JMIR Publications

ABSTRACT

Background:

Digital mental health training programs increasingly aim to strengthen stress regulation, resilience, and performance in high-demand environments. However, many existing solutions rely on static or self-paced content that does not adapt to an individual’s real-time mental state, limiting their effectiveness for both skill acquisition and transfer to real-world performance. Immersive virtual reality (VR)–based stress exposure training (SET), combined with physiological sensing and machine learning (ML), offers a scalable digital mental health training approach that supports active skills practice while also addressing operational readiness requirements in high-stress occupations.

Objective:

This study aimed to develop and validate supervised machine learning models to predict subjective task load during immersive VR-based stress exposure training and to identify and assess key physiological, behavioral, and psychological predictors of task load, in order to inform adaptive digital mental health training solutions that support both mental health skill development and operational readiness.

Methods:

Twelve participants completed a within-subject repeated-measures protocol involving immersive VR training scenarios designed to elicit emotional, cognitive, physical, and combined task demands, interspersed with baseline and recovery periods. Training-relevant data streams included heart rate variability (HRV), facial electromyography (EMG), perceived stress, demographic variables, and subjective workload measured using the NASA Task Load Index (NASA-TLX). Time-series features were extracted using a standardized multimodal data fusion and feature engineering framework. Predictive models were developed for six NASA-TLX domains using a cross-validated recursive random forest feature-selection and support vector machine (CV-rRF-FS-SVM) pipeline. Model interpretability was examined using Shapley Additive Explanations (SHAP).

Results:

Across all NASA-TLX domains (mental, physical, temporal demand, effort, frustration, and performance), ML models demonstrated strong predictive performance, with cross-validation and holdout receiver operating characteristic area-under-the-curve (ROC-AUC) values ranging from 0.78 to 0.99. Key predictors of task load included HRV indices (e.g., root mean square of successive differences (RMSSD), standard deviation 1 (SD1), interbeat interval (IBI)), facial EMG activity (particularly frontalis and corrugator muscle activation), perceived stress levels, VR scenario context, and participant age. SHAP analyses provided interpretable insights into how physiological and facial expression features differentially contributed to specific workload dimensions, supporting transparent inference for training applications.

Conclusions:

This proof-of-concept study demonstrates that subjective task load during immersive digital mental health training can be reliably inferred using explainable ML models and multimodal physiological data. Embedding these models within VR-based training systems has the potential to enable adaptive regulation of stress exposure and task difficulty, supporting personalized self-regulation and resilience skill development while maintaining relevance to operational readiness demands. This integrated approach advances digital mental health training solutions in support of bridging mental health skill acquisition with performance under real-world stress.