Multichannel Finger Pattern Recognition Using Single-Site Mechanomyography

Mechanomyography (MMG) has become a promising method for pattern recognition in upper-limb prostheses because mechanical signals can overcome the inherited unreliable drawbacks of the bioelectric electromyography (EMG) and electroencephalogram (EEG). However, the existing multi-channel pattern recognition methods require multi-site MMG because it is difficult to extract enough useful features from a single-site MMG. Therefore, this paper proposes a multi-channel finger pattern recognition using a single-site MMG signal. First, multi-mode high-dimensional features are extracted from the MMG signals, including statistical features, frequency-domain features, and acoustic features; second, the extracted features are further used for the pattern recognition using artificial neural networks (ANNs). In this study, 14 participants, including 6 males (3 athletes, 3 non-athletes), 6 females (3 athletes, 3 non-athletes) and 2 male amputees (wrist amputation and unilateral transradial) participated in the experiments. The real-time results show that the mean classification accuracy was 95.1% with 2.6 standard deviations (SDs) over the test folds (male athletes), 89.9%, 88.6% with 3.1, 3.7 SD (male non-athletes and female athletes, respectively) and 77.6% and 74.4% with 9.3 and 11.4 SD, respectively (wrist amputation, unilateral transradial) for seven movements. The results show that the proposed method has the capability to individually control fingers.