping and transfer learning to clinical or translational settings, potentially enabling early identification of stress-vulnerable individuals.

Behavioral phenotyping in immersive VR, combined with machine learning, can predict parasympathetic HRV responsiveness to subsequent stress. By integrating multiple locomotor features from exploratory tasks, the model provides a precise, respiration-robust predictor that generalizes across stressors and outperforms anxiety questionnaires and single behavioral metrics. This work lays the groundwork for scalable behavioral digital phenotyping tools to detect stress vulnerability and guide preventive strategies. Future directions include extending validation to females, benchmarking against established laboratory stressors (e.g., TSST, SECPT), longitudinal and clinical cohort studies, integration with wearable/mobile sensing, and iterative/transfer learning to improve prediction across populations and contexts.

• Sample restricted to males; generalizability to females remains to be tested.
• Benchmarking versus gold-standard non-VR stress paradigms (e.g., TSST, SECPT) was not conducted; comparative predictive and elicitation efficacy is unknown.
• Real-world applicability may be constrained by differences between VR-induced and naturalistic stressors, though generalization tests were encouraging.
• Wearable-derived short-term HRV reliability remains a challenge for real-life deployment; the study used controlled lab-grade ECG/pulse recordings.
• The predictive model relies on a specific set of VR tasks and feature engineering; external validation in diverse settings and with different hardware is needed.
• Longitudinal predictive validity for health outcomes or psychopathology was not assessed.