EOG correction: a new aligned-artifact average solution

Electroencephalography and clinical Neurophysiology 107 (1998) 395–401

EOG correction: a new aligned-artifact average solution Rodney J. Croft*, Robert J. Barry Department of Psychology, University of Wollongong, Wollongong, NSW 2522, Australia Accepted for publication: 15 May 1998

Abstract Objective: In the field of EOG correction, discrepancies have been found between the propagation rates for different types and frequencies of eye movement. However, Croft and Barry demonstrated that these differences can be explained by the affect of EOG magnitude on the correction procedure (Croft, R.J. and Barry, R.J. EOG correction: a new perspective. Electroenceph. clin. Neurophysiol., 1998, 107: 387–394). This study utilized a new ‘aligned-artifact average’ technique (AAA) to examine whether propagation is constant across eye movement types and frequencies, and tested the AAA as an EOG correction tool. Methods: Two experiments manipulated interference levels in real data sets to determine if interference affected propagation coefficients (Bs). The third tested real data for the effect of forward propagation of eye movement related neural potentials on Bs, and the fourth utilized computer simulations to assess the effectiveness of the new AAA correction procedure. Results: Interference was found to inflate B at low EOG amplitude, and its removal removed B variation and inflation. The forward propagation of eye movement related neural potentials had very little effect on B. The AAA procedure produced near perfect corrections of the simulated data, superior to a comparison method. Conclusions: EOG propagation is constant across eye movement types and frequencies, and thus only one correction coefficient should be calculated and applied to data. The AAA method provides a more accurate correction and makes possible, for the first time, the adequate correction of posterior sites.  1998 Elsevier Science Ireland Ltd. All rights reserved Keywords: Aligned-artifact average; Saccade; Blink

1. Introduction Electrical activity generated by eye movements propagates across the scalp and affects the EEG recorded from scalp sites. In order to obtain EEG free of eye movement related artifact, a number of correction procedures have been designed. These procedures remove a portion of the eye movement related potential (measured with the EOG) from the EEG, typically calculated using a regression procedure. Numerous researchers have concluded that eye movement related fields propagate differently for different types of eye movement and frequencies (for a discussion see Croft and Barry, 1998). However, the observed differences have been shown to be consistent with propagation that is constant across eye movement types and frequencies, suggesting the apparent differences are due only to failings of

* Corresponding author. Tel.: +61 2 42213742; fax: +61 2 42214163.

the regression procedure at small EOG/EEG ratios (Croft and Barry, 1998). That study demonstrated that propagation coefficients (Bs) are larger for small EOG amplitudes (corresponding to small EOG/EEG ratios), and that as the EOG amplitude increases, Bs decrease to what is believed to be the true propagation level. At relatively large eye movements it was also found that blinks and saccades produced similar Bs. It remained unclear whether changes in B as a function of EOG magnitude are artifactual, or represent real changes in the rate of propagation of eye movement related fields. The first part of the present study was designed to resolve this question. As the true propagation rate is unknown, we manipulated one hypothesized source of B inflation to ascertain if this affected B. The data from Croft and Barry (1998) were re-analyzed, controlling the effect of coherent interference in the EOG and EEG channels. A reduction in B due to a reduction in coherent interference would demonstrate its artifactual nature. Whereas Gratton et

0013-4694/98/$ - see front matter  1998 Elsevier Science Ireland Ltd. All rights reserved PII: S00 13-4694(98)000 87-X

EEG 97135

396

R.J. Croft, R.J. Barry / Electroencephalography and clinical Neurophysiology 107 (1998) 395–401

