Authors: Josh D Gubler, Connor D Olsen, Fredi R Mino, Mingchuan Cheng, Jacob A George
Mentors: Connor Olsen
Insitution: University of Utah
The long-term goal of this research is to investigate how lower sampling rates of electromyographic (EMG) signals affect the performance of classification and regression algorithms. EMG signals measure the electrical activity of muscle contractions. Myoelectric interfaces can classify or regress features generated from the EMG signal to control devices like prostheses, exoskeletons, robotic systems, or human-computer interfaces. Most of the power of the EMG signal is contained between 50 and 500 Hz, and most recording devices sample EMG at 1 kHz with a 5-15 Hz high-pass filter and a 375-500 Hz low-pass filter. As myoelectric devices become wireless and integrated with wearable technology, reducing the sampling rate can substantially reduce battery consumption and processing power. We sampled EMG data at 30 kHz from the forearms of three participants while they performed six gestures. We then downsampled to rates ranging between 50-1000 Hz and calculated various EMG features from the downsampled data. We found significant effects for both EMG feature and sampling rate on regression performance of a modified Kalman Filter (p < 0.05, two-way ANOVA). The mean-absolute-value and waveform-length EMG features performed significantly better at low frequencies (<250 Hz) in contrast to zero-crossing, slope-sign-change, and mean-frequency EMG features (p < .05, multiple pairwise comparisons). Sampling rate also had a significant impact on the classification accuracy of a k-nearest neighbors algorithm (p < 0.05, two-way ANOVA). However, sampling rate had no impact on classification accuracy for a continuous Convolutional Neural Network (CNN) (p > 0.05, two-way ANOVA). Future work will validate the effectiveness of this CNN as a control modality when using downsampled EMG from wearable sensors. If proficient control can be achieved from down sampled EMG, this could substantially improve battery life and make EMG a more practical biosensor for wearable devices.