Pattern recognition algorithms have been proposed as a way to control powered lower limb prostheses, specifically for transitioning between the different pre-programmed locomotion modes of the prosthesis (e.g., level ground walking, stair ascent, etc.). However, these algorithms cannot track changes in the statistical characteristics of input signals, and do not generalize well to novel users. Adaptive algorithms that can update their parameters by incorporating new data are a promising solution to these problems. The purpose of this work was to develop and evaluate an adaptive pattern recognition algorithm for controlling powered lower limb prostheses. The algorithm that was developed could track changes in electromyographic (EMG) signals (whose signal quality degrades over time) and could also generalize to novel users. To accomplish these tasks, this algorithm had to: 1) be able to both detect EMG disturbances and ignore the disturbed EMG data to prevent errors, 2) accurately label new patterns of data with the correct label representing the user’s intent, and 3) adapt system parameters to create an updated pattern recognition algorithm. We developed a metric for detecting disturbances in the EMG signals that was based on the probability of observing current EMG signals in comparison to the history of previously observed EMG signals. This metric could reliably detect a wide variety of disturbances, including electrode liftoff, electrode short-circuiting, and electrode shifts caused by donning and doffing the prosthesis. We developed a compensation technique to be used in the circumstance that EMG signal changes were detected: EMG signals were ignored and only embedded mechanical sensor signals were used to make locomotion mode predictions. This technique prevented many of the errors associated with these disturbances. A technique for automatically labeling new patterns of data was also developed: mechanical sensor data acquired after the user completes a stride with prosthesis could be accurately classified as one of the locomotion modes of the prosthesis. This technique, termed backwards estimation, accurately provided labels for new patterns of data with performance that was consistent across days and robust to changes in the user’s gait patterns. We evaluated the complete adaptive algorithm (comprised of EMG disturbance detection and labeling via backwards estimation) in a multi-day experiment with transfemoral amputee subjects. Amputee subjects ambulated with a powered lower limb prosthesis as the adaptive algorithm updated the model of EMG data in real-time. The results of this experiment demonstrated that an adaptive algorithm could be used to track EMG signal changes, resulting in low and consistent algorithm error rates over long-term use. A preliminary study on across-user adaptation also demonstrated that the model of mechanical sensor data could be updated, resulting in a system that learned from novel users and improved performance over time. The research in this dissertation is significant in the field of lower limb prosthetics because it provides a viable method to automatically update pattern recognition algorithms with new training data while the user is ambulating. This allows pattern recognition algorithms to incorporate new EMG information, and allows novel users to ambulate with prosthesis without participating in long experiments to collect training data. The impact of this research is that lower limb pattern recognition algorithms can maintain low error rates over long-term use.

Last modified

• 02/13/2018

Creator

• John Andreou Spanias

DOI

• https://doi.org/10.21985/N2M96G

Subject

• Biomedical Engineering

Keyword

• Prosthetics
• Adaptation
• Rehabilitation
• Engineering
• Machine Learning

Date created

• 2016-01-01

Resource type

• Dissertation

Rights statement

• In Copyright