Information theoretical analysis of high density electromyographic data for prostheses control

David Hofmann, Armin Biess, Janne Hahne, Bernhard Graimann, J. Michael Herrmann

Research output: Contribution to journalArticlepeer-review


Statistical properties of high density surface electromyographic (sEMG) signals for control of multifunctional myoprostheses are studied by an information-theoretic approach. We address the question, which electrodes and signal features contribute significantly to hand posture classification. For this purpose, we employ information theoretic and heuristic approaches and compare their performance. Methods: In order to assess the informational content of signals available for the control of a transradial hand prostheses we recorded 126 monopolar sEMG signals, see Fig. 1a. The data were obtained from able-bodied subjects performing repeatedly eight different static contractions (hand open and close, wrist flexion, extension, abduction, adduction, pronation and supination). For each electrode i the mutual information I(C, RMS(Eli ,T)) of the root mean square (RMS) with time window length T of its signal and the classes (hand postures) is calculated. Furthermore, we compute the pairwise mutual information I(RMS(Eli ,T), RMS(Elj ,T)). Using those measures, we construct an electrode selection algorithm and compare its performance in terms of classification accuracy with the heuristic electrode selection method sequential floating forward selection (SFFS). For classification linear discriminant analysis (LDA) is sufficient. Results: While SFFS generally outperforms the mutual information based algorithm, we find regimes for feature parameter T, where the performance of the mutual information algorithm reaches that of heuristic approaches. Figure 1. a) Location of the electrode array on the forearm of the participant. b) Classification performance trade-off between feature parameter T and selected number of electrodes by the mutual information algorithm. Performance is color coded ranging from a classification rate of 50% (blue) to 100% (red). c) Information (in bits) of RMS of signals Eli about hand posture ranges from 32.5 (blue) to 38.5 (red) with red regions indicating
Original languageEnglish
JournalFrontiers in Computational Neuroscience
Issue number132
Publication statusPublished - 2010

Fingerprint Dive into the research topics of 'Information theoretical analysis of high density electromyographic data for prostheses control'. Together they form a unique fingerprint.

Cite this