Developmental Differences in Neuromagnetic Cortical Activation and Phase Synchrony Elicited by Scenes with Faces during Movie Watching

Participants

MEG data were recorded from 22 adults (6 females, 3 left handed; mean age, 27.7 years; SD, 5.28 years) and 20 children (7 females, 1 left-handed, mean age, 9.5 years; SD, 1.52 years). A detailed description of the cohort is provided in Table 1 (the details regarding statistical comparison results provided in the table are described in the statistical analysis section below). None of the participants reported any psychiatric or neurologic conditions. All participants had normal or corrected-to-normal vision and reported no hearing problems. Informed consent was obtained from all participants. Additionally, informed consent was obtained by parents or legal guardians of child participants. This study was approved by the Research Ethics Board of Simon Fraser University.

MEG data collection

MEG data were collected using a 275-channel MEG system (CTF MEG Neuro Innovations) in an AK3B magnetically shielded room (Vacuumschmelze) at a sampling frequency of 1200 Hz. During data collection, participants viewed a short movie clip projected onto an MEG-compatible screen 40 cm above their eyes while lying in the supine position. The movie audio was presented via MEG-compatible earphones, which were inserted before the data collection, and a short test was conducted to ensure that each participant could hear sound clearly in both ears. Each participant viewed the movie clip twice with a short break between the presentations. All participants were instructed to pay attention to both movie presentations and to try to keep as still as possible. Head position in the dewar was recorded using three fiducial coils placed at the nasion and left and right preauricular locations. To enable coregistration between the MEG and structural MR image coordinate systems, each participant’s head shape and fiducial coil locations were digitized with a FASTRAK digitizer (Polhemus) before the MEG data collection.

MRI data collection

Most of the participants (22 adults and 18 children) attended the second session of data collection when a T1-weighted anatomic image was collected using a Philips 3T Ingenia CX scanner with the following settings: TE = 3.7 ms; flip angle = 8°; FOV = 256 × 242; matrix size = 256 × 242; slice thickness = 1 mm; number of slices = 213; sagittal orientation. Two child participants, however, did not return for the second session of data collection; thus, the T1-weighted image was not available for them. For these two participants, we selected the best matched MRI from the pool of the other 18 participants based on the mean distance between each Polhemus point and the closest point on the skull surface derived from the segmented T1 image using the Fieldtrip toolbox (Oostenveld et al., 2011). The same procedure was previously successfully applied in the children cohort of similar age (Kozhemiako et al., 2019, 2020).

Stimulus paradigm

The movie clip presented during MEG recording was 4 min and 32 s long and consisted of scenes from the movie “Charlie and the Chocolate Factory” (2005) directed by Tim Burton. Here we used “movie frame” as a term to describe a single image in a movie 24 frames/s) and “movie scene” to denote a movie segment (minimum of 15 movie frames long but usually lasting several seconds) composed of some number of similar frames. The movie clip was manually divided into scenes for the purpose of further scene classification. In addition, to ensure the precision of time-locked analysis, the first five frames of each scene were marked by a white circle that would appear on the black background (Fig. 1B). A photodiode was fixed at the location on the screen where the white circle was projected during movie presentation to detect changes in luminance. The photodiode signal was used to identify the precise time when each scene appeared on the screen. Each movie scene was manually classified into different categories which included the scenes containing primarily faces (51 occurrences). For the time-locked analysis, we combined the scenes from both movie presentations and used the time when each scene with a face appeared on the screen to define the epochs (102 1-s-long epochs in total). For PseudoZ calculation, other scenes picturing hand manipulations, distant human bodies, scenery, and nonliving objects (78 in total) were used to define epochs based on which the noise covariance was computed (described in detail in the PseudoZ subsection below). In addition, we conducted a control analysis to rule out the possibility that differences between groups were produced by some artifact shared between different conditions (e.g., signal leakage because of volume conduction). We used scenes with hands manipulations (58 in total) to define epochs for the control analysis.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Schematic illustration of study design and analysis workflow. A, Two movie presentations (4 min, 32 s each) composed of scenes from “Charlie and the Chocolate Factory” (2005) were presented to children and adult participants with a short break in between. During movie watching, MEG data were collected for each participant. We also recorded a photodiode signal registering the change in luminance at each new scene because of the white dot on the black background placed on the first five frames of each new scene. B, Scenes containing faces (51 faces per movie clip; total, 102 faces) were extracted, and the photodiode signal was used to define the 1000 ms epochs (prestimulus interval, 500 ms; poststimulus interval, 500 ms) with time at zero indicating the onset of face scenes. Subsequently, epochs were used to compute GFPm separately for children and adults. GFPm peaks were later used to define the time windows of interest. C, For each of the time windows of interest, we computed PseudoZ to assess cortical activation elicited by face scenes and intertrial phase synchronization to evaluate cortical connectivity time locked to face scene onset. D, Resulting measures were tested for significant differences between children and adult participants.

