Connectome-Based Individualized Prediction of Loneliness

Chunliang Feng; Li Wang; Ting Li; Pengfei Xu


Soc Cogn Affect Neurosci. 2019;14(4):353-365. 

In This Article

Material and Methods


Seventy-five healthy right-handed college students from Beijing Normal University (62 males and 55 singles; age 21.88 ± 3.01 years) without history of neurological or psychiatric disorder were recruited. The study was conducted in accordance with the 1964 Helsinki Declaration and its later amendments and was approved by the Ethics Committee of Beijing Normal University. Written informed consents were obtained from all participants.

Assessment of Loneliness

Loneliness was assessed using the Revised UCLA Loneliness Scale (Russell, 1996), which is a well-validated measure of general feelings of loneliness. The scale consists of 20 items, and each item is scored on a 4-point Likert scale ranging from 1 (never) to 4 (always). The higher scores on the scale indicate higher levels of loneliness.

NEO Personality Inventory-revised

Personality was assessed by the NEO personality inventory-revised (Costa Jr and McCrae, 1992). The scale consists of 120 items and assesses the five different dimensions of personality: neuroticism, extraversion, openness, agreeableness and conscientiousness. Each item is rated on a 5-point Likert scale ranging from 'strongly disagree' to 'strongly agree'.

Image Acquisition

Images were acquired on a Siemens 3-Tesla TRIO scanner at Beijing Normal University Imaging Center for Brain Research. The resting state scanning consisted of 150 contiguous echo-planar imaging (EPI) volumes using the following parameters: axial slices, 33; slice thickness, 3.5 mm; gap, 0.7 mm; repetition time (TR), 2000 ms; echo time (TE), 30 ms; flip angle, 90°; voxel size, 3.5 × 3.5 × 3.5 mm3 and field of view (FOV), 244 × 244 mm2. In addition, high-resolution structural images were acquired through a 3D sagittal T1-weighted magnetization-prepared rapid acquisition with gradient-echo sequence, using the following parameters: sagittal slices, 144; TR, 2530 ms; TE, 3.39 ms; slice thickness, 1.33 mm; voxel size, 1 × 1 × 1.33 mm3; flip angle, 7°; and FOV, 256 × 256 mm2.

All participants underwent a 5 min resting-state functional magnetic resonance imaging scanning, during which they were instructed to close their eyes, keep still, remain awake and not to think about anything systematically (see also Nooner et al.,2012). Several approaches were implemented to reduce the possibility that participants might fall asleep during the scan: (i) participants were explicitly instructed to close their eyes but not fall asleep during the resting-state scan; (ii) experimenters communicated with each participant immediately after the scan, and all participants responded promptly, indicating that they did not fall asleep; and (iii) the current study implemented rigorous criteria (see also 'image preprocessing') to exclude participants from further analyses based on their head motion. Therefore, it is likely that participants sleeping during the scan (therefore, lower control of head movements) were excluded from analyses in the current study.

Image Preprocessing

Neuroimaging data analyses were performed with the DPABI software package (Yan et al., 2016), which is a convenient software plug-in based on SPM12 ( The first 10 volumes of the functional images were discarded for signal equilibrium and participants' adaptation to scanning noise. The images were then realigned for head movement correction. Seven participants (6 males, 5 singles) were excluded from further analysis under the criteria of head motion exceeding 2.5 mm maximum translation, 2.5° rotation or mean frame-wise displacement exceeding 0.2 mm throughout the course of scans (Power et al., 2012; Yan et al., 2013). To normalize functional images, participants' structural brain images were first co-registered to their own mean functional images and were subsequently segmented. The parameters derived from segmentation were used to normalize each participant's functional images into the standard Montreal Neurological Institute space (MNI template, resampling voxel size was 3 × 3 × 3 mm3). Afterwards, the linear trends of time courses were removed, and a band-pass filtering (0.01–0 1 Hz) was applied to the time series of each voxel to reduce the effect of low-frequency drifts and high-frequency physiological noise (Biswal et al.,1995; Zuo et al., 2010). Subsequently, the images were spatially smoothed using a Gaussian filter to decrease spatial noise (4 × 4 × 4 mm3 full width at half maximum). Finally, three common nuisance variables were regressed out, including the white matter signal, the cerebrospinal fluid signal (Fox et al., 2005; Snyder and Raichle, 2012) and 24 movement regressors including autoregressive models of motion incorporating six head motion parameters, six head motion parameters one time point before and the 12 corresponding squared items (Friston et al., 1996).

