Data model conversion for independent component analysis to extract brain signals

This study addresses an empirical study for data model conversion when using independent component analysis (ICA) to extract brain event-related potentials (ERPs). We firstly prove that in theory there is no difference to perform ICA on the concatenated EEG recordings of a number of single trials and the averaged EEG recordings over those single trials. The general assumption for such conclusion is that mixing models of linear transformations do not change along single trials. Furthermore, we explicitly illustrate that an optimal wavelet filter based on properties of an ERP can convert the underdetermined model of EEG to at least quasi-determined one, but the optimal digital filter based on that ERP cannot make it, through empirical studies. Hence, we suggest combining an optimal wavelet filter and ICA together to extract desired brain signal from the averaged EEG recordings in the ERP study.