Automatic correction of ocular artifacts in the EEG: a comparison of regression-based and component-based methods

https://doi.org/10.1016/j.ijpsycho.2004.03.007Get rights and content

Abstract

A variety of procedures have been proposed to correct ocular artifacts in the electroencephalogram (EEG), including methods based on regression, principal components analysis (PCA) and independent component analysis (ICA). The current study compared these three methods, and it evaluated a modified regression approach using Bayesian adaptive regression splines to filter the electrooculogram (EOG) before computing correction factors. We applied each artifact correction procedure to real and simulated EEG data of varying epoch lengths and then quantified the impact of correction on spectral parameters of the EEG. We found that the adaptive filter improved regression-based artifact correction. An automated PCA method effectively reduced ocular artifacts and resulted in minimal spectral distortion, whereas ICA correction appeared to distort power between 5 and 20 Hz. In general, reducing the epoch length improved the accuracy of estimating spectral power in the alpha (7.5–12.5 Hz) and beta (12.5–19.5 Hz) bands, but it worsened the accuracy for power in the theta (3.5–7.5 Hz) band and distorted time domain features. Results supported the use of regression-based and PCA-based ocular artifact correction and suggested a need for further studies examining possible spectral distortion from ICA-based correction procedures.

Introduction

Ocular activity creates significant artifacts in the electroencephalogram (EEG, Fisch, 1991). Epochs contaminated by ocular artifacts can be manually excised, but at the cost of intensive human labor and substantial data loss. Alternatively, correction procedures can distinguish brain electrical activity from ocular potentials, using regression-based or component-based models (see Croft and Barry, 2000, Lins et al., 1993a, Lins et al., 1993b). However, the current literature lacks consensus about optimal correction procedures. This is due, in part, to the inherent challenges of developing and implementing accurate models. In addition, few studies have directly compared different methods for ocular artifact correction, and existing studies have focused almost exclusively on applications to event-related potential (ERP) research Lins et al., 1993a, Lins et al., 1993b. The goal of the current study was to use real and simulated EEG data to compare multiple regression, principal component and independent component methods of ocular artifact correction, with a particular emphasis on implications for spectral analyses of the EEG waveform.

Traditional ocular artifact correction procedures use a regression-based approach Elbert et al., 1985, Gratton et al., 1983. Regression analyses are used to compute propagation factors or transmission coefficients in order to define the amplitude relation between one or more electrooculogram (EOG) channels and each EEG channel. Correction involves subtracting the estimated proportion of the EOG from the EEG. One concern often raised about the regression approach is bidirectional contamination. If ocular potentials can contaminate EEG recordings, then brain electrical activity can also contaminate the EOG recordings. Therefore, subtracting a linear combination of the recorded EOG from the EEG may not only remove ocular artifacts but also interesting cerebral activity. In order to reduce the cerebral activity in the EOG, Lins et al. (1993b) suggested low-pass filtering the EOG signal used to compute regression coefficients. However, they recognized that low-pass filtering removes all high frequency activity from the EOG signal, both of cerebral and ocular origin. In the current paper, we introduce a new filtering approach for regression-based correction using Bayesian adaptive regression splines DiMatteo et al., 2001, Wallstrom et al., 2002. This approach uses a locally defined nonlinear filter to remove high frequency activity when the amplitude fluctuations are small and retain high frequency activity when the amplitude fluctuations are large. Such adaptively filtered EOG essentially isolates activity typically associated with ocular artifacts and removes cerebral activity. The use of such adaptive filtering prior to applying regression correction may substantially reduce problems from bidirectional contamination.