Data preprocessing

To control for the effects of excessive head movement, time points with movement exceeding a 10 mm threshold (in any direction for any of the three fiducial coils) were detected, and the epochs that overlapped in time with excessive movement were excluded from the analysis. For the included epochs, the median of head movement was computed for each participant to investigate potential group differences. Next, the data were visually inspected to identify and exclude bad channels. A bandpass filter from 1 to 150 Hz was applied to the data, and line noise was removed by applying a notch filter at the power line frequency of 60 Hz and harmonics. Independent component analysis was applied, and components composed of oculographic (e.g., blinks, saccades) and cardiac artifacts were removed after visual inspection. One second epochs were later extracted based on the stimulus onset timing spanning a 0.5 s prestimulus interval and a 0.5 s poststimulus interval.

Sensor level time-locked responses

We investigated sensor level time-locked responses to scenes with faces to define the time windows of interest for whole-brain activation and phase synchronization analysis. The Fieldtrip toolbox was used to low-pass filter at 40 Hz and downsample the epochs to 300 Hz to reduce computational load for the statistical analysis (described below; Oostenveld et al., 2011). Then the epochs time locked to the stimulus onset were averaged in children and adults. A magnetic counterpart of global field power, global field power magnetic (GFPm) was computed as a root mean square across all MEG sensors and baseline corrected using the time interval from −200 ms to stimulus onset.

Source modeling

The T1-weighted MR image of each participant was realigned based on the coordinates of fiducial coils and segmented to extract cortical, skull, and scalp surface using the Fieldtrip toolbox. A single-shell head model was generated based on each participant’s inner skull surface. The geometry of cortical surface for each participant was reconstructed using Freesurfer, resulting in a cortical mesh of 4002 vertices per hemisphere on the standard space (Fischl et al., 1999). The forward solution was generated for each vertex from the cortical mesh using the Fieldtrip toolbox.

PseudoZ calculation

As a proxy to brain activation, we computed PseudoZ for each vertex on the cortical mesh using a scalar beamformer for sources oriented orthogonal to the cortical surface (but see Nunes et al., 2020). Essentially, PseudoZ reflects the ratio of event-related source power at a given time to the source power at time of no interest. In our case, the time of interest was defined as 0–500 ms of all epochs when faces were displayed on the foreground (102 in total). Noise covariance was computed based on other epochs, including distant bodies, scenery, and nonliving objects (78 in total). Such an approach essentially contrasts the activation pattern elicited during the scenes with faces to the activation elicited during other scenes with no faces on the screen. Alternatively stated, PseudoZ values will be high only if the activation was specific to scenes with faces and was not present during other scenes. To estimate dynamic changes in evoked activation, PseudoZ was computed for four time windows of interest delineated based on visual inspection of GFPm peaks observed in both groups (Fig. 2) and previous literature reporting the most common components elicited by face stimuli (Puce et al., 2013). This resulted in the following time windows of interest: 50–120 ms approximately corresponding to the M100 component; 120–190 ms approximately corresponding to the M170 component; 190–330 ms approximately corresponding to the M250 component; and 330–470 ms approximately corresponding to the M400 component, consistent with previously reported ERFs elicited by face stimuli.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Difference in the time-locked responses to face scenes during movie watching between child and adult groups. The PLS analysis of GFPm across time points from 50 to 500 ms indicated significant group differences (p = 0.018). GFPm in adults (in green) and children (in gray) were averaged across all MEG channels and face epochs. The shadow represents the SE, and star markers illustrate time points where adults had higher GFPm values compared with children based on PLS z score threshold (z score, >2), whereas circle markers represent the opposite (z score, < −2). The rectangular shaded areas illustrate the time windows chosen for further analysis: 50−120, 120−190, 190−330, and 330−470 ms.

