Electromyography Analysis by Classification Complexity Estimation

Intuitive control based on myoelectric pattern recognition (MPR) can be used in clinical applications such as prosthetic limbs and Phantom Limb Pain treatment. Electromyography (EMG) patterns representing limb movements are learned by a pattern recognition algorithm to enable classification of future EMG observations. These EMG patterns are commonly constituted by descriptive features extracted from raw EMG. The complexity of the classification task is highly influenced by both the selection of such features and the differentiation between movements in the raw EMG. A reliable estimation of classification complexity would facilitate selection of features and elimination of detrimental EMG patterns. Two such algorithms, Separability Index and Nearest Neighbor Separability, were found to be highly correlated with classification accuracy and enable efficient feature selection for three classifiers commonly used for MPR (Linear Discriminant Analysis, Multi-Layer Perception and Support Vector Machine). The algorithms were implemented in the data analysis and feature selection modules of BioPatRec, an open source tool developed at Chalmers University of Technology for development and benchmarking of algorithms in MPR. The implementation included dedicated graphical user interfaces to ease visualization. This thesis deepens the understanding of the complexity of MPR and provides tools for prediction of classification performance and analysis of MPR applications.