Another class of methods is based on decomposing the EEG and EOG signals into spatial components, identifying artifactual components and reconstructing the EEG without the artifactual components. For example, Lins et al. (1993b) and Lagerlund et al. (1997) used principal component analysis (PCA) to identify the artifactual components. In addition, the dipole modeling technique of Berg and Scherg, 1991a, Berg and Scherg, 1991b, Berg and Scherg, 1994 can use PCA to compute topographies of eye activity. Statistically, PCA decomposes the signals into uncorrelated, but not necessarily independent, components that are spatially orthogonal. More importantly for the purpose of artifact correction, the PCA components may be thought of as formed sequentially to maximize the remaining variance. In particular, the first component is formed to have the largest variance, and therefore easily isolates large amplitude ocular artifacts. Such components can be identified and selected automatically by comparing them with the EOG signal, as we demonstrate in this paper. A newer approach uses independent component analysis (ICA), which was developed in the context of blind source separation problems to form components that are independent Bell and Sejnowski, 1995, Comon, 1994, Jutten and Herault, 1991 ICA has been applied to correct for ocular artifacts, as well as artifacts generated by other sources Jung et al., 1998a, Jung et al., 1998b, Jung et al., 2000, Vigário, 1997. Compared to PCA, ICA removes the constraint of orthogonality and forces components to be approximately independent rather than simply uncorrelated. However, the ICA components lack the important variance maximization property possessed by PCA components. In addition, ICA requires the user to manually select components for correction, thus creating challenges for implementing automated correction routines.

In the current study, we compare the performance of regression-based and component-based correction procedures on real and simulated EEG data of varying epoch lengths. In particular, the correction procedures we consider are two time-domain regression methods (i.e., with and without adaptive filtering), automated and manual PCA, and ICA. Our simulation study allows us to verify the effect of the adaptive filter on correction performance and observe the role of epoch length on the regression methods and automated PCA. In addition, we quantify the effects of each method on the spectral parameters of the EEG. Our aim is to describe the relative strengths and weaknesses of the different methods for effectively removing ocular artifact, without distorting the spectral parameters of the EEG by correction.

Section snippets

Subjects

Twelve adult subjects were selected from subjects participating in a larger study of childhood-onset depression (Miller et al. (2002)). Half of the subjects had a history of childhood-onset depression and half had no history of major psychopathology. Within each diagnostic group, half of the subjects were female. The ages of the subjects range from 21 to 35, with a mean of 28. There were no significant age differences for the diagnostic and gender groups.

Recordings

EEG and EOG was recorded during three

Simulation results

The time and frequency domain errors are summarized in Table 1. The main findings are: adaptive EOG filtering reduces alpha and beta band errors, short epochs reduce frequency domain errors but enlarge time domain errors, and the component-based methods generally yield smaller theta band errors. Overall, the methods and epoch lengths with low error rates are REG-ADAPT with short and long epochs, and PCA-AUTO with short epochs, depending upon the EEG feature of interest.

There is little artifact

Careful adaptive filtering improves correction when using regression in the time domain

We have argued that the use of an appropriate adaptive filter can improve correction when using a time domain regression approach to EEG correction. Adaptive filtering reduced the alpha band error rates by 63% and 69% when using 60- and 1-s epochs, respectively, and reduced the beta band error rates by 76% for both epoch lengths. It is important to emphasize, however, that adaptive filtering cannot completely remove the concern about bidirectional contamination. Low frequency contamination

References (27)

  • I DiMatteo et al.

    Bayesian curve fitting with free-knot splines

    Biometrika

    (2001)
  • B.J Fisch

    Artifacts

  • A Hyvärinen

    Fast and robust fixed-point algorithms for independent component analysis

    IEEE Transactions on Neural Networks

    (1999)
  • Cited by (265)

    • EEG-based emotion recognition: Review of commercial EEG devices and machine learning techniques

      2022, Journal of King Saud University - Computer and Information Sciences
      Citation Excerpt :

      The pre-processing stage deals with removing unwanted artifacts, such as EOG, ECG, EMG interferences, and electromagnetic and frequency types of interferences. Various methods have been presented to eliminate EOG and EMG-related noise (Brunia et al., 1989; Gratton et al., 1983; Wallstrom et al., 2004). The electromagnetic and frequency interferences mainly emerge at high frequencies, so it is an apt approach to use low-pass or band-pass filtering to cut off unwanted frequency bands and only keep EEG-related frequency.

    View all citing articles on Scopus

    Support for the current work was provided by NIMH Program Project MH56193, NSF VIGRE Award DMS-9819950, NIMH Postdoctoral Fellowship MH18951, NIMH Seed Award MH30915 and a NARSAD Young Investigator Award.

    View full text