Article Figures & Data
Figures
- Figure 1.
Analysis of event segmentation, self-report, and recall. A, Upper panel, The starting point of each phase as identified by the independent participants after participating in a workshop, including the span of each phase and the phase’s names. The y-axis of this panel indicates the expected engagement evoked by the narrative plot. A, Middle panel, The raters’ z-normalized engagement ratings. The hot colors denote higher scores, and the cold colors indicate lower scores. A, Bottom panel, The z-transformed group-averaged engagement ratings. The dashed vertical lines represent the start of the new phase and the end of the previous phase. B, The participants’ recall similarity for each separate phase. In this case, the falling action and denouement phases were combined. The matrices show the similarities between the subjects, with the hot colors indicating higher similarity scores. The horizontal bars show the SDs of the subjects’ recall similarities. The stars indicate that there was a significant difference (p < 0.001) between the SDs of the last phase as compared with other phases.
- Figure 2.
The procedure for calculating the dynamic intersubject correlation (dISC) and the results thereof. A, The steps in preprocessing the EEG data. Each participant’s data went through a preprocessing stage in which the raw data were de-noised. After this, the volume conduction effect was corrected by implementing the current source density approach and then downsampling and z-transforming. B, The steps for calculating the dISC for three participants (from a total of 32) and one channel (out of all the channels). The defined window (indicated by the gray) was applied to every channel for all participants. After this, we calculated the correlation between the participants’ corresponding channels for all participant pairs. Thus, this figure demonstrates the correlations for all three participant pairs as an example. The correlation between each window pair is one sample point of pairwise correlations (shown by the purple, blue, and brown arrows at the top of panel). Sliding the window throughout the entire signal, we obtained a correlation signal for the three participant pairs. The average for these correlation signals was then used in the next step. C, The procedure for the statistical analysis that was conducted using a permutation test. First, we calculated the actual correlation between the averaged dISC and group-averaged engagement rates. Then, we phase-randomized the group-averaged correlations 10,000 times and calculated the correlation between the averaged dISC and each of these phase randomized engagement ratings. Thus, a null distribution and an actual correlation was obtained. D, The results from the comparison of the actual correlation to the null distribution. The orange, cyan, and, purple violin plots correspond to the FC5, FC1, and CP2 channels, respectively. The location and the obtained p-values are reported in this panel. The black horizontal lines represent the observed correlations. E, The results from the comparison of the actual correlation to the null distribution, separately for each phase of narrative. The first row shows the actual correlation against the null distribution and the second row indicate the channel locations which their correlation is reported. The phases in which dISC was significantly correlated with engagement ratings (rising, crisis, and climax) are shown with colorful violin plots and channel locations. The phases in which dISC failed to significantly predict the engagement ratings (exposition, falling, and denouement) are plotted in gray. Total of two channels (CP5 and FP2) in rising, three channels (Pz, Cz, and FP2) in crisis, and two channels (FT8 and FC6) in climax were the predictor channels. For the nonsignificant phases (exposition, falling and denouement), we just plotted three examples of the channels (C4, FC6, and CP5) to show where the actual correlation is located against the null distribution.
- Figure 3.
The schematic representation of the dynamic predictive model. A, The sample of preprocessed signals and sliding window (gray squares). B, The functional connectivity among all channels was obtained from the signals occurred within each window. C, Out of each functional connectivity matrix, three graph features, node degree, clustering coefficient, and betweenness centrality, were extracted. D, The procedure of the predictive model. In “Step 1,” we concatenated all the functional connectivity matrices (or graph features) of all subjects and divided it to a train and test sets (the leave-one-subject-out procedure). In “Step 2,” we trained a support vector regression model using the functional connectivities of all subjects but one to predict the averaged engagements and tested the model on the leaved-out subject. The output of Step 2 is the predicted engagement rating. By calculating the correlation between actual engagement and predicted engagement, we obtained the actual correlation. In “Step 3,” we shuffled the engagement ratings as presented in Figure 2 and repeated Step 2 to predict these shuffled ratings. By calculating the correlation between the predicted engagement and each shuffled engagement scores and repeating it for all the 1000 shuffled engagement scores, we obtained a null distribution of correlations. By testing the actual correlation on this null distribution, we obtained the significancy level of the prediction. Note that we implemented this procedure for functional connectivity features and all three graph features separately.
