direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments


Search for publications

Journal Publications by GRK members

Improved (I)CA-noise elimination of electrophysiological data using band-pass filtered components
Citation key Goergen2009
Author Görgen, K., Bosman, C., Womelsdorf, T., Oostenveld, R., Fries, P.
Year 2009
ISSN 1662-5137
DOI 10.3389/conf.neuro.10.2009.03.257
Journal Front. Comput. Neurosci.
Number 00257
Abstract Removal of signal components unrelated to the signals of interest, so called 'noise', is a general problem in many types of data. One method for identifying and removing noise (e.g. eye movement artifacts or power-line noise) from electrophysiological data is independent component analysis (ICA) (e.g. Jung et al., 1998). A shortcoming of this method is that it does not guarantee a complete separation of physiological signals and noise sources into different components. Thus, those components containing obvious noise contributions might also contain physiological data. This phenomenon is called 'leakage'. Rejecting these components might therefore not only clean the noise but may additionally and erroneously take out parts of the physiological data. Here, we suggest that this leakage can be reduced using prior knowledge about the spectral features of the noise. After identification of the noise components, we applied band-pass filtering to these components at the frequencies of the noise before subtracting the back-projected components from the data. This effectively prevents deletion of physiological signals in frequency bands beyond the a priori specified noise frequency bands. To test this method, we applied it to simulated data and physiological signals obtained from ECoG recordings in macaque monkey. Our results demonstrate that this method (a) is as efficient as the standard ICA correction, (b) also works for data available only as short time segments ('trials'), (c) results in a clearer noise time-course allowing the subtraction of less non-noise activity, and (d) does not - in contrast to the standard unfiltered ICA method - change the power spectra of the cleared signals in spectral regions outside the noise regions. Additionally, we present a simple and computationally efficient method to find noise components with sharp frequency tunings (e.g. line noise), which can be used independently of or prior to ICA to improve its performance. To achieve this, we band-pass filter the original signal prior to component identification (which can consequently be done using standard component analysis techniques, e.g. PCA/ICA). We thereby reduce the influence of components whose main contribution lies in other frequency bands, thereby rendering the noise components more salient to the component identification algorithm. This method is especially useful when no clear noise-components can be detected using standard ICA on the unfiltered data. In summary, we show that ICA noise elimination is improved using prior knowledge about the spectral features of the noise, ultimately enhancing the signal-to-noise ratio of neurophysiological data recordings.
Link to original publication Download Bibtex entry

To top

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions