Compact Deep Learning Models and Auxiliary Methods for Robust Myoelectric Control


Myoelectric pattern recognition-based upper limb prostheses measure electromyographic (EMG) signals from the residual limb and learn to identify muscle activity patterns that correspond to intended gestures. To train an accurate pattern recognition controller, it is essential that the training signals typify signals measured in real-world scenarios. When these conditions are met, clinical systems enable accurate and intuitive prosthesis control. However, routine usage of a prosthesis gives rise to signal nonstationarities that cause dataset shifts (ie. changes in the joint distribution of classifier input and output). These shifts reduce classification accuracy and render control ineffective. In this dissertation, I examine common sources of dataset shift that affect myoelectric pattern recognition and propose clinically feasible approaches to improve control robustness. First, I investigate the effects of limb position and external load on real-time pattern recognition control and show that a modified training data collection protocol can eliminate these effects in amputee users. Next, I combine data augmentation and deep learning techniques to build classifiers that are tolerant to multi-channel signal noise originating from the electrode-skin interface. Finally, I quantify dataset shift across long-term prosthesis usage and use continual learning with deep neural networks to reduce classifier recalibration frequency. Together, these methods provide a foundation for clinical implementations of advanced deep learning controllers that are robust under dataset shift.

Alternate Identifier
Date created
Resource type
Rights statement