Recognition of forearm muscle activity by continuous classification of multi-site mechanomyogram signals

Recent studies on identifying multiple activation states from mechanomyogram (MMG) signals for the purpose of controlling switching interfaces have employed pattern recognition methods where MMG signal features from multiple muscle sites are extracted and classified. The purpose of this study is to determine if MMG signal features retain enough discriminatory information to allow reliable continuous classification, and to determine if there is a decline in classification accuracy over short time periods. MMG signals were recorded from two accelerometers attached to the flexor carpi radialis and extensor carpi radialis muscles of 12 able-bodied participants as participants performed three classes of forearm muscle activity. The data were collected over five recording sessions, with a ten-minute interval between each session. The data were spliced into 256ms epochs, and a comprehensive set of signal features was extracted. A pattern classifier, trained with continuously acquired signal features from the first recording session, was tested with signals recorded from the other sessions. The average classification accuracy over the five sessions was 89±2%. There was no obvious declining trend in classification accuracy with time. These results show that MMG signals recorded at the forearm retain enough discriminatory information to allow continuous recognition of hand motion across multiple (>90) repetitions, and the MMG-classifier does not show short-term degradation. These results indicate the potential of MMG as a multifunction control signal for muscle-machine interfaces.