Improving the Robustness of Pattern Recognition-Based Myoelectric Prostheses

An upper-limb amputation is a life-changing procedure severely impacting the individual's ability to perform every-day tasks. Prosthetic devices have been designed to provide some relief to these individuals. Myoelectric prostheses have received significant attention in recent years as they have been designed to look more natural and provide the user with enhanced degrees of freedom (DoF) over the traditional body-powered prostheses. Myoelectric prostheses rely on the acquisition and processing of signals attributable to muscle activity within the residual limb. This signal is known as the electromyogram (EMG), representing the surface recording of electrical activity of muscle fibers. EMG signals recorded from an array of electrodes on forearm are used as signals to decode patterns of dexterous hand movement. Specifically among the myoelectric-based control schemes, pattern recognition-based approaches provide the user with immediate access to multiple DoFs. This functionality promotes a more intuitive experience over those schemes which provide the user with access to only one DoF at a time. In the case of pattern recognition-based prostheses, a set of features are extracted from the EMG signal recorded from the user's residual limb from one or more sites and patterns of muscle activity are ``learned''. Subsequently, when a known pattern of muscle activity is observed, the prosthetic device is actuated and moved appropriately. Despite the advantages of pattern recognition-based prostheses over other prosthetic control methods, the robustness of the control scheme in real-world use remains an issue. Many factors experienced during real-world use have been shown to negatively impact the ability of the system to correctly predict the intended action of the user. This work is dedicated to enhancing the robustness of pattern recognition-based myoelectric prostheses thereby making the devices more reliable, useful, and accepted by those using the devices. The specific focus of the work is to improve the robustness of the devices pertaining to variations in limb position. In the introduction, background information regarding myoelectric prostheses and previous work to improve their robustness is presented. Following this introductory information and having established the need to address the robustness of myoelectric prostheses, an analysis of the effect limb position has on extracted features of EMG from able-bodied subjects is discussed. Subsequently, a thorough investigation of this effect is presented through an experiment conducted with able-bodied subjects and amputee subjects both wearing and not wearing their prostheses. It is found that a particular strategy of training in multiple positions should be employed to optimally reduce the negative effects of the limb's position on EMG features. Specifically, it is concluded that when using Linear Discriminant Analysis to classify time-domain features of EMG during discrete hand and wrist actions, a two-stage position specific classification method does not outperform a system in which data from multiple positions are aggregated to form a single classifier. After a discussion of the impact of this research, directions of future research are suggested with supporting preliminary experimentation