Source reconstruction and intertrial phase synchronization

Using a scalar beamformer, we computed spatial filters to reconstruct the time series of each vertex (see Nunes et al., 2020). For the purpose of dimensionality reduction, we used the regions of interest (ROIs) defined by the MultiModal Parcellation Atlas (MPA) to compute intertrial phase synchronization (Glasser et al., 2016). In total, MPA outlines 362 cortical areas. In our analysis, we included 360 cortical areas excluding a medial wall ROI in both hemispheres. Note that although the number of ROIs exceeds the number of channels, in our analysis we do not set an assumption that the signals in ROIs are linearly independent. Glasser et al. (2016) also provided the list of 22 groups encompassing ROIs with similar location, anatomic, and functional properties. We use these groups to label the ROIs (see Figs. 4, 5). Those include the following: primary visual cortex (PVC); early visual cortex (EVC); dorsal stream visual cortex (DSVC); ventral stream visual cortex (VSVC); MT+ complex and neighboring visual areas (MT); somatosensory and motor cortex (SMS MOTOR); paracentral lobular and mid-cingulate cortex (PCL MCC); premotor cortex (PMC); posterior opercular cortex (POC); early auditory cortex (EAC); auditory association cortex (AAC); insular and frontal opercular cortex (Ins FOC); medial temporal cortex (MTC); lateral temporal cortex (LTC); temporo-parieto-occipital junction (TPOJ); superior parietal cortex (SPC); inferior parietal cortex (IPC); posterior cingulate cortex (PCC); anterior cingulate and medial prefrontal cortex (ACC MPC); orbital and polar frontal cortex or PFC; inferior frontal cortex (IFC); and dorsolateral prefrontal cortex (DLPFC).

Before sample phase-locking value (sPLV) calculation, each ROI time course was estimated by averaging the time courses of all vertices within the ROI boundaries accounting for arbitrary polarity as described in the study by Nunes et al. (2020). Then spectral decomposition using a continuous wavelet transform centered at 15 frequency bins spaced on a logarithm scale from 4 to 40 Hz (we decided to avoid high gamma frequency because of its high sensitivity to muscle artifacts) was performed, and sPLV between the signals of ROI pairs was computed for each sample from –200 to 500 ms as follows: sPLVi,j,t,f=1N∑k=1NeΔϕx,y(t,f), where N is the number of epochs, and Δϕ is the phase difference between x and y at time t and frequency f. As in the PseudoZ analysis, sPLV was then averaged within four time windows (50–120, 120–190, 190–330, and 330–470 ms) and baseline corrected (baseline interval from −200 to stimulus onset) to remove non-task-related spurious synchronization.

Statistical analysis

Mean-centered partial least squares (PLS) analysis (McIntosh and Lobaugh, 2004; Krishnan et al., 2011) was selected to perform statistical testing of group differences between children and adult participants because of its numerous advantages. First, it is a multivariate approach that considers multiple features at once when testing group differences and produces a single p-value providing a method for addressing multiple comparisons in a manner well suited for neuroimaging studies. Second, it is very tolerant of high collinearity in data that is characteristic of neuroimaging. Third, PLS uses permutation and bootstrapping techniques to test the significance and robustness of the results and, thus, can be used even if the data do not meet the assumption of normality. Fourth, there is no need to specify a predefined group contrast when using the mean-centered PLS as it generates a data-driven group contrast fully based on existing data. In our analysis, we used PLS toolbox (version 6.15.1) implemented in MATLAB (https://www.rotman-baycrest.on.ca/index.php?section=84).

