General Method

Electrophysiological recordings

EEG recordings were made using an EGI 128-channel system. The system consists of geodesic sensor nets available for different head sizes and a high input impedance amplifier. EEG was recorded continuously by means of an EGI high density array with 128 electrodes (Electrical Geodesics, 1998). The vertex (recording site Cz) was chosen as reference. Electrode positions as projected onto a plain are shown in Figure II.1. As suggested for the Electrical Geodesics high input impedance amplifier, impedances were kept below 50 kΩ. All channels were preprocessed on-line by means of 0.1 Hz high- and low-pass filtering, with low-pass cut-off frequency being set to values well below Nyquist frequency. Sampling rates for the experiments reported below were either 250 or 500 Hz.

Data reduction and analysis

For artifact rejection and artifact correction, data were submitted to a procedure developed by Junghöfer and collaborators (Statistical Correction of Artifacts in Dense Array Studies, SCADS, in press). In traditional psychophysiological approaches to artifact correction, epochs containing artifacts or contaminated sensors are completely removed from the data. The SCADS method employed in the studies described below uses a different procedure based on statistical parameters of the data. In a first step, this procedure detects recording channel artifacts using the recording reference (i.e. Cz). Subsequently, global artifacts are detected using the average reference. In a next interactive step, distinct sensors from particular trials are removed on the basis of the distribution of their amplitude, standard deviation and gradient. The information of eliminated electrodes is replaced with a statistically weighted spherical interpolation from the full channel set. In a last step, the variance of the signal across trials is computed to document the stability of the average waveform. The limit for the number of approximated channels was set to 20 channels. With respect to the spatial arrangement of the approximated sensors, it was ensured that the rejected sensors are not located within one region of the scalp, because this would make interpolation for this area invalid. Single epochs with excessive eye-movements and blinks or more than 20 channels containing artifacts were discarded. All further analyses were performed using the average reference.

Figure II.1: Sensor layout of the EGI 128-channel system. Sites roughly corresponding to sites of the international 10-20 system are also shown

2 1

As a measure of global cortical activity and assessment of ERP components and their respective time course, the Global Power g(t) of the ERP was computed according to

g(t)=

xj(t)2 ⋅sj(t)−1 j=¹

128∑

128⋅ ^s_j^(t)−¹

j =1 128∑

(1)

where xj(^t) = voltage at sensor j and time t

and sj(t) = standard deviation of voltage at sensor j and time t across trials

Statistical analysis

When comparing multiple recording sites using linear Analysis of Variance (ANOVA) models, a common problem is the fact that an amplitude difference between two conditions is not associated with an additive increase or decrease at all sites. Rather, the effect of amplitude is multiplicative with respect to the distance from a given source. Consequently it has been proposed to use normalization procedures for the evaluation of interactions including the SITE factor (see below). In the present thesis, ERP and spectral data were normalized as proposed by McCarthy and Wood (1985): For each condition and time window, the minimum and maximum values across electrodes were determined and the normalized value n at each electrode j was computed according to

n

j^{(t) =}

x

^j(t)−min

max −min (2)

Where x_j(t) = Potential at sensor j and time t

For assessment of main effects, ANOVAs on the original parameters were computed. In order to enhance reliability of the sampled voltages for each site, electrodes were grouped into regional means prior to statistical analysis. Respective ANOVA structures and channel groups are

given in the experiment sections of this thesis.

Spectral analysis

Two different methods have been used in order to extract the induced GBA from the EEG epochs. In experiments A and B, the induced GBA was obtained using a Fast Fourier Transform (FFT) approach proposed by Feige, which is included in the avg_q software (Feige, 1996). This approach computes the power spectrum in a manner similar to that described by Makeig (1993).

In addition, it enables the user to employ a modular approach to spectral analysis. In the present work, the time-locked activity (i.e., the ERP) was subtracted from the signal in a first step.

Subsequently, the epochs were demeaned and detrended. A Welch-tapered analysis window was shifted across the recording epoch, thus determining the time resolution of the spectral estimate.

The power spectrum for each of these analysis windows was determined using two FFT windows with a user-determined overlap that can be adapted to the requirements of the experiment. The resulting power spectra were averaged across epochs. In general, one major drawback of FFT approaches to time-frequency analyses is due to the Heisenberg uncertainty principle: An increase in time resolution t is associated with reduced frequency resolution f and vice versa.

More precisely, t f = 1/4ð. In addition, FFT windows that allow for an assessment of effects in the EEG alpha range may be too long to detect fast and short oscillations in the higher ranges. In experiments C and E, complex Morlet wavelets were used for frequency analysis to overcome the above-mentioned problems with constant FFT window length. This procedure has the advantage that the time resolution for high frequencies is better compared to low frequency ranges, where frequency resolution is better, but time resolution is coarse. Thus, this technique is especially suited for detecting induced high-frequency oscillations that may occur during brief periods of time.

