Dimension reduction: Additional benefit of an optimal filter for independent component analysis to extract event-related potentials

► An optimal filter (OF) conforms to the linear superposition rule. ► An OF is designed according to the temporal and spectral properties of a desired ERP. ► An OF does not change coefficients of the linear transformation model of ERP data. ► Number of sources in ERP data may be reduced by an OF. ► An OF might make the underdetermined model of ERP data be determined. The present study addresses benefits of a linear optimal filter (OF) for independent component analysis (ICA) in extracting brain event-related potentials (ERPs). A filter such as the digital filter is usually considered as a denoising tool. Actually, in filtering ERP recordings by an OF, the ERP’ topography should not be changed by the filter, and the output should also be able to be modeled by the linear transformation. Moreover, an OF designed for a specific ERP source or component may remove noise, as well as reduce the overlap of sources and even reject some non-targeted sources in the ERP recordings. The OF can thus accomplish both the denoising and dimension reduction (reducing the number of sources) simultaneously. We demonstrated these effects using two datasets, one containing visual and the other auditory ERPs. The results showed that the method including OF and ICA extracted much more reliable components than the sole ICA without OF did, and that OF removed some non-targeted sources and made the underdetermined model of EEG recordings approach to the determined one. Thus, we suggest designing an OF based on the properties of an ERP to filter recordings before using ICA decomposition to extract the targeted ERP component.