Nineteen healthy men without any history of neurological and psychiatric illnesses participated in the study (23.18 ± 1.92 years). Only men subjects were recruited to minimize the impact of potentially confounding variable according to gender differences in the cortical processing of visual emotional stimuli, (which is not direct interest of our study).21 All subjects were undergraduate students of Yonsei University and had normal or corrected-to-normal vision.

The visual stimuli were constructed from the International Affective Picture System (IAPS).22 The selection of images was performed based on the valence and arousal levels reported by the participants. We selected 100 images that may induce positive emotion (valence, 6.92 ± 0.56; arousal, 5.52 ± 0.75), 100 images that may induce negative emotion (valence, 2.48 ± 0.45; arousal, 5.66 ± 0.72), and 200 non-emotional pictures (valence, 4.97 ± 0.51; arousal, 3.65 ± 1.04). The IAPS stimuli used in this experiment for each condition are specified in Table 1. The number of neutral stimuli was set to be equal to the sum of positive and negative stimuli, so that the numbers of emotional and neutral stimuli were matched. Valence levels of the two emotional categories (positive and negative) were found to be significantly different (independent t-test, t(198) = 69.78, P < 0.001, Bonferroni corrected), but there was no significant difference in arousal levels of them (independent t-test, t(198) = −1.60, P = 0.330, Bonferroni corrected).

The experiment task was divided into four blocks of 12.11 ± 0.25 minutes. Each block consisted of 100 trials, including 25 positive, 25 negative, and 50 neutral visual stimuli. Five-minute resting periods were given between two blocks. Neutral pictures were always presented between emotional pictures, in order to avoid the effect of emotion induced by preceding stimulus. Fig. 1 shows the experimental task and the temporal structure of a single trial. The pictures were presented for 2,000 ms, which was set as the epoch of interest for the analysis. A fixation mark (‘+’) was presented for 1,000 ms before each stimulus. The subjects were instructed to determine whether they liked or disliked the visual stimulus as soon as possible, but to keep the intention of preference (like or dislike) for another 2,000 ms after the stimuli was extinguished while the fixation mark was displayed. Then, the subjects were to respond whether they ‘liked’ or ‘disliked’ by either ‘○’ or ‘×’ as shown in Fig. 1 as soon as possible within 2,000 ms. The location of ‘○’ and ‘×’ was randomized in order to minimize the prediction of the direction, and motor preparation.

EEGs were recorded at a sampling rate of 500 Hz using a 64 channel Ag/AgCl electrode cap (actiCAP; Brain Products, GmbH, Munich, Germany) according to the extended 10-20 system. The impedances of all electrodes were kept below 10k ohm. The reference and ground electrodes were placed at FCz and AFz, respectively. A bandpass filter (0.03–100 Hz) and 60 Hz notch filter was applied. EEGLAB was used for the preprocessing of brain potential.23 The single-trial waveforms were excluded from further analysis if they were severely contaminated by drift or transient artifacts due to head/electrode motion. Since we had the knowledge on the ground truth of valence of each visual stimulus, we further excluded the trials in which the subject's response was apparently incorrect. Each image in the International Affective Picture System (IAPS) has official normative rating on emotional condition. The rating can be considered to reflect normal emotional response (i.e., ground truth), and thus, each image can be classified into positive/negative/neutral states based on the rating. The subjects were asked to report whether they liked or disliked the stimuli. A trial was excluded from further analysis, if the response was not in accordance with the emotional state which was determined from the rating. For example, when a subject answered that he liked an apparently negative picture, we considered he showed abnormal emotional response due to some reason including abnormal personal preference, personal experience, or inattentiveness to the stimulus. The number of remaining trials was 89.05 ± 13.88, 88.37 ± 7.50, 187.63 ± 14.56 for positive, negative and neutral stimuli, respectively. Stereotyped artifacts such as eye movement or EMG were removed using independent component analysis (ICA).24 Subsequently, the EEGs were re-referenced according to the average reference.

