Spatial Feature Extraction for Classification of Nonstationary Myoelectric Signals

We compare classifiers for the classification of myoelectric signals and show that the performance can be improved by using spatial features that are extracted by independent component analysis. The obtained filters can be interpreted as reflecting the spatial structure of the data source. We find that the performance improves for several preprocessing algorithms, but it affects the relative performance for various classifiers in different ways. A critical performance difference is especially seen when non-stationary signal regimes during the onset of static contractions are included. Although a practically utilizable performance appears to be reached for the present data set by a certain combination of classification and preprocessing algorithms, it remains to be further optimized in order to keep this level for more realistic data sets.