al. (1983) subtracted DC levels from each 1200 ms epoch before calculating B from their conjunction, we subtracted DC levels from every 78 ms sub-epoch, which substantially reduces coherent interference from DC shift at small EOGs. This method is designated AM, as opposed to the Grattonstyle method1 (GM) used in Croft and Barry (1998). We expected AM to produce smaller Bs than GM at small EOG ranges. Note that if B inflation at small EOG/EEG ratios is artifactual, it does not follow that large EOG/EEG ratios necessarily produce accurate Bs. The relevant experiments have only examined frontal sites, where the affect of forward propagation and coherent interference is small relative to B, but at posterior sites B is substantially smaller, and thus such an artifact has a larger affect. Although blinks have a large EOG/EEG ratio, this may not be large enough posteriorly to negate the effect of interference. To overcome this posterior sites problem, we can create EOG averages, time-locked to certain eye movements, and calculate B from these averages. We use the phrase ‘aligned-artifact averages’ (AAA) to distinguish them from traditional event-related potentials, since the AAAs are not necessarily time locked to an event: they may be aligned, for example, with a blink maximum. By definition, all that should be left in the AAAs of the two channels are the potentials created by the eye movements themselves, and any neural potentials related to these eye movements (such as described by Berg and Davies, 1988). The impact of coherent interference would thus be eliminated from the computations, and B should approximate true propagation. Because EEG is minimized in this approach, the EOG/EEG ratio is maximized, and should become large enough for accurate B calculation from small EOG at all sites.2 This AAA procedure allows a test of the low EOG/EEG ratios as the cause of B variation (and error) at frontal sites, since removal of EEG should remove B variation (and error). Section 2 of this paper includes a re-analysis of data from Croft and Barry (1998), where blinks and saccades produced similar Bs at moderate EOG magnitudes but with significant B variation within the non-blink artifact types. According to our thesis, this variation should disappear as EEG is removed from the two channels by creating AAAs for each of the eye movement types. For each artifact type, we compared the mean of the individual Bs to the mean of the Bs derived from each subject’s AAAs, and to the B derived from the AAAs of the combined subjects (Section 3). As the number of epochs included in the mean increases so will the ratio of EOG/ 1 This Gratton-style method calculates a correction coefficient via a simple linear regression over a 1 s epoch, with EEG as the dependent variable and EOG as the independent variable. No attempt is made to remove forward propagation or coherent interference from the data. 2 All experiments in this paper utilize the vertical EOG channel only, and the results are therefore only generalizable to one-channel EOG correction. However, the principles discussed and the new correction procedure outlined are generalizable to any number of EOG channels.

EEG, and Bs from each artifact type should approach the same value. It was further hypothesized that the ‘final’ B would be close to the Bs calculated for the individual blinks, where the EOG explained a large percentage of the measured EEG due to the large EOG/EEG ratio. It was noted above that the EOG AAAs would contain only the eye movement related potentials and neural potentials related to these eye movements that propagate forward to the EOG electrodes. It is assumed that forward propagation will be slight and thus not affect B significantly. Section 4 of this paper investigated this assumption by removing a portion of the activity at Fz from the EOG, and recalculating B. The difference between the two estimates of B indicates the proportion of B due to forward propagation. Because we cannot at present determine the true EEG signal or true propagation, Section 5 of this paper uses computer simulations with various EOG/EEG ratios to compare the relative efficiency of the traditional Gratton-style and the AAA correction procedures. A known proportion of simulated EOG was added to a randomly generated EEG series, and then simulated forward propagation and coherent interference were added to both the EOG and EEG series. Contaminated EEG series were then corrected using either the Gratton-style or AAA procedure. Correlations were used to compare the success of the methods. It was hypothesized that because the AAA procedure minimizes EEG interference, it would be relatively unaffected by the EOG/EEG ratio, and would perform consistently better than the Gratton-style procedure.

2. Experiment 1 2.1. Method 2.1.1. Procedure EEG from Fz, relative to linked ears, and the vertical EOG, were recorded from 5 volunteers whilst they performed various eye movement and blinking tasks, with data visually selected so as not to include more than one eye movement type in the epoch. The subjects, data acquisition and procedure are described in Expt. 2 of Croft and Barry (1998). 2.1.2. Data analysis Each epoch collected in Croft and Barry (1998) was extended to include 1600 ms before and after the artifact peak. Each epoch was divided into 78 ms segments (20 data points) and for each channel, the mean of each segment was subtracted from each data point within that segment. For each subject and artifact type, temporally corresponding 78 ms adjusted segments were combined. For example, after subtracting the mean, the first 78 ms segment from each of 20 blinks were combined together (sequentially) within-subject. This produced a total of 400 data point

397

R.J. Croft, R.J. Barry / Electroencephalography and clinical Neurophysiology 107 (1998) 395–401 Table 1

Means and standard deviations for both methods of analysis, for each of the eye movement types, and the general eye movement types Artifact type

Fig. 1. Bs are shown as a function of EOG range, for both the Gratton-style method and the alternate method of analysis.

