Multivariate AR modeling of electromyography for the classification of upper arm movements

Objective We compared the performance of two feature extraction methods for multichannel electromyography (EMG) based arm movement classification. One method was to use a scalar autoregressive model (sAR) for each channel. Another was to model all channels as a whole by a multivariate AR model (mAR). Methods The classified arm movements included elbow flexion, elbow extension, forearm pronation and internal shoulder rotation. Six-channel bipolar EMG signals were collected from four electrodes fixed on the biceps, triceps, brachioradialis and deltoid. Fifteen two-channel and four three-channel configurations were formed out of these six-channel signals for a comparison of different channel combinations. Leave-one-out cross-validation was adopted for evaluating the classification performance using a parametric statistical classifier. Results We processed a total of 216 EMG segments obtained from repeated 18 performances by three normal subjects. mAR model based feature set achieved a better classification accuracy than sAR did for each configuration. Moreover, significance of improvement was greater than 0.95 for those configurations which consisted of EMG channels that were close spatially. Conclusions The stronger the cross-correlation among EMG channels the more improvement of classification accuracy one would expect from using a mAR model.