Optimization of an independent component analysis approach for artifact identification and removal in magnetoencephalographic signals

Abstract

Objective: To propose a noise reduction procedure for magnetoencephalography (MEG) signals introducing an automatic detection system of artifactual components (ICs) separated by an independent component analysis (ICA) algorithm, and a control cycle on reconstructed cleaned data to recovery part of non-artifactual signals possibly lost by the blind mechanism.

Methods: The procedure consisted of three main steps: (1) ICA for blind source separation (BSS); (2) automatic detection method of artifactual components, based on statistical and spectral ICs characteristics; (3) control cycle on ‘discrepancy,’ i.e. on the difference between original data and those reconstructed using only ICs automatically retained. Simulated data were generated as representative mixtures of some common brain frequencies, a source of internal Gaussian noise, power line interference, and two real artifacts (electrocardiogram=ECG, electrooculogram=EOG), with the adjunction of a matrix of Gaussian external noise. Three real data samples were chosen as representative of spontaneous noisy MEG data.

Results: In simulated data the proposed set of markers selected three components corresponding to ECG, EOG and the Gaussian internal noise; in real-data examples, the automatic detection system showed a satisfactory performance in detecting artifactual ICs. ‘Discrepancy’ control cycle was redundant in simulated data, as expected, but it was a significant amelioration in two of the three real-data cases.

Conclusions: The proposed automatic detection approach represents a suitable strengthening and simplification of pre-processing data analyses. The proposed ‘discrepancy’ evaluation, after automatic pruning, seems to be a suitable way to render negligible the risk of loose non-artifactual activity when applying BSS methods to real data.

Significance: The present noise reduction procedure, including ICA separation phase, automatic artifactual ICs selection and ‘discrepancy’ control cycle, showed good performances both on simulated and real MEG data. Moreover, application to real signals suggests the procedure to be able to separate different cerebral activity sources, even if characterized by very similar frequency contents.

Introduction

When studying the cerebral activity through neurophysiological techniques, such as electroencephalography (EEG) and magnetoencephalography (MEG), the main goal is the discrimination of different sources of electrical activity. The first step in this direction is to separate the ‘cerebral’ sources from ‘the others’ generated by non-cerebral sources, these latter often of such intensities that the former ones are hidden (Zappasodi et al., 2001). Typically, a MEG sensor measures the mixture of original source signals with an adjunctive noise (Del Gratta et al., 2001). This mixture comes from both the ‘wanted’ sources and those ‘to be discharged’ (called artifacts).

Independent component analysis (ICA) is a promising approach that can be useful for the elimination of artifacts and noise from biomedical signals. Since the pioneering work of Scott Makeig and collaborators (Makeig et al., 1996) it is generally accepted that ICA is a good tool for isolating artifacts in EEG and MEG data (Delorme and Makeig, in press, Vigario et al., 1997, Ziehe et al., 2000, Ziehe et al., 2001, Cao et al., 2000), but not much has still been achieved, except for the EEG recordings (Vorobyov and Cichocki, 2002, Delorme et al., 2001), about the criteria for selecting artifactual components in practical ICA applications. The aim of this study is to present an automatic detection procedure that can assist researchers in classifying the obtained independent components (ICs) as cerebral or non-cerebral (artifacts and noise) sources based on IC statistical properties (kurtosis, entropy) and spectral characteristics (Power Spectrum Density=PSD).

Moreover, the present method allows better signal reconstruction, i.e. the retro-projection of the automatically retained ICs: the difference between the original recorded data and the reconstructed ones is computed and called ‘discrepancy.’ Whenever a non-artifactual part of discrepancy is identified on the basis of its spectral characteristics, it is added to the reconstructed signal.

Section 2 describes the signal-generating model and each step of the proposed blind noise-reduction procedure; Section 2.1 makes some considerations about mathematical ICA assumptions in a MEG data context and presents the ICA algorithm used; Section 2.2 details the automatic detection method based on statistical and spectral component characteristics. Section 2.3 presents simulated and real data chosen to test the effectiveness and coherence of the proposed method. In Section 3 the results of each step of the procedure are provided. Finally, in Section 4, the significance of the results is discussed.