I have not come across any reference that states the ideal number of ICA components. But EEGLAB has plugins like ADJUST toolbox, which can help identify eye-movement and other non-neural artifacts using multiple parameters (topography pattern, reccurence across trials, etc). This can help us remove artifactual ICs in a data-driven manner. Another plugin FASTER does most of the data reduction in a fully automated manner (from preprocessing before ICA to ICA to ADJUST toolbox application).

Another approach that we used in our study was to use a PCA based automated artefact removal algorithm of "clean_raw" plugin before ICA analysis. This helps in avoiding artifactual transients from influencing ICA analysis, and drastically reducing ICA computation time. So the resulatnt ICA components were then used for source localisation using DIPFIT plugin, and the ICs that did not fit inside the head space and having residual variance>40% (poor fit) could be removed from further analysis. 

Hope this helps...

Thank you Dr Arun

Aha..Thank you Stephen for your answer

Dear Dr. Kabbara,

number of ICs has been fiercely debated in FMRI research, in which several groups provided computational heuristics to estimate optimal number. I don't know how well these applications can be readily applied to EEG data given different numbers of channels and noise structures (also, algorithms are applied before IC decomposition), but two potentially interesting methods are:

BIC / MDL in GIFT package (Matlab) by Vince Calhoun and colleagues (http://mialab.mrn.org/software/gift/index.html)

Laplacian method in probabilistic ICA (Melodic, part of FMRIB's FSL package) (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/MELODIC)

However, recent empirical studies suggest to estimate larger numbers of components (typically a higher number than the computational heuristics would suggest) in order to capture interesting but less strong data patterns. Sounds reasonable to me, as ICA overfitting is useually less costly in terms of 'errors' than underfitting. Accordingly, FMRI data can be analysed with 80+ numbers of components. Note, though, that many such approaches in FMRI are aimed at detecting components of interest, instead of finding noise components to remove. 

Finally, the ICAsso package by Aapo Hyvarinen and colleagues may be of interest: It uses self-organizing clustering and resampling techniques to run ICAs of a dataset multiple times and extract components that replicate across repeated decompositions - thereby finding out which components are strong and reliable across decompositions (this works, as the seed point for ICA is random). Fabrizio Esposito and others used this approach to get an empirical sense of number of optimal components, that is, the set size at which most components replicate at repeated decompositions. The ICAsso toolbox (Matlab-based) has also been applied to MEG and EEG data, so you should be able to make it work for you.

Again, my ICA expertise is mainly with FMRI data, but I hope this is helpful to you.

References:

ICAsso: Himberg et al. Neuroimage 2004: 1214-1222 (http://www.cs.helsinki.fi/u/ahyvarin/papers/Himberg04reprint.pdf)

Assessing replicability across decompositions:

Esposito et al. Neuroimage 2005: 193-205; van de Ven et al. Human Brain Mapping 2008: 1450-1461

Kind regards,

Vincent.

Thank you Vincent for your helpful answer.