RSFC Feature Extraction

In the current study, network nodes were defined by using a functional brain atlas, derived from a graph theory-based parcellation algorithm that maximized the similarity of the voxel-wise time series within each node (Shen et al., 2010; Shen et al., 2013). The atlas includes 268 nodes spanning the whole brain including cerebellum and brainstem (Figure 2A). Notably, the 268-node atlas comprises nodes with more coherent time series than those defined by the automatic anatomic labeling atlas and thus represents an improvement over anatomical parcellation schemes because anatomical boundaries do not always match functional ones (Shen et al., 2013).

Figure 2.

Macroscale regions used for characterizing contributing connectivity. (A) The 268 nodes. (B) Twenty macroscale brain regions. (C) The connectivity patterns selected by the prediction model, plotted as number of connections within each macroscale regions. (D) Connections plotted as number of edges within and between each pair of macroscale regions. L, left; R, right; PFC, prefrontal; Mot, motor; Ins, insula; Par, parietal; Tem, temporal; Occ, occipital; Lim, limbic; Cer, cerebellum; Sub, subcortical; Bsm, brainstem.

For each participant, the time course of each node was computed by averaging the blood oxygen level-dependent signal of all of its constituent voxels at each time point. Second, network edges were defined as functional connectivity between each pair of nodes, calculating as the correlation (Pearson's r) between time courses of each pair of nodes. Fisher's r-to-z transformation was then implemented to improve the normality of correlation coefficients, resulting in a 268 × 268 symmetric connectivity matrix that represented the set of edges/connections in each participant's resting-state connectivity profile (Finn et al., 2015; Rosenberg et al.,2016).

Exploratory Correlation Analysis

An exploratory correlation analysis was implemented across all participants to examine the relevance of RSFC to loneliness. Specifically, Pearson correlation between each edge in the connectivity matrices and loneliness scores was computed across participants. The resultant r values were forward to a threshold of P < 0.05 (Finn et al., 2015; Rosenberg et al., 2017; Rosenberg et al., 2018) and separated into a positive tail (i.e. positive correlation between strength of edge and loneliness scores) and a negative tail (i.e. negative correlation between strength of edge and loneliness scores). Therefore, connections in the positive tail (hereafter referred to as 'positive network') and negative tail (hereafter referred to as 'negative network') were selected by correlations with loneliness scores rather than positive or negative functional connections themselves (see also Rosenberg et al., 2016; Beaty et al., 2018; Hsu et al., 2018). Afterwards, a single aggregate metric of network strength was employed to characterize degree of connectivity in the positive and negative tails for each participant. That is, positive network strength was computed by summing the edge strengths (i.e. Z scores) for all the edges in the positive tail. Similarly, negative network strength was computed by summing the Z scores of all the edges in the negative tail. Lastly, the positive and negative network strengths were correlated with loneliness scores. Notably, results of this analysis were for display purpose, and no statistical tests were performed (Kriegeskorte et al., 2009; Kristensen and Sandberg, 2017). Furthermore, conclusions on the relationship between positive/negative network strengths and loneliness were not derived from this analysis, but instead were based on results from cross-validation detailed below. In other words, this analysis was conducted to illustrate an overview of data before formal prediction analysis (see also Rosenberg et al., 2016).

Prediction Analysis Using Cross-validation

To determine whether network strength predicted loneliness in unseen individuals, a leave-one-out cross-validation (LOOCV) was used to evaluate the out-of-sample prediction performance. Specially, N-1 participants were used as the training set and the remaining one was used as the testing sample, where N is the number of the participants. During the training procedure, predictive networks were defined and employed for calculating positive and negative network strengths as described in the exploratory correlation analysis. Afterwards, simple linear models were constructed to respectively relate positive and negative network strengths to loneliness scores in the training set. During the testing procedure, each testing participant's strengths of positive and negative network was normalized using the parameters acquired during training procedure, and then the trained models were used to predict the testing participant's loneliness score (Finn et al., 2015; Rosenberg et al., 2016; Shen et al.,2017). The training and testing procedures were repeated N times such that each participant was used once as the testing participant.