pairs per regression. Bs were then calculated for the combined segments, for each artifact type and subject, producing 5 Bs per artifact type, for each of the 40 segment-pairs. The ranges for EOG/EEG from each of the segments were also computed. 2.2. Results A scatterplot of B versus EOG range for both this AM and the GM of Croft and Barry (1998) is given in Fig. 1. There is no between-method difference in Bs in the 160.0–728.0 mV eye blink range (t(226) = 0.48, P = 0.628), nor in the 66.0– 159.8 mV saccade range (t(267) = 1.09, P = 0.279). The GM mean in the sub-eye movement range is greater than the AM (t(164.25) = 6.26, P , 0.0001, using a t test for un-equal variances). Means are displayed in Table 1. There was an effect of eye artifact type on B for GM, F(3,383) = 18.89, P , 0.001 but not AM, F(3,597) = 0.255, P , 0.858 (means given in Table 1). Using Levene’s test, B variability was greater in the sub-eye movement range for GM than AM (F = 135.609, P , 0.001), but not in the eye movement range (F = 1.367, P = 0.243).

B means Gratton method

Alternate method

Grouped types Fixations Saccades Blinks

0.324 (0.265) 0.139 (0.064) 0.132 (0.043)

0.185 (0.112) 0.148 (0.060) 0.119 (0.060)

Individual types Cued blinks Spontaneous blinks Looking down Looking up

0.137 0.126 0.164 0.111

0.130 0.119 0.162 0.143

(0.043) (0.042) (0.048) (0.066)

(0.048) (0.048) (0.060) (0.059)

averaged to give another AAA (AAA100; displayed in Fig. 2). Bs were then calculated using standard regression techniques for each artifact type for AAA1, AAA20 and AAA100. The means of the AAA1 Bs and AAA20 Bs were then calculated. 3.2. Results A scatterplot for each artifact type shows that, as the number of epochs in the AAA increases, Bs of different artifact types converge (Fig. 3). One-way ANOVAs with artifact type as independent variable revealed significant differences among AAA1 Bs (F(3,394) = 24.12, P , 0.0001), but not AAA20 Bs (F(3,19) = 0.2651, P = 0.8496). It is not appropriate to test for differences where there is only one B per group, but it can be seen in Fig. 3 that the range of Bs is smaller at AAA100 than at AAA20, suggesting that not enough EEG had been removed at AAA20 for optimal estimation.

4. Experiment 3 4.1. Method and data analysis Data from Expt. 1 were re-analyzed: For each artifact

3. Experiment 2 3.1. Method and data analysis Data from Expt. 1 were re-analyzed: 1-s epochs, designated AAA1, were selected from each 3200 ms epoch, such that artifact peaks occurred at 157 ms (to maximize the movement-to-background ratio). For each subject and type, the 20 epochs were averaged to give an AAA (AAA20), and for each type, the 5 subjects’ AAA20s were

Fig. 2. AAAs for the 4 artifact types are shown. The AAAs consist of the epochs from all subjects combined, aligned at point ‘0’ on the X-axis.

398

R.J. Croft, R.J. Barry / Electroencephalography and clinical Neurophysiology 107 (1998) 395–401 Table 2 Correction coefficients from the AAA100s before (B) and after (adjusted B) the residual neural potentials have been removed

Fig. 3. Mean Bs and standard deviations are shown for individual epochs, and mean epochs (AAA20 and AAA100). Different artifact types are drawn separately, and their convergence can be seen as the number in the AAA increases.

type, AAA100 Bs were used to obtain the corrected EEG (CEEG) equal to the original Fz channel EEG minus a proportion (B) of the EOG. A portion (B) of the resultant CEEG was then removed from the EOG channel to give an estimate of the ‘pure’ EOG.3 Uncorrected Fz was then re-gressed on the ‘pure’ EOG to give a ‘pure’ estimate of B (propagation not influenced by forward propagation of neural potentials) for each eye movement type. 4.2. Results Table 2 compares the ‘pure’ estimates of B with the AAA100 estimates of B for each artifact type. The deviation, expressed as percentages of the AAA100 Bs, was maximal for the ‘up’ saccades (1.4%). In contrast, the range of AAA100 type Bs was 0.007, or 4.9% of the maximum B. This indicates that in the AAA method of correction, the forward propagation of neural potentials to the EOG does not substantially affect B.