The present procedure has been proposed by Bertrand and coworkers (1994) and is described in detail in the respective publications (e.g. Tallon-Baudry et al., 1997, 1998). In brief, complex Morlet wavelets g can be generated in the time domain for different analysis frequencies f0 according to

g( t, f

₀

) = ^{A' e}

⁻

t² 2σt2

e

^{2iπf0t (3)}

with A’ depending on the parameter σ f , specifying the width of the wavelet in the frequency domain, the analysis frequency f0 and the user-selected ratio m:

with

Thus, given a constant ratio m, the width of the wavelets in the frequency domain, σ f , changes as a function of the analysis frequency f0. In the present experiment, we used a ratio of m

= 7, because this ratio provides adequate time resolution and frequency accuracy in a time range between 10 and 100 Hz, while introducing no distortion due to differences in the integral size of wavelets at different frequencies in the frequency domain. Both the signal and the wavelets, computed using equation (3), were transformed to the frequency domain and multiplied separately for different analysis frequencies. After retransformation, the energy of the signal at different frequencies was obtained as the complex modulus (magnitude) of the real and imaginary part of the result matrix. Information on phase-locking of oscillatory responses can be obtained by averaging the normalized complex result of the convolution of signal and wavelet across single epochs. The modulus of this complex value that can lie between 0 and 1 can be statistically tested using the circular Rayleigh test (see Tallon-Baudry et al. 1997).

Source space analysis

Given sufficient spatial sampling associated with a large number of electrodes, it is possible to use source localization techniques for estimating the origins of the scalp-recorded signal. In the present thesis, we used a distributed source modeling approach: cortical sources of ERPs were estimated using the Minimum norm estimate (MNE), first proposed by Hämäläinen and Ilmoniemi (1984). The MNE is an inverse method for reconstructing the primary current that underlies an extracranially recorded time-locked brain potential. The procedure is based on the

A

'

= σ

_f ²

π

³

^m

f

₀

π

m

=

^f0

σ

f

(4)

(5)

assumption that the data vector d, which contains the recorded scalp potential at given electrode sites can be described as the product of the leadfield matrix L, specifying the electrode sensitivity to the sources and the source current vector j (see e.g. Grave de Peralta Menendez et al. 1997), plus a noise component ε.

Since L and d are known, and ε is treated as if to be estimated with an acceptable accuracy, the MNE for j can be determined as the mathematically unique solution of this equation which minimizes the squared current density (j²=min)². This solution is obtained by multiplying the pseudoinverse of the leadfield matrix L with the data. Given the high number of electrodes and the presence of noise, spatial regularization is performed with the factor λ (termed

‘regularization lambda’ in this thesis). To this end, Tikhonov-Phillips regularization for matrices is applied during pseudo-inversion of the leadfield matrix L, i.e., L + λ I is pseudo-inverted (Hauk et al., 1998), with I = identity matrix. A main property of this algorithm is the fact that sources that do not contribute to the measured scalp potential are omitted. Consequently, a priori information about the number or locations of cortical sources is not required for this analysis. For computation of the MNE, hypothetical dipoles having three orthogonal orientations at 1384 evenly distributed locations were assumed. Thus, a total of 4152 dipole components were used for the leadfield matrix and the resulting source vector.

A further important issue is the role of the depth of the source for scalp distribution and accuracy of inverse solutions. The MNE algorithm used here addresses this question in the following manner: A three-dimensional source space consisting of three concentric shells (80%, 60%, 40% of electrode radius) was computed as a rough approximation of the brain volume.

However, estimates of the MN method for deep sources are not independent from superficial ones, which is due to the general constraints imposed by the so-called inverse problem of bioelectric processes. Nevertheless, the solutions for the different shells can be considered as indices of the true current source differing with respect to their degree of blurring and depth sensitivity. Accordingly, Hauk and coworkers have shown for the biomagnetic case that deeper shells are associated with less suppressions of deep sources, but more blurring occurs (Hauk et al., 1998). In the present work, we report the solutions for the shell at 60% electrode radius

2 Alternative possibilities include minimizing the absolute value of j , or |j| (L’-norm), among others.

d = ^Lj + ε

⁽⁶⁾

throughout, as this was regarded a compromise between blurring and depth sensitivity. 129 locations on this shell were selected to correspond to the recording sites of the EGI system, and the respective MNE amplitudes at these sites were computed for illustration and further analysis.

