A Filter approach for myoelectric channel selection

For the control of upper limb prostheses, machine learning algorithms are increasingly common for disriminating different patterns of the surface myoelectric signal (MES). Sophisticated myoelectric controllers usually record data from multiple bipolar channels, placed on muscle groups of interest. The appropriate number of channels is a delicate question. One usually tries to minimize the amount of channels required while maintaining reasonable classification performance. This paper presents a filter approach to the channel selection problem by exploiting properties of the principal component analysis. Out of a set of channels measured on the patient, a ldquogoodrdquo subset is selected for further processing by a pattern recognition algorithm. The method is applied to data recorded from an amputee who has undergone targeted muscle reinnervation (TMR) surgery. It is shown that the amount of channels can be reduced with only minor decrease of classification performance.