5. Experiment 4 5.1. Method 5.1.1. Simulations To simulate true EEG (TEEG), 50 1000 data-point series 3 This method was utilized to attempt to remove the effect of forward propagation. It assumes that the rate of forward propagation will be the same as the rate of backward propagation, and thus that the contaminated EOG is composed of the eye movement related potentials and a portion (B) of the EEG.

Artifact type

B

Adjusted B

Change in B

Cued blink Spontaneous blink Looking down Looking up

0.143 0.136 0.143 0.141

0.143 0.136 0.143 0.143

0.000 0.000 0.000 0.002

were randomly generated (Fig. 4). To simulate true EOG (TEOG), a 1000-data-point series was created to represent 20 50 data-point blinks (Fig. 4). The TEOG series was then multiplied by a factor of either 1, 2.5, 5, or 7.5 (forming series designated EOGa, EOGb, EOGc and EOGd). To simulate forward propagation, each of the 50 TEEG series was multiplied by 0.20 and added to each of the 4 TEOG series. This created 50 measured EOG (MEOG1) series for each of the 4 EOG magnitudes. To simulate backward propagation, each of the 4 TEOG series was multiplied by 0.20 and added to each of the 50 TEEG series. This created 4 measured EEG (MEEG1) series for each of the 50 EEG series. A 1000 data-point DC simulation was designed (Fig. 4) and this was added to both MEEG1 and MEOG1, to give MEEG2 and MEOG2. This resulted in 50 series at each EOG magnitude, for each of the EEG and EOG types. In addition, the DC series was added to each of the TEEG series (designated TEEG + DC). 5.2. Statistical methods 5.2.1. Gratton-style Forward propagation (FP) condition. At each EOG magnitude regressions were performed on the 50 pairs of data series with MEEG1 as dependent variable and MEOG1 as independent variable. The resultant Bs were used to correct the corresponding MEEG1 series (giving CEEG1).4 r2 values were calculated between corresponding CEEG1 and TEEG series. Forward propagation and coherent interference (FP/CI) condition. At each EOG magnitude, regressions were performed on the 50 pairs of data series with MEEG2 as the dependent variable and MEOG2 as the independent variable. The resultant Bs were used to correct the corresponding MEEG2 series (giving CEEG2). r2 values were calculated between corresponding CEEG2 and TEEG + DC series. 5.2.2. AAA correction method FP condition. MEEG1 and MEOG1 series were divided into 20 50 data-point segments, and the mean at each point of the 20 segments was computed. Regressions were then calculated on these averaged series with MEEG1 as depen-

4

The correction takes place as follows: CEEG = MEEG − (B p MEOG).

399

R.J. Croft, R.J. Barry / Electroencephalography and clinical Neurophysiology 107 (1998) 395–401 Table 4

Gratton-style and AAA method r2 values for both the forward propagation and forward propagation plus coherent interference conditions, for each of the EOG/EEG ratios EEG/EOG ratio

Fig. 4. Examples of the simulated data are shown at the EOG/EEG ratio of 1. Displayed are 40 data-point EOG and EEG series, as well as a portion of the DC shift.

dent variable, and MEOG1 as independent variable. The resultant Bs were used to correct the corresponding 1000 data-point MEEG1 series (giving CEEG3). r2 values were calculated between corresponding CEEG3 and TEEG series. FP/CI condition. MEEG2 and MEOG2 data series were divided into 20 50 data-point segments, and the mean at each point of the 20 segments was computed. Regressions were then calculated on these averaged series with MEEG2 as dependent variable, and MEOG2 as independent variable. The resultant Bs were used to correct the corresponding 1000 data-point MEEG2 series (giving CEEG4). r2 values were calculated between corresponding CEEG4 and TEEG + DC series. 5.2.3. General The following tests were performed for both the FP and the FP/CI conditions: directional matched pairs t test compared the Gratton and AAA correction procedure Bs and r2 values for each EOG magnitude. Because 4 comparisons were made in each condition, a Bonferroni adjustment changed the significance level required from 0.05 to 0.0125.

