Fig 1.
We utilized two publicly available fMRI datasets of participants watching Sherlock (N = 16) [47] and Friday Night Lights (N = 35) [48]. Additionally, we collected behavioral ratings of the two episodes from a separate group of participants. Each participant is presented with one movie clip and is instructed to continuously rate how the clip is making them feel in terms of either valence (i.e., positive to negative) or arousal (i.e., not aroused to aroused). The 120 participants in the rating study were equally divided between conditions. Thus, there were 30 participants per pairing of an affective dimension (valence or arousal) and movie (Sherlock or Friday Night Lights). We blurred the Images in the figures for copyright reasons; in the experiment, movies were shown at high resolution. We built within and across dataset CPMs to predict moment-to-moment affect ratings from dynamic whole-brain functional connectivity patterns.
Fig 2.
Affective experience is synchronized across participants in all four experimental conditions.
A. Participants’ subjective affective experience fluctuates over time during naturalistic movie watching. The red and blue solid lines indicate average arousal and valence time courses respectively. The gray areas indicate the standard deviation of ratings across participants at each time point. Each condition includes 30 participants. B. Histograms of the similarity of each individual’s subjective ratings to the group-averaged rating with the individual left out. Higher mean r values indicate stronger affective synchrony.
Fig 3.
Dynamic functional connectivity predicts fluctuations in arousal but not valence.
CPM performance in predicting arousal (A) and valence (B) within-dataset (the left panel) and between-dataset (the right panel). The y-axis represents the predictive accuracy, as measured by Pearson’s correlation between the model predicted time course and the observed group-average time course. Each datapoint in the box plot represents the predictive accuracy of each round of cross-validation. The black horizontal lines show the mean r value computed from the average of the Fisher-z transformed individual-subject r values. The gray half-violin plots show the null distribution of 1000 permutations, generated by phase-randomizing the observed group-average time course before training and testing the models. *p < 0.05, **p < 0.01, n.s.: p > 0.05, as assessed by comparing the empirical mean predictive accuracy against the null distribution.
Fig 4.
Functional anatomy of arousal.
A. Arousal networks for Sherlock, Friday Night Lights, and their overlap. Each cell represents the proportion of selected FCs among all possible FCs between each pair of functional networks. The cells in the upper triangle represent the high-arousal network (red), and the cells in the lower triangle represent the low-arousal network (blue). Network pairs with above-chance selected FCs are indicated with an asterisk (one-tailed t-test, FDR-corrected p < .05). B. Visualization of the overlap arousal network. Each node represents a brain region. The lines connecting two nodes show their functional connection (red: high-arousal network; blue: low-arousal network). C. Connectome-based model trained on the overlap arousal network generalized to two more fMRI datasets, Merlin and North by Northwest. The left figure indicates how well the model could predict the group averaged experiences of arousal. We averaged the model predictions across participants watching the same movie and computed the correlation between the group-average model predictions (predicted arousal) and group-average arousal ratings from a separate group of individuals (observed arousal). The gray bands indicate the standard deviation of predicted arousal across participants at each time point. The right figure shows the predictive accuracy of the across-dataset prediction. The model was trained on a minimal set of FCs that predicted arousal in both Sherlock and Friday Night Lights, and tested separately on Merlin and North by Northwest. **p < .01, as assessed by comparing the empirical mean predictive accuracy against the null distribution.