Registration of EMG Electrodes to Reduce Classification Errors due to Electrode Shift

Abstract

Non-invasive recording of EMG signals from the arm of a typical subject or amputee has been popularized in control of a variety of devices, including upper limb prostheses. One of the most difficult challenges of using external recording devices, such as the Myo Armband, is the need to retrain a movement classifier due to the shift in positions and electrode location around the arm. Electrode shift causes distortion of the features to be extracted for classification and makes previous training unusable. For amputees, this means retraining movement classifiers several times per day. In this experiment, the Myo Armband is used to test the ability to predict the degree of electrode shift from the electrode sites used to originally train a classifier in order to correct by the detected shift and continue to use the same classifer, instead of training a new one. The Myo Armband was rotated around the arm of subjects with intact limbs as they performed six commonly used movements. The mean absolute value of each electrode was used to characterize the response at each electrode site. Shifts in orientation between one position and a new position were identified by minimizing the mean-squared error of their characteristic movement profiles. The correct shift was identified across subjects using only 0.25 s of data with over 90% accuracy using the “open” or “wrist supinate” grips. New movements at a shifted location were classified using the feature vectors of a previously collected training set and accounting for the shift; classification error averaged 95.7 ± 0.4%, indicating a possibility for real-time correction of electrode shift error.