Open-Environment Evidential Learning for Reliable Myoelectric Locomotion Prediction

Accurate locomotion prediction in dynamic environments is crucial for lower-limb exoskeletons to provide walking assistance. Surface electromyography (sEMG)-based deep learning models demonstrate great potential for decoding user intent, yet most existing approaches are developed under restrictive closed-world assumptions. These limitations hinder adaptability to ambiguous or unexpected locomotion modes in real-world deployment, increasing the risk of unsafe control decisions. To address this gap, we propose the Open-Environment Evidential Learning (OEEL) framework, which integrates evidential deep learning and out-of-distribution augmentation to enhance the reliability of predictions. Evaluations show that it achieved an average prediction accuracy of 96.04% across eight subjects for five locomotion modes in closed environments. More critically, under open-environment conditions, OEEL maintained 79.53% accuracy on known classes and improved failure detection performance by 26.67% (measured by the risk-coverage curve) over a conventional convolutional neural network. By reliably quantifying prediction uncertainty and detecting novel or ambiguous movements, OEEL provides a critical framework to advance myoelectric control systems toward real-world lower-limb exoskeleton applications, where safer operation in unpredictable environments is paramount.