Using myoelectric signals to classify prehensile pattern

People want to live independently, but too often disabilities or advanced age robs them of the ability to do the necessary activities of daily living (ADLs). Finding relationships between electromyograms measured in the arm and movements of the hand and wrist needed to perform ADLs can help address performance deficits and be exploited in designing myoelectrical control systems for prosthetics and computer interfaces.;This dissertation presents the results of applying several machine learning techniques to discover the electromyogram patterns present when performing typical fine motor functional activities used to accomplish ADLs. The primary data in this research is from electromyogram and accelerometer signals collected from the arms and hands of several subjects while they performed typical ADLs involving grips or movements of the hand and wrist. Four approaches were developed and tested. One involved classification of 100 ms individual signal instances. The second and third approaches used a symbolic representation called SAX to approximate signal streams. The second created an affinity matrix approach to model the co-occurrence of SAX symbols and classes to classify based on multiple adjacent signal values. The third used nearest neighbor classification with Dynamic Time Warping (DTW) as a distance measure to classify entire activity segments. A fourth approach used a Hidden Markov Model (HMM) to classify continuous movement segments by applying a 'belief' calculation that uses that instance's signal reading as the observation model, the belief values of the previous instance's classes, and estimated transition probabilities. Accelerometer data were systematically used to aid in labelling the data since it clearly indicates the start and stop of dynamic movements.;The findings reported here support the view that grips and movements of the hand can be distinguished by combining electrical and mechanical properties of the task to an accuracy of 76.72% for 47 classes in a segmented approach and 75.09% in a continuous movement approach. Converting the signals to a symbolic representation and classifying based on larger portions of the signal stream improves classification accuracy. More precise labelling and applying the belief calculation gave credible results for the more complex continuous movement scenario. Classification errors were in all approaches predominantly concentrated within particular grip family groups. This is both clinically useful and opens the way for an approach to help simulate hand functional activities. With improvements it may also prove useful in real time control applications.