Estimating no. of sources to carry out BSS using ICA

If I want to apply Independent Component Analysis in order to carry out filtering of EEG signals(say from 20 electrodes), I need to input the number of sources to the algorithm.

How am I supposed to make a reasonable of estimate of how many sources combine to form the EEG signals that I end up recording?

There is no real way of knowing apriori the no. of sources... so how does one go about solving this problem? Ideas anyone?

Answer 1

[Image Analyst]

I don't understand. Are you saying that somehow your 20 EEG signals all got added together to form one signal and now you need to extract back out all 20 individual signals? Why would that happen? How could they all get combined like that in the first place?

Answer 2

[Rohan]

No. I mean to say that many different sources might combine in an unknown manner to yield signals on the scalp. I record 20 such signals which can be assumed to be a linear combination of n different sources. These n sources combine linearly in different proportions to yield 20 instances of EEG signals recorded from the scalp.

The question is how does one estimate the number of sources? What I am trying to solve is just an instance of the famous cocktail party problem.

Eg: ECG signals contaminate the EEG signals and also end up getting recorded in the EEG signals. One can use BSS implemented via ICA to get separate the ECG signal from the EEG signal. This cant be done using standard spectral filtering due to overlapping spectral ranges.