- Figure 4.
The engagement level predictions for the whole narrative. A, The testing of the actual correlation against the null distribution. Each violin plot refers to the prediction results for each EEG feature. The colored violin plots refer to statistically significant prediction features, and the gray plots refer to the features that were not statistically significant. The dark red violin plot refers to the dFC feature, and dark green plot refers to the BC feature. The light to dark gray spectrum refers to the CC, ND, and EEG features, respectively. The gray circles in the violin plots show the actual correlation obtained in each cross-validation fold with reference to each participant. The red horizontal lines show the observed correlation. The stars indicate that the corresponding feature had been a predictor that was chosen because of it being significantly higher than a chance occurrence. B, The predictor FCs that are either positively (red) or negatively (blue) correlated with engagement ratings across brain regions (F = frontal, C = central, P = parietal, T = temporal, O = occipital). Each cell represents the average number of times that each region played a predictor role. The light to dark colors (red and blue) refer to the strength of their score, which is the number of times they were considered a predictor divided by the length of all possible connections with other channels. To show the positive and negative features simultaneously, we used an upper triangle for positive and a lower triangle for negative features. The original features are symmetrical. C, the predictor FC features and channels on the scalp from three perspectives. The predictor FCs that were averaged in panel B are shown individually here in that the positively correlated items are indicated by red and the negatively correlated items are indicated by blue. The predictor channels obtained from the BC features are represented as colored dots, with bigger dots denoting higher proportions.
- Figure 5.
The results for predicting group averaged engagement ratings using dISC-bands. The violin plots refer to the null distribution of the correlation value (r) between surrogated engagement scores and the calculated dISC-bands, only for the significant channels. The observed r, which is the correlation between actual engagement scores and dISC-bands, is also indicated black lines over violin plots. The topo-plots indicate the observed r of the channels that could significantly predict the engagement scores, on the scalp. The p-values are reported above each violin plot, and all of them are corrected. In dISC-δ and dISC-θ, five channels could significantly predict the group-averaged engagement ratings. The spatial activity of the predictor channels can be seen on the topo-plot beside the violin plot. In dISC-β, two channels could significantly predict the engagement scores and the spatial distribution of the predictor channels can be seen in the topo-plot.
- Figure 6.
See Figure 4. The predictions of the engagement ratings of each phase using the corresponding features. The black horizontal lines show the observed correlation, and the violins show the null distribution. A, The phase prediction results for all phases using dFC features. It shows that the prediction in the last two phases (falling action and denouement) was significantly higher than could be explained by chance. B, The prediction results using BC features and the significant prediction that occurred in the last two phases. C, The predictor FCs across regions and their proportion scores for each phase. D, The predictive FCs and the channels that the BCs played as predictor roles for each phase.
Tables
Phase Starting point (s) Duration (s) Number of participants who agreed (%) SD of all responses (s) Exposition 0 151 19 (100%) 0 Rising action 151 149 13 (68.42%) 95.83 Crisis 300 45 10 (52.63%) 65.92 Climax 345 115 12 (63.16%) 32.37 Falling action 460 22 17 (89.47%) 23.74 Denouement 482 24 18 (94.74%) 4.05 This table shows the time stamps for starting and duration made by independent raters. The choice of the starting time point of each phase of the dramatic arc was obtained from the data of 19 participants that indicated these moments. From the starting time points, the duration of each phase was calculated. The number of participants that indicated the same time points statistically and the standard deviation (SD) of all responses are indicated in the last two columns, respectively.
In this issue