III. Experiments

Experiment A: Gamma-band activity and event-related potentials during perceptual shifts of an ambiguous sad/happy face figure3

Introduction

Vision involves the perception of organized wholes in addition to the perception of individual object attributes. The process of forming a meaningful percept includes integration of object properties and characteristics. A traditional approach used to study visual perception exploits the properties of ambiguous figures (Attneave, 1971; Kanizsa and Luccio, 1995). In the present study, we used an ambiguous figure that when rotated biased the participants’

perception to that of either a sad or happy face. The advantage of this procedure is that while the visual input changes continuously, the visual system alternates between two distinct visual experiences, that of either a happy or sad face. In addition, processes of emotional perception can be assessed without using different stimulus material representing affective categories.

Given this stimulus, we hypothesized that GBA would be associated with the perception of the sad and happy faces rather than with simply viewing the continuously moving stimulus.

Furthermore, it was expected that differences of GBA amplitude or topography should be associated with switches from sad to happy facial expressions, compared to the switch from happy to sad expressions.

Materials and Methods Participants

11 right-handed undergraduate university students (5 women, 6 men; age range from 23-29, mean age 25) with normal or corrected vision consented to participate. They received class credits or a small financial bonus for participating.

Experimental Design and Stimuli

The ambiguous stimuli used were a schematic face drawing that can be perceived as either sad or happy depending on orientation (Fig. III.A.1, B), and a modified version of the Rubin vase, which can be perceived as either two faces or a vase (Fig. III.A.1, A). The

3 This part of the thesis largely corresponds to the paper of Keil A , Müller, MM, Ray, WJ, Gruber, T & Elbert T (1999): Human gamma-band activity and perception of a gestalt. J Neuroscience 19, 7152-7161

experimental design comprised three conditions: Condition 1 - non-ambiguous versions of the stimuli were presented in a static form such that perceptual shifts were not possible, i.e., the two faces in the Rubin illusion alone (Fig. III.A.1, C), the vase alone (D), the sad face alone (E) or the happy face alone (F). Condition 2 - perceptual shifts could occur independently of stimulus orientation as would be the case when the Rubin vase is rotated; Condition 3 - perceptual shifts were elicited by the rotating happy/sad face.

Figure III.A.1: ambiguous and control figures used in the present study

The order of these conditions was counterbalanced across individuals. All figures were drawn in black ink on square paper cards of 17.4 cm length, shown at 200 cm distance from the viewer. The figures formed a visual angle of 5° both horizontally and vertically. Illumination was held constant at 20 cd/m². To reduce exploratory eye movements, a fixation point was marked in the middle of each figure. The stimuli were mounted on a rotation device which was hidden from the subject by a black background. Thus, there was nothing in the subject’s field of view but a black/white stimulus in front of a black background. In the rotation trials, stimuli were rotated in a clockwise direction with a rotation speed of 12 revolutions/minute with one revolution lasting 5 seconds. A signal was sent to the EEG trigger channel after each full rotation of the stimuli at the vertical orientation of the happy face and the Rubin vase.

The starting orientation of the rotating figures was randomized across participants.

Procedure

Subjects completed a short form of the Edinburgh Handedness Questionnaire (Oldfield, 1971) and signed an informed consent form. At the beginning of the experimental session, the figure reversal phenomenon was demonstrated using the ambiguous figures. After application of the electrode array, subjects were seated in a comfortable chair and were asked to look steadily at the fixation point. During EEG recordings, each non-ambiguous figure (see Figure III.A.1) was presented twice without rotation for 60 seconds each (condition 1). Then, both the Rubin vase (condition 2) and the sad/happy figure (condition 3) were presented twice for 2.5 minutes in continuous rotation. In a separate block, the rotating stimuli were shown again for 2.5 minutes each and the participants were asked to press a response key each time they experienced a change of percepts. No EEG recordings were made during this task to avoid contamination with movement artifacts or movement related potentials. The order of conditions and stimuli was pseudo-randomized. After each stimulus presentation, i.e. after a 1 or 2.5 minute viewing period, respectively, subjects looked at a white board in the recording chamber and reported if movement aftereffects (i.e. the waterfall effect – objects in the field of view seem to move in the opposite direction of the observed moving stimuli) or afterimages were present. To ensure that these effects did not interfere with processing of subsequent stimuli, presentation of consecutive figures was delayed until the participants reported absence of any aftereffects.

Electrophysiological recordings

EEG was recorded from 128 electrodes using an Electrical Geodesics system, as described in section II. This electrode montage includes sensors for the recording of vertical and horizontal EOG. Data were digitized at 250 Hz using Cz as reference. Impedances were kept below 50 kΩ. All channels were preprocessed on-line using 0.1 Hz high-pass and 100 Hz low pass filtering. Epochs of 5000 ms length were obtained, thus containing one complete revolution of the rotating stimuli. Further data processing was performed off-line by the procedure proposed by Junghöfer and coworkers (Junghöfer et al., in press). Using this procedure, 32% of epochs were rejected, resulting in an average of 41 epochs in conditions 2 and 3, respectively. Average reference data were used for further analyses.