Gratton method

AAA method

r2

SD

r2

SD

0 010 0 005 0 003 0.002

0.999 0.999 0.999 0.999

0.002 0.001 0.001 0.001

0.993 0.994 0.994 0.994

0.003 0.002 0.001 0.001

Forward propagation 1:1 0.970 2.5:1 0.994 5:1 0.998 7.5:1 0 999

Forward propagation/coherent interference 1:1 0.862 0.027 2.5:1 0.969 0.008 5:1 0.987 0.005 7.5:1 0.991 0.003

5.3. Results 5.3.1. FP condition Mean B and r2 values for the Gratton and AAA procedures are given in Tables 3 and 4, respectively. At all EOG magnitudes AAA had larger r2 values than the Gratton method: t(49) . 2.49, P = 0.008 (Fig. 5), and produced Bs closer to the true propagation of 0.20; t(49) . 17.22, P , 0.001 (Fig. 6). 5.3.2. FP/CI condition Mean B and r2 values for the Gratton and AAA methods are shown in Tables 3 and 4. At each EOG magnitude AAA had larger r2 values than the Gratton method; t(49) . 8.64,

Table 3 Gratton-style and AAA method Bs for both the forward propagation and forward propagation plus coherent interference conditions, for each of the EOG/EEG ratios EEG/EOG ratio

Gratton method

AAA method

B

SD

B

SD

0.023 0.010 0.005 0.003

0.207 0.201 0.200 0.200

0.025 0.010 0.005 0.003

0.209 0.202 0.201 0.200

0.025 0.010 0.005 0.003

Forward propagation 1:1 0.334 2.5:1 0.223 5:1 0.206 7.5:1 0.202

Forward propagation/coherent interference 1:1 0.473 0.021 2.5:1 0.254 0.010 5:1 0.214 0.005 7.5:1 0.206 0.003

Fig. 5. Mean r2 values (and standard error bars) for the corrected and original EEG series are displayed as a function of EOG/EEG ratio, for both the Gratton-style procedure and the AAA correction method.

400

R.J. Croft, R.J. Barry / Electroencephalography and clinical Neurophysiology 107 (1998) 395–401

P , 0.001 (see Fig. 5), and produced Bs closer to the true level of 0.20; t(49) . 101.00, P , 0.001 (see Fig. 6). The maximum discrepancy between actual and measured mean propagation levels (as a percentage of the true propagation level) was 136.5% for the Gratton-style procedure, and 4.5% for the AAA correction procedure.

6. General discussion Experiment 1 clearly demonstrated that coherent interference inflates B and causes greater within-subject variability when B is calculated from epochs with little EOG activity. Reduction in this effect caused a reduction in B variation and inflation (Fig. 1). That is, B is artificially inflated at lowpower EOG, and correction procedures based on low-power EOG will produce distorted EEGs. A residual B variation and inflation remained after removal of coherent interference. In this study Bs were reduced from a mean of 0.32 using the Gratton-style procedure, to a mean of 0.18 using the alternate method, but this is still significantly higher than the blink mean of 0.14 hypothesized by Croft and Barry (1998) to be the true propagation. As will be discussed later, the results of Expt. 2 show that residual B inflation was artifactual and not a true change. Thus, regarding residual inflation, since the only artifact removed in Expt. 1 was a portion of DC, the remaining inflation may be due to the forward propagation of EEG to EOG, other sources of coherent interference, or the remaining effect of the DC variation within the 78 ms sub-epochs. It was claimed above that AAA correction provided a means of eliminating almost all sources of contamination, and that it was the only method appropriate for correcting

Fig. 6. Calculated Bs are displayed as a function of EOG/EEG ratio, for both the Gratton-style procedure and the AAA correction method.

