Optimising extended reality training: a Bayesian network approach to predicting learning and transfer outcomes

Abstract The present work sought to develop a deeper understanding of the complex interactions between features of extended reality (XR) training interventions and successful training outcomes using a Bayesian Network (BN) model. We extracted data from 41 previous independent-group XR training studies, focusing on technology features (e.g., body visualization, visual fidelity, haptic feedback, personalization) and training design elements (e.g., continuous challenge, explicit feedback, training duration). We sought to elucidate the causal relationships between these variables to predict training success (both learning of the training task and transfer to new contexts) in sensorimotor tasks. The structure of our BN model was informed by expert knowledge, prior literature, and causal reasoning. The joint probability distribution over the model was then learned from the data extracted from previous studies. The model’s performance was examined using cross-fold validation, yielding high precision (85%) and accuracy (76%). Performing probabilistic inference on the final model revealed that haptic feedback and body visualization provided little benefit for learning of the training task, but did improve transfer of training. Meanwhile, continuous challenge emerged as the most influential training design feature for increasing the probability of transfer. Overall, the technology and training design features that were beneficial for transfer were often not beneficial for task learning, highlighting a critical distinction between short-term performance gains and the long-term retention and adaptability of skills. These findings contribute to a comprehensive understanding of how specific technology and training design features influence XR training success, paving the way for new theoretically-grounded applications of immersive technology.