Figure III.A.2: Layout of the electrode array. Electrodes in the shaded clusters, corresponding with sites of the international 10-20 system, were grouped for statistical analysis (see text). Frontal electrodes are shown at the top of the figure.

Data analysis ERP

A 30 Hz low-pass filter was applied to the data prior to all ERP analyses. ERPs associated with switching between the sad and happy face orientation of the rotating face stimulus were examined in time segments time-locked to horizontal orientation of the sad/happy stimulus. The averaged event-related potential in the 350 ms time window following attainment of a horizontal orientation was analyzed. The 150 ms segment following the trigger signal indicating vertical stimulus orientation after completion of a full rotation was selected as the baseline for the ERP analyses. No qualitative changes of the percept were expected to occur in this time range. In addition, the signal in this time window appeared to be stable across subjects upon visual inspection. However, the observation of continuously rotating stimuli may be associated with a steady-state response at the frequency of the rotating

figures. Therefore, an additional ERP analysis was computed using as baseline the mean voltage of the averaged ERP across the entire time range. In each case, the selected baseline mean was subtracted from the ERP data. For the purpose of statistical analysis, the values from 128 electrodes were summed into twelve regional means based on recording sites of the international 10-20 system (Fp1, Fp2, F3, F4, T3, T4, C3, C4, P3, P4, O1, O2; see Fig.

III.A.2). These were organized into two within-subject factors in the analyses of variance (ANOVA): HEMISPHERE (left vs. right) and SITE (anterior-inferior, anterior-superior, medial-inferior, medial-superior, posterior-inferior, posterior-superior). Accordingly, ERP differences between the two rotating figures, the two time windows (350 ms following 90°

and 270° orientations) and the 12 spatial means were evaluated by means of a univariate ANOVA with the within-subject factors being FIGURE (Rubin vase vs. sad/happy figure), HEMISPHERE (left vs. right), SITE and ORIENTATION (90° (happy to sad) vs. 270° (sad to happy)). For the evaluation of interactions including the SITE factor, the ERP was normalized as proposed by McCarthy and Wood (1985). The rotating Rubin vase data were included in these analyses as a control measure. As expected, there was no visible event-related potential in the Rubin vase condition. Therefore, the potential obtained in this condition was assumed to reflect baseline noise and brain activity during the continuous observation of a rotating object.

Data analysis gamma band power (GBP)

Spectral analysis of the EEG data was performed using a Fast Fourier Transform (FFT) algorithm included in the avg_q analysis software developed by Feige (1996). A Welch-tapered analysis window of 384 ms (96 sample points) was shifted in steps of 42 sample points across the recording epoch, thus providing a time resolution of 168 ms. The power spectrum for each analysis window was determined using two FFT windows containing 64 sample points, with an overlap of 32 data points. The resulting power spectra with a frequency resolution of 3.9 Hz were averaged across epochs. For the rotation conditions, time information on spectral power was further collapsed into four time windows centered around orthogonal orientations of the rotating stimuli at 0 (happy face up), 90 (switch happy-sad), 180 (sad face up) and 270 degrees of rotation (switch sad-happy). These orthogonal time windows contained information from four overlapping 384 ms FFT windows, i.e. a total of 720 ms. For non-rotating figures, spectra were averaged across recording epochs and all FFT windows before further analysis. Gamma band power (GBP) was examined in

two frequency ranges, 29-45 Hz and 55-71 Hz, to avoid the possibility of 50 Hz electrical interference.

Following the suggestion of Pulvermüller and coworkers (1997), we also analyzed an additional frequency range (72-97 Hz) where electromyographic (EMG) power usually peaks for facial and head muscles (Cacioppo et al., 1990). Although effects in this frequency range might possibly be due to neuronal activity (Eckhorn et al., 1993; Kreiter and Singer, 1996), an absence of effects in this high band would indicate that effects in lower bands are unlikely to

Following the suggestion of Pulvermüller and coworkers (1997), we also analyzed an additional frequency range (72-97 Hz) where electromyographic (EMG) power usually peaks for facial and head muscles (Cacioppo et al., 1990). Although effects in this frequency range might possibly be due to neuronal activity (Eckhorn et al., 1993; Kreiter and Singer, 1996), an absence of effects in this high band would indicate that effects in lower bands are unlikely to

Im Dokument Oscillatory Brain Activity in Human Affective Stimulus Processing (Seite 38-71)