A new method for unsupervised analysis of spontaneous MEG/EEG data: combination of projection pursuit and parallel factor analysis
-
1
University of Helsinki, Finland
-
2
Helsinki Univ of Tech, Finland
Background: Analysing spontaneous brain activity, for example in resting state, requires exploratory or unsupervised data analysis methods. Independent component analysis is widely used in fMRI (van de Ven et al, HBM, 2004; Beckmann et al, Philos Trans Royal Soc, 2005) but its application to spontaneous MEG/EEG may not be straightforward. Recently, parallel factor analysis has been applied on a time-frequency representation (Miwakeichi et al., NeuroImage, 2004). Here, we develop a new computational method for analyzing spontaneous MEG/EEG, combining ideas from projection pursuit, parallel factor analysis, and independent component analysis.
Methods: We model each MEG/EEG channel as a sequence of time-frequency elements similar to wavelets. In contrast to most methods, we learn the form of such atoms from the data. The data are thus modelled in the spirit of trilinear (or three-way) analysis methods such as parallel factor analysis. First, each channel is divided into time windows, giving data with three indices: X(c,w,t) where c is the channel, w is the index of the window, and t is the time index inside the window. The method is a three-way generalization of projection pursuit: We search for linear combinations of the channels which are as "structured" as possible. Such linear combinations are effectively matrices, indexed by w and t. We consider a matrix to be structured if it has a low rank or is close to a low-rank matrix. A very low-rank matrix, in effect, leads to a structure in which a small subset of the learned time-frequency elements is repeated with the same time courses for all the channels involved in that linear combination. This means that the time courses and energy distributions across the channels are separable in the same way as in parallel factor analysis. The method has been implemented as maximization of an index of "structuredness". The method thus finds a number of components, each of which is characterized by a distribution across the channels, and further analysis allows associating a temporal envelope and frequency distribution to each component.
Results: Preliminary analysis (see Figure) of resting-state MEG data (204 planar gradiometers) from a single subject shows that the method finds meaningful components, including parieto-occipital 10-Hz rhythms, as well as Rolandic mu rhythms with frequency peaks around 11-12 Hz and 22-23 Hz (components 2 and 5). In many cases, a component describes activity in one hemisphere only. Shown are 10 components in the order of "structuredness".
Discussion: The method is generally applicable to resting-state or natural-stimulation MEG and EEG as an exploratory tool. It can also be applied on data gathered in more conventional paradigms to search for any surprising effects or artefacts that are not found by traditional analysis methods. A Matlab implementation of the method will be made publicly and freely available.
Conference:
Neuroinformatics 2008, Stockholm, Sweden, 7 Sep - 9 Sep, 2008.
Presentation Type:
Poster and Short Oral Presentation
Topic:
Electrophysiology
Citation:
Hyvarinen
A,
Parkkonen
L,
Ramkumar
P and
Hari
R
(2008). A new method for unsupervised analysis of spontaneous MEG/EEG data: combination of projection pursuit and parallel factor analysis.
Front. Neuroinform.
Conference Abstract:
Neuroinformatics 2008.
doi: 10.3389/conf.neuro.11.2008.01.097
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
25 Jul 2008;
Published Online:
25 Jul 2008.
*
Correspondence:
Aapo Hyvarinen, University of Helsinki, Helsinki, Finland, aapo.hyvarinen@helsinki.fi