Prosthesis control using a nearest neighbor electromyographic pattern classifier

Abstract

A prosthesis control strategy using a nearest neighbor electromyographic pattern classifier was investigated with both a real time microprocessor-based controller and offline computational facilities. Four active electrodes for myoelectric signal amplitude detection were interfaced with a microcomputer for data logging and pattern classification. A nearest neighbor algorithm correctly identified arm motions as belonging to one of six pattern classes from 72 percent to 100 percent of the time. There were five vectors for each class in the look-up table.

The nearest neighbor pattern classifier was compared to a minimum error rate Bayes classifier under the assumption that the probability densities were distributed as a multivariate normal distribution. Comparable error rates were obtained with the same data vectors.

A condensed nearest neighbor classifier was constructed to determine what minimum number of vectors was necessary in the look-up table. This minimum number of vectors ranged from two to six for the majority of the classes. Larger numbers of vectors were placed in the look-up table for classes that were more difficult to classify.