2021: Demystifying Independent Component Analysis (ICA) (Plenary)

in an easy to digest manner for those who are less mathematically inclined. ICA is increasingly gaining popularity in the EEG-processing community, yet not many industry professionals truly understand what the analysis is doing “under the hood”. This has led to a multitude of debates on whether or not ICA should be utilized for the EEG at all [Friston, 1998]. This talk aims to provide concrete information about the theory and usage of ICA with respect to the EEG. To begin, this talk will go over the basic matrix equation that ICA algorithms attempt to solve: S = WX (where S is the matrix of independent components, W is the “Mixing matrix”, and X is the raw EEG) [Langlois et al, 2010]. Then the talk will discuss the 5 key assumptions that provide the foundation of ICA and discuss whether the EEG is a suitable subject for independent component analysis under these assumptions [Ullsperger & Debener, 2010]. The 5 assumptions are as follows: 1 – Statistical independence between each source, 2 – The mixing matrix must be square and full rank, 3 – No external noise, 4 – Data must be centered, and 5 – Source signals must not be gaussian. The gaussian assumption is difficult to truly prove when it comes to an EEG, so we will discuss why that is and how ICA can still be implemented [Onton & Makeig, 2006]. After understanding the general ICA assumptions, we will compare the unique underlying procedures and assumptions of the three most widely used ICA algorithms in the field: Infomax, fast-ICA, and AMICA [Palmer at al, 2011]. These comparisons will be accompanied by ICA examples created in EEGLAB, ISync, and WinEEG. To conclude, the presentation will go over some important clinical considerations for those who want to implement ICA, such as the amount of data required for a good recording, how to maximize the accuracy of your results, and when ICA may fail.
Presented by: Kody Newman