Pearson correlation coefficient (r) and mean squared error (MSE) between actual and predicted loneliness scores were used to evaluate the accuracy of prediction. The permutation test was applied to determine whether the obtained metrics were significantly better than expected by chance. Specially, we permuted the loneliness scores across participants without replacement 1000 times, and each time re-applied the above LOOOCV prediction procedure. This resulted in a distribution of correlation (r) and MSE values reflecting the null hypothesis that the model did not exceed chance. The number of times the permuted value was greater than (or with respect to MSE values, less than) or equal to the true value plus one was then divided by 1001 providing an estimated P-value for both the correlation coefficient (r) and observed MSE.

Contributing Network in the Prediction of Loneliness Scores

To characterize the neural substrates of the contributing network, the network was defined as the set of edges that were present in the every iteration of the LOOCV described above. Afterwards, the 268 nodes were grouped into 10 macroscale brain regions, including the prefrontal lobe (46 nodes), motor lobe (21 nodes), insular lobe (7 nodes), parietal lobe (27 nodes), temporal lobe (39 nodes), occipital lobe (25 nodes), limbic lobe (36 nodes), cerebellum lobe (41 nodes), subcortical lobe (17 nodes) and brainstem lobe (9 nodes) (Finn et al., 2015; Rosenberg et al., 2016). The number of edges between each pair of macroscale regions was then calculated. Furthermore, the importance of individual nodes was measured as the number of their connections (Rosenberg et al., 2016; Beaty et al., 2018). The connectivity patterns of the top 20 most highly connected nodes were illustrated.

Validation Analysis With Different Cross-validation Schemes

Main results were further validated with different cross-validation schemes (i.e. 2-fold, 5-fold and 10-fold). Taken the 2-fold cross-validation as an example, all participants were divided into two subsets, in which one subset was used as the training set, and the remaining one was used as the testing set. Training set was normalized and used to train a linear prediction model, which then was used to predict scores of the normalized testing data. The normalization of testing data used the normalizing parameters acquired from training data. This procedure was repeated twice, so that each subset was used as testing set once. Finally, the correlation r and MSE between the true and predicted scores were calculated across all participants. As the full data set were randomly divided into two subsets, the performance might depend on the data division. Therefore, the 2-fold cross-validation was repeated 100 times, and the results were averaged to produce a final prediction performance. A 1000 times permutation test was applied to test the significance of the prediction performance.

Control Analyses

Several control analyses were implemented to further examine the significance of predictions of our models despite potential confounds of age, gender, relationship status (single vs in a romantic relationship) and motion. In these analyses, new predictive networks were constructed by employing those edges whose partial Pearson correlation with loneliness scores while controlling for confounding variables (e.g. motion) passed the P < 0.05 threshold (see also Shen et al., 2017; Hsu et al., 2018). Finally, head motion was further controlled for in the data preprocessing, such that volumes with an FD > 0.5 mm, along with the immediately preceding volume and two subsequent volumes, were considered micromovement-containing volumes, and each of these volumes was modeled as a separate regressor in nuisance covariates regression (Yan et al., 2013; Power et al., 2014).

Relationship of Personality With Loneliness and Associated Network Connectivity

The associations between loneliness and five personality dimensions (neuroticism, extraversion, openness, agreeableness and conscientiousness) were estimated with a linear regression, with the loneliness as the dependent variable and five personality dimensions as predictors. Since the regression analysis revealed reliable association of loneliness with neuroticism and extraversion (see also Results section), we examined whether networks contributing to the prediction of loneliness were capable of predicting neuroticism and extraversion. In these analyses, connectivity features selected by the prediction model of loneliness were forward to the predictive models for these personality scores. In other words, these analyses examined whether loneliness-related predictive networks were also associated with neuroticism and extraversion. Finally, control analyses were conducted to examine whether RSFC-based model could still predict loneliness after controlling for neuroticism and extraversion (for details, see also 'Control analyses' section).