Towards Robustness in Pattern Recognition Based Myoelectric Prosthesis Control

Commercial myoelectric prosthetic systems are based on relatively simple signal acquisition and processing. Academic research has indeed been developing powerful algorithms with promising results (e.g., pattern recognition methods and linear or non-linear transformations [1-4]), but there are still no commercial systems available using these approaches because they are not robust enough for clinical applications. This is due to the fact that academic research often assumes ideal conditions which are not present during daily life activities [5]. These algorithms fail when changes in arm or electrode position occur. Therefore, we evaluate the changes in myoelectric activation patterns under these two forms of non-stationarities and propose heuristic signal acquisition and processing methods to alleviate their effects on classification accuracy, adapting the methods proposed by [6] and [7].