On Independent Component Analysis based on Spatial, Temporal and Spatio-temporal information in biomedical signals

Abstract

Independent Component Analysis (ICA) as a Blind Source Separation technique has been used in biomedical signal processing applications for over a decade now. A common goal for ICA is in de-noising multiple signal recordings, for artefact removal and source separation and extraction. ICA decomposes a set of multi-channel measurements into a corresponding set of underlying sources using the assumption of independence between the sources as the separation criterion. Most commonly, ICA is applied as ensemble ICA (E-ICA), where a series of spatial filters are derived from the multi-channel recordings giving rise to independent components underlying the measurements. Where single channel recordings only are available or desirable it is not possible to apply the standard E-ICA model. In previous work we have introduced a Single-Channel ICA (SC-ICA) algorithm that can extract multiple underlying sources from a single channel measurement. Whereas E-ICA utilizes spatial information in the multi-channel recordings, SC-ICA utilizes wholly temporal information to inform the separation process. The two algorithms have differing underlying assumptions for the separation process. A natural extension is to combine the information inherent in both spatial and temporal recordings through the use of a Spatio- Temporal ICA (ST-ICA) algorithm. Here we review three implementations of these ICA algorithms as outlined above, and as applied to biomedical signal recordings. We show that standard implementations of ICA (E-ICA) can be lacking when attempting to extract complex underlying activity. SCICA performs well in separating underlying sources from a single measurement channel, although it is clearly lacking in spatial information, whereas ST-ICA uses both temporal as well as spatial information to inform the ICA process. ST-ICA results in information rich Spatio-Temporal filters which allows the extraction of independent sources which are both quasi-spectrally overlapping as well as having very similar spatial profiles — both of which are not possible in SC-ICA and E-ICA respectively.Keywordsblind source separationindependent component analysisspatial ICAtemporal ICAspatiotemporal ICA