Classification of low level surface electromyogram using independent component analysis

There is an urgent need for a simple yet robust system to identify natural hand actions and gestures for controlling prostheses and other computer-assisted devices. Surface electromyogram (SEMG) is a non-invasive measure of the muscle activities but is not reliable because there are a multiple simultaneously active muscles. This study proposes the use of independent component analysis (ICA) for SEMG to separate activity from different muscles. A mitigation strategy to overcome shortcomings related to order and magnitude ambiguity related to ICA has been developed. This is achieved by using a combination of unmixing matrix obtained from FastICA analysis and weight matrix derived from training of the supervised neural network corresponding to the specific user. This is referred to as ICANN (independent component analysis neural network combination). Experiments were conducted and the results demonstrate a marked improvement in the accuracy. The other advantages of this system are that it is suitable for real time operations and it is easy to train by a lay user.