Intermittent brain network reconfigurations and the resistance to social media influence

Abstract Since its development, social media has grown as a source of information and has a significant impact on opinion formation. Individuals interact with others and content via social media platforms in a variety of ways, but it remains unclear how decision-making and associated neural processes are impacted by the online sharing of informational content, from factual to fabricated. Here, we use EEG to estimate dynamic reconfigurations of brain networks and probe the neural changes underlying opinion change (or formation) within individuals interacting with a simulated social media platform. Our findings indicate that the individuals who changed their opinions are characterized by less frequent network reconfigurations while those who did not change their opinions tend to have more flexible brain networks with frequent reconfigurations. The nature of these frequent network configurations suggests a fundamentally different thought process between intervals in which individuals are easily influenced by social media and those in which they are not. We also show that these reconfigurations are distinct to the brain dynamics during an in-person discussion with strangers on the same content. Together, these findings suggest that brain network reconfigurations may not only be diagnostic to the informational context but also the underlying opinion formation.

: EEG sensor position. The topo plot shows the sensor montage. The EEG recording system model used was B-Alert R X24 wireless sensor headset (Advanced Brain Monitoring, Inc., Carlsbad, CA, United States) with 20 channels.

Node allegiance
To further understand the differences between those that did and did not change their opinion in terms of network reconfigurations, we explored how the network nodes, particularly the nodes that showed significant difference in flexibility between two groups, (significant nodes, Figure 2) changed their functional allegiances. Allegiance is defined as the fraction of the total time two nodes are in the same community. First, we calculated the allegiance metric between all the node pairs and then compared them between two groups. None of the node pairs (including or excluding the significant nodes) showed a significant difference between groups. Moreover, on average, some of the node-pairs showed higher allegiances for those without a change in opinion while some showed higher allegiances for those who changed their opinion. In Figure S2, we show average allegiance differences between two groups.
Yellow entries in the matrices represent higher allegiances for those with no change.
In topographical plots we show these differences only for the significant nodes. Figure S2: Node allegiance differences. The matrices show average allegiance differences between the two groups (no change and change). The topographic plots show the mean allegiance differences for sensors which showed significant flexibility changes between the two groups (as discussed in Figure 2). The links in yellow (blue) indicate a higher (lower) allegiance value for the no change(change) group.

Effects of temporal window sizes
To explore the effects of different temporal window sizes we compared the mean wPLI temporal coefficient of variation (CoV) for the opinionators with and without opinion change during the social media platform interaction, the results are summarized in Fig.   S3. As a general effect of the increase in the window size we observe higher values and broader distributions of the mean temporal CoV. The differences between the two groups captured on the gamma band are present for temporal windows larger than 10s, indicating that smaller temporal windows do not capture the temporal scale of the dynamics that reflect the different processes occurring in the two groups.

Opinion changes on Scenario 2 time in jail
A considerable amount of subjects (24 subjects) that changed opinions on scenario 2 changed only the time in jail. Since our group division on opinion change binarizes this response, a concern that might arise is if just a small change on the time in jail would classify a subject as having a change in opinion when there actually very little change (<1 year sentencing change, for example). To address this concern we plot the distribution of differences of time in jail between questionnaire 1 and 2, this is shown in Figure S4. All differences observed accounts for at least 10 years which is a substantial time and was subsequently considered an opinion change. Besides the changes in years in prison for a convicted murder, we also observed 9 subjects that showed opinion changes involving non-numerical values (death penalty and life sentence), the labels for those options were standardized to avoid identification of false differences due to typos and differences on capital letters.

Example articles presented to subjects:
As the posts presented to participants were inspired by real articles, below is a nonexhaustive list of some of content that was linked from the posts, from the travel https://www.nytimes.com/2018/10/02/reader-center/donate-indonesia-tsunamiearthquake-victims.html

Comparison of link weights using wPLI and PLV
To disentangle the effects of PLV and wPLI, we first inspect the proportion of overlap at a variety of thresholds across wPLI and PLV. Figure S5 below visually depicts the proportion of edges that are overlapping after a binarization of the connectivity matrix for PLV and wPLI. Interestingly, within this dataset, at the extremes of connectivity, where t is the threshold of PLV and wPLI (t < 0.1 || t > 0.8), PLV and wPLI show very similar connectivity patterns; however, at the lower range (t > 0.15 && t < 0.6) there is substantial uniqueness of these connectivity patterns. We conclude that only the midrange connectivity values have variable amounts of contribution from the common source problem, or volume conduction. In the figure below, we plot the mean, standard deviation, and coefficient of variation across time. For the mean, as the frequency increases, the overall mean decreases relative to the lower frequencies, suggesting that higher frequencies are less likely to be ambiguous in volume conduction effects or the "common sources problem"; however, the mean wPLI of the lower frequencies (e.g., delta, theta, alpha) is within the range of ambiguous sources. Figure S5: wPLI and PLV differences on the link weights. A Shows the comparison of wPLI and PLV edges present as we increase the threshold for considering an edge as present. B-D show, respectively the mean, standard deviation and coefficient of variation of the wPLI weights for both groups (opinion change and no change). Albeit the fact that the region of the minima is contained in the mean wPLI distribution range, the wPLI coefficient of variation shows that the distribution has a high variance. Figure S6: A-C show, respectively the mean, standard deviation and coefficient of variation of the PLV weights for both groups.
The statistics of the weight distributions calculated based on PLV are presented in Figure S6, the distributions cover a larger range of values when compared with wPLI with mean PLV distributed in a range of higher values than wPLI especially for the higher frequency bands. From these two figures, considering the high overlap between metrics at the extreme values and the high variability in PLV that, it is unlikely that volume conduction significantly contributes to our results. Figure S7: The platform and general setup. Subjects were presented with curated online content to elicit decision making processes (left) whilst instrumented with EEG (middle) and allowed to freely scroll through the content. High level alpha differences between rest and platform engagement are shown (right).