Filtering Strategies for Robust Myoelectric Pattern Classification

Abstract

Recent investigations into the use of real-time, pattern recognition based myoelectric control systems have shown excellent results in terms of classification accuracy and limb controllability under clinical supervision. Longer term, continuous use appears to be subject to deterioration in classification accuracy and usability due to factors including electrode displacement, electrode/skin interface impedance, and user variability. In this work, a simple filtering strategy for improved robustness to external noise is introduced. Recorded signals are digitally filtered to remove noise vulnerable frequencies while retaining discriminatory myoelectric information for classification.