PLS analysis uses singular value decomposition (SVD) to extract latent variables (LVs), which express the most variance in the data. For mean-centered PLS, data are organized in a 2D matrix (number of participants × features) for each group (children and adult participants). Subsequently, the data matrices are mean centered and averaged across the participants of each group, resulting in two vectors. They are then transposed and combined into a 2D matrix (features × number of groups) that is subjected to SVD. That is why the number of LVs is always equal to the number of groups for mean-centered PLS. Each LV is composed of the following three elements. (1) The first is a singular vector that represents a data-driven group contrast. (2) The second is a singular value that indicates the total variance of the data expressed by this LV. The latter is used to test the LV significance using a global permutation test, as follows. First, a null distribution is created by randomly permuting participants between the group data matrices, mean-centering and averaging the permuted group data matrices. Then SVD is performed to extract LVs based on permuted data, and their singular values are saved. After repeating the above steps, n number of times (in our case, we used 3000 permutations for each PLS analysis), the p-value for the original un-permuted data is computed based on how many randomly generated singular values were larger than the original singular value divided by the number of permutations. In other words, we compute the probability of observing an LV based on a randomly permuted sample that captures the same or more variance as the LV of the original sample. (3) The third is that a singular vector is produced that expresses SVD loadings for each data feature (e.g., time points for GFPm, vertices for PseudoZ, and pairs of ROIs for PLV). Series of bootstraps were used later to generate bootstrap ratios for each data feature that demonstrate the contribution of each data feature to the group contrast and could be interpreted as a z score. Specifically, we used 3000 bootstrapping rounds for each PLS analysis. In the Results section, we report spatial distribution of z scores on the brain surface (Fig. 3) or as a connectivity matrix (Figs. 4B, 5B). The z score threshold of −2 and +2 approximately corresponds to the 95% confidence interval (McIntosh and Lobaugh, 2004).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Spatial distribution of brain activation induced by faces during movie watching in adults and children. PseudoZ was averaged across four time windows (50−120, 120−190, 190−330, and 330−470 ms), which were defined based on the time-locked responses illustrated in Figure 1A (please note that this is a qualitative illustration, and PseudoZ values are not statistically thresholded).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Differences in PseudoZ activation elicited by faces during movie watching in adults versus children. A, PLS z score distribution during the 120−190 ms time window. B, PLS z score distribution during 190−330 ms time window. C, PLS z score distribution during the 330−470 ms time window. Areas in red correspond to the positive z score, meaning higher activation in adults compared with children. Areas in blue correspond to negative z score and represent higher activation in children compared with adults. Padj – PLS p-value adjusted for multiple comparisons.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

Adults display higher intertrial synchrony compared with children while viewing faces during the 120−190 ms time window. A, Number of connections with the highest (>2, in red) or lowest (less than −2, in blue) z score across frequency bins. B, The z score distribution across connections for frequency bin with the highest number of connections with z score above the selected threshold (6.7 Hz). C, D, Connectivity matrix for adults (C) and for children (D) at the same frequency bin.

In total, nine PLS tests were run for the face-processing analysis. We used a single PLS analysis to compare the GFPm at each time point of the poststimulus time interval between 50 and 500 ms between children and adults. We used a separate PLS analysis for each time window (i.e., four altogether) to compare PseudoZ between groups at all vertices. Similarly, four PLS analyses were performed testing connectivity differences between groups during four time windows. The same number of PLS tests were run for the control analysis. Although we had two LVs for each PLS test as a result of two group comparison, in our analyses we always focused on the first LV because it explained the largest proportion of variance. To account for multiple PLS comparisons, the p-values for PseudoZ activations and connectivity comparisons were adjusted for multiple comparisons using the Bonferroni method.

To test whether male/female or left-handed/right-handed participant proportions were different between children and adult samples, we used a χ² test. To test whether there were differences in age, head size, head motion, and the number of included epochs, we used a two-tailed t test. The results are presented in Table 1. We also used a two-tailed t-test to perform a univariate analysis of GFPm across all channels averaged within 50-120 ms and 120-190 ms to target M100 and M170, respectively.