A Neural Signature of the Bias Toward Self-Focus

Abstract

People are remarkably self-focused, disproportionately choosing to think about themselves relative to other topics. Self-focus can be adaptive, helping individuals fulfill their needs. It can also be maladaptive, with self-focus a risk and maintenance factor for internalizing disorders like depression. Yet, the drive to focus on the self remains to be fully characterized. We discovered a brain pattern that when spontaneously brought online during a quick mental break predicts the desire to focus on oneself just a few seconds later. In Study 1 (19 female and 13 male human subjects), we identified a default network neural signature from pr-trial activity that predicts multiple indicators of self-focus within our sample. In Study 2 (588 female and 498 male human subjects), we applied our neural signature to independent resting-state data from the Human Connectome Project. We found that individuals who score high on internalizing, a form of maladaptive self-focus, similarly move in and out of this pattern during rest, suggesting a systematic trajectory toward self-focused thought. This is the first work to "decode" the bias to focus on the self and paves the way toward stopping maladaptive self-focus in its course.

Keywords: default mode network; fMRI; self; social cognition.

Figure 1.

A, The choice task (used to derive the pre-self pattern) included three parts that repeated. First was a pretrial jittered rest period (2.5–6 s; mean, 4.5 s); second was the choice activity where participants chose who (themselves, a designated friend, or Biden) they wanted to think about in a later task (5 s); third was a shape-matching task aimed at moving participants' minds off their last choice. These three parts repeated 54 times in run 1 and 54 times in run 2 for a total of 108 trials. Our analysis examined neural activity during the pretrial jittered rest in relation to the subsequent task choice and response time. B, The first fMRI run that the participants completed was an 8-min-long resting-state scan. The 8 min were broken up into four, 2-min-long sections. After each 2 min section, participants had 32 s to rate the extent to which they were thinking about themselves, others, the future, and the past. These ratings were made on a continuous scale with “not at all’ on one end, “completely” on the other end, and “somewhat” in the middle of the scale. C, The final task participants completed was a typical self-reflection task. In this task participants were instructed to rate how well an adjective describes their personality in different social roles (Friend, Student, Significant Other, Son/Daughter, and Worker) on a scale from 1 to 4 using button boxes. We used the Big 5 list of 100 adjectives (Goldberg, 1993a). The rating trials were 4 s long and the jittered rest was 1–3 s, with mean of 2 s. There were two runs of 101 trials each, with a total of 202 trials.

Figure 2.

A, Participants choose to think about themselves significantly more than their self-nominated friend [χ² (1, N = 2,823) = 114.69, p < 0.001] or Biden [χ² (1, N = 2,320) = 495.34, p < 0.001]. B, Participants are faster in their decisions to think about themselves in comparison with a friend (β = 0.11, standardized β = 0.02, t₍₃₄₂₉₎ = 4.68, p = 0.003) and Biden (β = 0.08, standardized β = 0.03, t₍₃₄₃₅₎ = 2.93, p < 0.001). *** indicates p < 0.001; ** indicates p < 0.005.

Figure 3.

The default network's core subsystem during pretrial rest periods predicted the decision to choose to think about the self with 83% accuracy. This accuracy was significantly better (p < 0.001) than the average classification accuracy in the generated null distribution, which was 49%. A, The default network's core subsystem. B, The multivoxel pattern generated by the SVM to differentiate rest activity that proceeds self-choice rather than other-choice. Red indicates regions where activity predicts self-choice and blue indicates regions where activity predicts other-choice. The pattern shown was created with all subjects although classification accuracy was computed with a sixfolds approach.

Figure 4.

Applying the pre-self pattern to a resting-state scan with experience sampling. A, Our analytic approach started with a within-subjects, TR-to-TR multivariate pattern similarity analysis, assessing the similarity between (1) the pre-self pattern and (2) subjects’ resting-state scan pattern in the default network core subsystem. Those correlation values were then averaged for the 2 min rest sections and Fisher z-transformed. A linear mixed model assessed how self-reported thought content (self, other, future, and past) as well as the section of rest affected the mean correlation. B, In the results graph, the y-axis displays the normed mean correlation of the resting-state neural patterns (in the default network core) with the multivoxel classifying pre-self pattern generated (in the default network core). The x-axis is the self-reported self-focus rating that follows each of the 2 min sections of rest from 1 = “not at all” to 5 = “completely.” The dark gray bars reflect the confidence intervals based on standard error (± 1.96 * SE). We found that the strength of the default network core subsystem multivoxel pattern during rest was significantly related to self-reported self-focus (β = 0.19, standardized β = 0.09, t_(105.1) = 2.03, p = 0.045). In other words, the default network core multivariate pattern derived from short, jittered rest to predict subsequent choice to think about the self, is also able to decode self-focused thought during resting state.

Figure 5.

Relationship between pre-self pattern and active self-reflection pattern over the resting-state scan (A) depicts the approach using an 8 TR delay: the pre-self pattern (blue) was correlated with each TR (i.e., 1 s) of the resting-state scan and the active self-reflection pattern (orange) was also correlated with each TR of the resting-state scan. Linear mixed models (Bates et al., 2015) assessed if pre-self pattern correlation strength predicted active self-reflection pattern correlation strength 0–20 TRs later (here 8 TRs later). This allowed us to assess whether the presence of the pre-self pattern temporally predicted the presence of the active self-reflection pattern at multiple lag times. B, The results for a single subject using an 8 TR delay, demonstrating that the presence of the pre-self pattern (blue) predicts the presence of the active self-reflection pattern (orange) 8 s later. Note that the active self-reflection pattern strength visualized in orange has been shifted in time (n + 8 s) to help visualize its relationship with the pre-self pattern.

Figure 6.

Intersubject representation similarity analysis (IS-RSA) with Anna Karenina model: high internalizations have a similar pre-self pattern time course during a resting-state scan. The Anna Karenina model testing whether high internalizers move in and out of the pre-self pattern similarly over the course of a resting-state scan. A, Similarity in the timing of the pre-self pattern strength was computed for all pairs of Human Connectome Project participants and compared with (B) the theoretical model that high internalizers show a similar time course to one another while low internalizers show idiosyncratic time courses. C, A nonparametric, Mantel permutation test (see Materials and Methods) showed the internalizing Anna Karenina model was significant.

Figure 7.

Intersubject representational similarity analysis (IS-RSA): Example subjects' pre-self pattern instatement time course for our IS-RSA. We compared the time course of individuals' pre-self pattern instatement strength with an Anna Karenina model of internalizing such that the higher a pair's internalizing rank, the higher their similarity in pre-self pattern instatement timing, whereas the lower their internalizing rank, the more idiosyncratic their timing. A shows two high-internalizing subjects (90th percentile) pre-self pattern instatement time course while B shows two low-internalizing subjects (10th percentile) time course.