An open source MATLAB toolbox Brainstorm was used to calculate the cortical current source density on cortical surface.25 Boundary elements method was used for the forward problem. International Consortium for Brain Mapping (ICBM) 152 template was used as default anatomy.26 The head model was computed with a symmetric boundary element method (BEM) using OpenMEEG implemented in Brainstorm toolbox.27 The number of vertices in scalp, outer skull, inner skull layer were all set to 1922, and the conductivities of them were set to 1, 0.0125, 1 respectively. The thickness of skull was set to 4 mm. The current density at 15,002 points on cortical surface was calculated by solving the inverse problem with weighted minimum norm estimation.28 The cortical surface was homogeneously parcellated into 500 points using surface clustering implemented in Brainstorm toolbox,29 for the whole-brain analysis without prior knowledge. Finally, current density time-series on 500 cortical points were obtained.

Conventional analysis of event-related potential (ERP) utilizes prior knowledge of electrodes and time which are highly likely to show significant difference between emotional conditions. However, our purpose here is to find out the critical time and cortical region from the ERCD time-series itself, without using prior knowledge. This requires statistical comparisons of cortical current density at many points in time-cortex space, resulting in a multiple comparison problem involving a very large number of simultaneous comparisons. In our case here, there exist 250,000 (= 500 vertices × 500 time points) multiple comparisons. We adopted a nonparametric cluster-based permutation test17 to deal with this problem. As shown in Fig. 2A and B, this method tries to find a set of continuous spatiotemporal points which show significant changes in ERCD according to conditions, which is referred to as a ‘cluster.’17

The detailed procedure is as follows: 1) ERCD averaged over all trials at all 500 cortical points, during −100 to 1,000 ms (550 time samples) are calculated; 2) Baseline correction of the averaged ERCD time-series is performed by subtracting average ERCD within −100 to 0 ms; 3) Statistical comparisons of three valence conditions are performed by one-way ANOVAs at 500 × 500 points in time-cortex space (500 temporal points, excluding 50 points within −100 to 0 ms baseline period); 4) The points with large F values, corresponding to P < 0.05, are selected; 5) The clusters of critical spatiotemporal points are defined by combining the points if two significant points are consecutive in time or within 30 mm in space17; 6) The ‘mass’ of a cluster is defined as the sum of F values within the cluster; 7) Surrogate data are obtained by random permutation of three conditions within each subject (5,000 times), and the null distribution of cluster mass is obtained by repeating steps 1)–6) above for the surrogate data; 8) The cluster masses found in step 6) are compared to the highest 5% of null distribution in step; 7) to determine whether they are significantly large, i.e., whether corresponding point corresponds shows significant difference between emotional conditions.

For the nonparametric cluster-based permutation test described above, the Mass Univariate Toolbox of Groppe et al.30 was modified to enable the ANOVA in addition to the t-test. The temporal epoch for the analysis was set to 400–800 ms, according to the earliest and latest time in the significant clusters determined as described above. The significant clusters were further parcellated to anatomical regions following the automated anatomical labeling (AAL) of Tzourio-Mazoyer et al.31 after taking time-average over the temporal epoch described above. The ERCDs of the vertices belonging to each parcel, i.e., each region of interest (ROI), were averaged to obtain ERCD time series for the parcel. Here the clusters were grouped to 6 ROIs (among 116 in the AAL database31) including bilateral central, cingulate, and inferior temporal areas.

Finally, two-way repeated-measures ANOVA was performed to analyze averaged ERCD amplitude over 400–800 ms of each region of interest. Within-subject factors included emotional conditions (3 levels) and ROIs (6 levels). Post-hoc t-test was performed to compare ERCDs between emotional conditions for each ROI. The P value was corrected by false discovery rate (FDR)32 method, proposed by Benjamini and Yekutieli.33 Statistical results of ROIs which showed significant ERCD differences among the conditions were displayed for simplicity.

This study was approved by the Yonsei University Wonju Institutional Review Board (approval No. 2011-21). Written informed consent was obtained from each subject prior to the experiment.