@CONFERENCE\{IMM2008-05545,
    author       = "P. M. Rasmussen and M. M{\o}rup and L. K. Hansen and S. M. Arnfred",
    title        = "Model Order Estimation for Independent Component Analysis of Epoched {EEG} Signals",
    year         = "2008",
    keywords     = "EEG; Event related potentials; Independent component analysis (ICA); Molgedey Schuster; {TDSEP};Model Selection, Cross validation.",
    booktitle    = "Biosignals 2008 International Conference on Bio-inspired Systems and Signal Processing.",
    volume       = "",
    series       = "",
    editor       = "",
    publisher    = "",
    organization = "",
    address      = "",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/5545-full.html",
    abstract     = "In analysis of multi-channel event related {EEG} signals indepedent component analysis (ICA) has become a widely used tool to attempt to separate the data into neural activity, physiological and non-physiological artifacts. High density elctrode systems offer an opportunity to estimate a corresponding large number of independent components (ICs). However, too large a number of ICs leads to overfitting of the {ICA} model, which can have a major impact on the model validity. Consequently, finding the optimal number of components
in the {ICA} model is an important problem. In this paper we present a method for model order selection, based on a probabilistic framework. The proposed method is a modification of the Molgedey Schuster (MS) algorithm to epoched, i.e. event related data. Thus, the contribution of the present paper can be summarized as follows: 1) We advocate {MS} as a low complexity {ICA} alternative for {EEG}. 2) We define an epoch based likelihood function for estimation of a principled unbiased ’test error’. 3) Based on the unbiased test error measure we perform model order selection for {ICA} of {EEG}. Applied to a 64 channel {EEG} data set we were able to determine an optimum order of the {ICA} model and to extract 22 ICs related to the neurophysiological stimulus responses as well as ICs related to physiological- and non-physiological noise. Furthermore, highly relevant high frequency response information was captured by the {ICA} model."
}