[Eeglablist] Applying Independent Component Analysis (ICA) to EEG Data with Mixed Characteristics

[Chista N]

 Dear EEGlab Community,


I am seeking insights into the application of Independent Component
Analysis (ICA) on EEG data with mixed characteristics.


I understand two critical criteria for applying ICA:

1. ICA expects the data to be stationary, implying the same statistical
model generates all data samples.

2. Sufficient samples are needed to complete ICA training.

My questions are:

1. Between these criteria, which holds greater priority?
2. If neither criterion is met, what are the implications for the ICA
results? Are they entirely unreliable?


Regarding the first criterion, I know that EEG data is typically not
wide-sense stationary (WSS) across its entirety, but ICA is commonly
applied to EEG data of durations longer than a few seconds.

 How does this duration affect the requirement for stationary input data?

Expanding on this, consider EEG recordings under different conditions, such
as a 20-second rest followed by 60 seconds of stimulation. Can ICA be
applied to the entire dataset? (Based on what I know that is common
practice)

If dividing the data into separate parts isn't feasible due to insufficient
samples, what is the recommended approach?


In my specific case, I have a less than 1-minute signal (Fs=140 Hz)
comprising a seizure activity lasting around 40 seconds and a DC-like
activity lasting less than 10 seconds.

How should I proceed with applying ICA to this data?


I appreciate all your help
Best