Limitations of ICA for Artefact Removal

This paper reports analysis of the limitations of using independent component analysis (ICA) for biosignal analysis especially artefact removal. The possible difficulty is that there are limited number of electrodes (recordings) making it an overcomplete problem (non-square ICA). The other difficulty is the distribution of biosignal being close to Gaussian. These two properties of the signals may make these outside the standard ICA application. This paper reports that ICA is able to successfully separate the biosignals if the number of recordings are not less than the number of sources. If that is not the case, ICA separates artefact component only when the corresponding artefact is predominant. The experiments demonstrate that the results are not reliable and hence the authors recommend that caution should be exercised before using ICA for such applications