Pilot comparison of surface vs. implanted EMG for multifunctional prosthesis control

Abstract

The classification accuracies of controllers utilizing EMG input from six surface and ten intramuscular recordings are compared. In addition, the effect of including autoregressive (AR) parameters into the input sets is examined. The average accuracies from four subjects are reported. It was observed that surface recordings based solely on amplitude data did not perform well (21.1% error) but adding AR coefficients increased this accuracy substantially (10.3%). The intramuscular recordings performed comparably to the surface recordings with AR coefficients using all ten (13.2%) and a smaller set of six (12.1%) channels of intramuscular data. The subset of six channels was selected using multinomial logistic regression. It was observed that adding the AR coefficients to the intramuscular recordings also produced an improvement in classification accuracy for the six (92.8%) and ten (93.7%) channel input sets. To our knowledge this is the first work in more than three decades that explores the use of intramuscular EMG for the control of upper-limb prostheses and this work demonstrates that it is possible to achieve a decrease in classification error of nearly 40% by using intramuscular recordings.