posterior sites (where artifactual contamination is similar in magnitude to the EOG propagation itself). Experiment 2 provided validation of this thesis, in that at moderate EOG magnitudes, the AAA method (by removing most of the artifact), removed most of the B variation. For example, ‘up’ and ‘down’ eye movements had significantly different Bs (0.11 and 0.16, respectively) at moderate EOG/EEG ratios, but as we increased the number of epochs in the AAAs (and thus removed much of the contaminating EEG), the values approached equality (0.1409 vs. 0.1428). The initial variation was thus not due to true differences in propagation, and when the mechanism of variation was nullified by the increased EOG/EEG ratio, the variation vanished. It should be noted that B values calculated from the up and down AAAs were similar to the mean of the individual blink Bs. This suggests that individual blink EOG/EEG ratios (mean = 7.3) are large enough to give adequate B estimates for sites close to the eyes. Experiment 2 also showed that residual B inflation was artifactual. This is because AAA correction removed only activity unrelated to the eye movements and so B changes due to this procedure are due to removal of activity unrelated to eye movements, which may thus be identified as artifactual in nature. It is concluded that inflation at low EOG power is artifactual. Reported B differences for blinks and saccades can therefore be explained in terms of lower EOG power typically used to calculate saccade Bs. Blink Bs are invariably calculated from large EOG power, and saccade Bs generally from data that are simply lacking blinks. Saccade Bs may therefore be calculated from data with very little EOG activity, causing B inflation. In a similar fashion, different Bs reported for different EOG frequencies may be explained in terms of different magnitudes of the different frequencies. That is, frequencies of eye movement producing large Bs may correspond to low-power EOG, and frequencies producing small Bs may be associates with high-power EOG. Thus low-power EOG should not be used to calculate EOG correction coefficients, as it will result in inflated and variable Bs. Procedures that calculate coefficients to correct that data set itself often use data that are of low EOG power. Such procedures should be avoided. It was noted earlier that eye movement AAAs contain contamination due to the forward propagation of neural potentials related to these eye movements themselves, time-locked to the eye movements and not removed in the averaging. Experiment 3 found that removal of this contamination affected B very little. B variation between artifact types decreased with removal of EEG from the EOG, but the maximum change was only 1.4% of the original B (for the ‘up’ condition). To translate into real data changes, if we take the ‘up’ saccade case (the maximal B change) and assume a 300 mV eye movement, then the amount of EEG subtracted from Fz would change from 42.3 mV to 42.9 mV when the effect of neural potentials is removed. As the variation in B between artifact types was approximately

R.J. Croft, R.J. Barry / Electroencephalography and clinical Neurophysiology 107 (1998) 395–401

4.9% of the maximum B, this variation due to eye movement related neural potentials (only a quarter of that due to the variation between eye movement types) is not sufficient to concern the correction process. Experiment 4 showed that AAA correction was superior to the Gratton-style procedure, with greater advantage at smaller EOG/EEG ratios. In terms of B accuracy and true-to-corrected EEG concordance, AAA correction performed better at all EOG/EEG ratios. The greatest difference was at the ratio 1:1 (equivalent to fixations or small saccades). At this ratio, 99.29% of the variance of the true EEG was shared by the AAA-corrected EEG, compared with only 86.18% for the Gratton-style corrected EEG. The AAA correction procedure is therefore appropriate to use for posterior sites since B is calculated after the removal of a large portion of the EEG, and is thus based on an increased EOG signal (relative to the EEG). Bs calculated by this method thus will not be affected significantly by coherent interference or forward propagation. We conclude that procedures employed previously have been overcorrecting at posterior sites, and thus have been introducing artifact into the EEG. This new procedure can be implemented either on- or offline. Off-line an EOG correction program could collect epochs time-locked to blink maxima, create AAAs for the EOG and EEG channels, calculate Bs from these, and apply them to the raw data. On-line, a pre-experiment calibration procedure is required, where a number of the same eye movements are performed. Bs calculated from this could then be used to correct the ongoing EEG. The latter procedure assumes that there will be no within-subject B fluctuations during the experimental session itself.

401

It should be noted that AAA correction did not correct perfectly in the simulations. This is because the correction was performed on raw data contaminated by forward propagation and coherent interference. Instead of removing a portion of the eye movement related voltages, it removed a portion of these plus a portion of the forward propagation plus a portion of the coherent interference. This effect is small, however, and the near perfect r2 values in Table 4 show that the AAA correction method is adequate. To conclude, eye movement related potentials propagate across the scalp in a uniform fashion, independently of artifact type. The different propagation coefficients previously reported for saccades, blinks, and different frequencies have been shown to be due to a deficiency of the regression equation that becomes significant only at low EOG power. Coefficients calculated from low EOG power and/or at posterior electrode sites using previous methods are likely to be highly variable and inflated. Their corresponding ‘corrections’ are likely to introduce EOG artifact and should be avoided. A more accurate correction procedure is to create AAAs in both the EOG and EEG channels time-locked to the eye movements, and to correct the raw EEG using Bs calculated from these. References Berg, P. and Davies, M.B. Eyeblink-related potentials. Electroenceph. clin. Neurophysiol., 1988, 69: 1–5. Croft, R.J. and Barry, R.J. EOG correction: a new perspective. Electroenceph. clin. Neurophysiol., 1998, 107: 387–394. Gratton, G., Coles, M.G.H. and Donchin, E. A new method for the off-line removal of ocular artifact. Electroenceph. clin. Neurophysiol., 1983, 55: 468–484.