Learning Transient Electromyographic Activities during Lower-limb Motion Initiation using Gaussian Mixture Hidden Markov Models

Real-time decoding of locomotion initiation processes using surface electroencephalogram (sEMG) is an emergent topic with significant theoretical and practical implications in e.g. human interaction with intelligent assistive robots. An open yet important question is how machine learning using dynamic Bayesian networks can learn the early, prior-to-motion transient EMG activities – not only in the moving leg but also in the stance leg. To address this question, this preliminary study designs and evaluates a Gaussian Mixture Model - Hidden Markov Model (GMM-HMM) in the context of predicting 3 classes of human lower-limb movements, namely, left, middle, and right kicks. Using these unsupervised learning models on EMG recordings from the stance leg only, Bayesian classification achieved an average prediction accuracy of 86.80% in three out of five test subjects. We also analyzed the hidden-state-transition patterns during the motion initiation process. This study demonstrated the feasibility of using unsupervised dynamic Bayesian learning models to capture the predictive dynamic EMG activities in the stance leg during locomotion initiation.