Deep learning for improved myoelectric control

Abstract

Advancement in the myoelectric interfaces have increased the use of myoelectric controlled robotic arms for partial-hand amputees as compared to body-powered arms. Current clinical approaches based on conventional (on/off and direct) control are limited to few degree of freedom (DoF) movements which are being better addressed with pattern recognition (PR) based control schemes. Performance of any PR based scheme heavily relies on optimal features set. Although, such schemes have shown to be very effective in short-term laboratory recordings, but they are limited by unsatisfactory robustness to non-stationarities (e.g. changes in electrode positions and skin-electrode interface). Moreover, electromyographic (EMG) signals are stochastic in nature and recent studies have shown that their classification accuracies vary significantly over time. Hence, the key challenge is not the laboratory short term conditions but the daily use. Thus, this work makes use of the longitudinal approaches with deep learning in comparison to classical machine learning techniques to myoelectric control and explores the real potential of both surface and intramuscular EMG in classifying different hand movements recorded over multiple days. To the best of our knowledge, for the first time, it also explores the feasibility of using raw (bipolar) EMG as input to deep networks. Task are completed with two different studies that were performed with different datasets. In the first study, surface and intramuscular EMG data of eleven wrist movements were recorded concurrently over six channels (each) from ten able-bodied and six amputee subjects for consecutive seven days. Performance of stacked sparse autoencoders (SSAE), an emerging deep learning technique, was evaluated in comparison with state of art LDA using offline classification error as performance matric. Further, performance of surface and intramuscular EMG was also compared with respect to time. Results of different analyses showed that SSAE outperformed LDA. Although there was no significant difference found between surface and intramuscular EMG in within day analysis but surface EMG significantly outperformed intramuscular EMG in long-term assessment. In the second study, surface EMG data of seven able-bodied were recorded over eight channels using Myo armband (wearable EMG sensors). The protocol was set such that each subject performed seven movements with ten repetitions per session. Data was recorded for consecutive fifteen days with two sessions per day. Performance of convolutional neural network (CNN with raw EMG), SSAE (both with raw data and features) and LDA were evaluated offline using classification error as performance matric. Results of both the short and long-term analyses showed that CNN and SSAE-f outperformed the others while there was no difference found between the two. Overall, this dissertation concludes that deep learning techniques are promising approaches in improving myoelectric control schemes. SSAE generalizes well with hand-crafted features but fails to generalize with raw data. CNN based approach is more promising as it achieved optimal performance without the need to select features.