Segregation, integration and balance in resting-state brain functional networks associated with bipolar disorder symptoms

Abstract

Bipolar disorder (BD) is a serious mental disorder involving widespread abnormal interactions between brain regions, and it is believed to be associated with imbalanced functions in the brain. However, how this brain imbalance underlies distinct BD symptoms remains poorly understood. Here, we used a nested-spectral partition (NSP) method to study the segregation, integration, and balance in resting-state brain functional networks in BD patients and healthy controls (HCs). We first confirmed that there was a high deviation in the brain functional network toward more segregation in BD patients than in HCs and that the limbic system had the largest alteration. Second, we demonstrated a network balance of segregation and integration that corresponded to lower anxiety in BD patients but was not related to other symptoms. Subsequently, based on a machine-learning approach, we identified different system-level mechanisms underlying distinct BD symptoms and found that the features related to the brain network balance could predict BD symptoms better than graph theory analyses. Finally, we studied attention-deficit/hyperactivity disorder (ADHD) symptoms in BD patients and identified specific patterns that distinctly predicted ADHD and BD scores, as well as their shared common domains. Our findings supported an association of brain imbalance with anxiety symptom in BD patients and provided a potential network signature for diagnosing BD. These results contribute to further understanding the neuropathology of BD and to screening ADHD in BD patients.

Keywords: bipolar disorder; calibrated; fMRI; functional balance; functional connectivity; nested-spectral partition.

FIGURE 1

Brain network deviation toward higher segregation in BD patients. (a) Group‐averaged FC network for HC and BD groups, visualized using BrainNet viewer (Xia et al., 2013) with a binarizing threshold of 0.62. The connectivity densities were also provided (p = .176; see figure S2 for more comparisons). (b) The partition of seven functional systems. (c) Comparisons of network integration component $H_{\mathit{\text{In}}}$, segregation component $H_{\mathit{Se}}$, and balance indicator $H_B$. (d) Regions with significantly changed $H_{\mathit{\text{In}}}^i$, $H_{\mathit{Se}}^i$ and $H_B^i$ (p < .05, uncorrected). The color bar represents the difference between the BD and HC groups. (e) Differences in $H_{\mathit{\text{In}}}$, $H_{\mathit{Se}}$ and $H_B$ in different systems between the HC and BD groups (black dot: p > .05).

FIGURE 2

Brain network balance is associated with lower anxiety in BD patients. A quadratic regression model (i.e., y ~ x ² + x) was used to identify relationships between anxiety scores and (a) integration component $H_{\mathit{\text{In}}}$, (b) segregation component $H_{\mathit{Se}}$ and (c) balance indicator $H_B$. (d) Linear relationship between $\left|H_B\right|$ and anxiety scores.

FIGURE 3

Balanced segregation and integration better predict BD symptoms. (a) Correlations between symptom scores and predicted scores using different brain measures for anxiety, energy, daydream, and mood. As comparisons, the results for degree (Deg; sum of weighted FC) and participant coefficient (PC) were also provided. (b) The best predictions for BD symptoms were statistically tested by using the permutation test (10,000 times). *p < .05, **p < .01, ***p < .001. (c) The weights of regions in the best prediction model were mapped to the brain surface. (d) The mean weight of the regions in systems in the best predictions.

FIGURE 4

Predictions of BD and ADHD scores. (a) Predictions of BD sumscores. From the left panel to the right panel: Prediction performances, the best prediction, and the weights in seven systems. (b) Predictions of ADHD scores. (c) he selected features in (a) were used to predict ADHD scores, and the features in (b) were used to predict BD scores. (d) PCA was applied to obtain the dominant common score, which was subsequently predicted through the machine‐learning approach.