On the identification of an ICA algorithm for auditory evoked potentials extraction: A study on synthetic data

Abstract

Extracting characteristics and information from Auditory Evoked Potentials recordings (AEPs) involves difficulties due to their very low amplitude, which makes the AEPs easily hidden by artifacts from physiological or external sources like the EEG/EMG, blinking, and line-noise. To tackle this problem, some authors have used Independent Component Analysis (ICA) to successfully de-noise brain signals. However, since interest has been mainly focused on removing artifacts like blinking, not much attention has been paid to the quality of the recovered evoked potential. This is the AEP case, where literature reports interesting results on the de-noising matter, but without an objective evaluation of the AEP finally extracted (and the influence of different implementations or configurations of ICA). Here, to study the performance of three popular ICA algorithms (FastICA, Ext-Infomax, and SOBI) at separating AEPs from a mixture, a synthetic dataset composed of one Long Latency Auditory Evoked Potential (LLAEP) signal and the most frequent artifacts was generated. Next, the quality of the independent components (ICs) estimated by such algorithms was measured by using the AMARI performance index (Am), the signal interference ratio index (SIR), and the time required to achieve separation. Results indicated that the FastICA implementation, with the symmetric approach and the power cubic contrast function, is more likely to provide the best and faster separation of the LLAEP, which makes it suitable for this purpose.