Loneliness and brain rhythmic activity in resting state: an exploratory report

Abstract

Recent studies using resting-state functional magnetic resonance imaging have shown that loneliness is associated with altered blood oxygenation in several brain regions. However, the relationship between loneliness and changes in neuronal rhythm activity in the brain remains unclear. To evaluate brain rhythm, we conducted an exploratory resting-state electroencephalogram (EEG) study of loneliness. We recorded resting-state EEG signals from 139 participants (94 women; mean age = 19.96 years) and analyzed power spectrum density (PSD) and functional connectivity (FC) in both the electrode and source spaces. The PSD analysis revealed significant correlations between loneliness scores and decreased beta-band powers, which may indicate negative emotion, attention, reward, and/or sensorimotor processing. The FC analysis revealed a trend of alpha-band FC associated with individuals’ loneliness scores. These findings provide new insights into the neural basis of loneliness, which will facilitate the development of neurobiologically informed interventions for loneliness.

Introduction

Loneliness is a negative emotional state resulting from a perceived discrepancy between desired and actual social relationships (Peplau and Perlman 1982, Hawkley 2022). This feeling is subjective, and individuals may perceive themselves as socially isolated even if they are not objectively isolated (Cacioppo and Cacioppo 2018). Loneliness is not a clinical diagnosis, but it is associated with a wide range of poor health outcomes, including both physical and mental (Hawkley 2022, Mann et al. 2022). Therefore, loneliness is an excellent index of psychiatric risk in healthy populations. Identifying high-risk populations and developing neurobiologically informed interventions for loneliness require a search for mechanistic underpinnings and biomarkers (Lam et al. 2021, Farzan 2024). Resting-state functional magnetic resonance imaging (fMRI) studies have demonstrated that loneliness is associated with altered blood oxygenation in various brain regions. However, the relationship between loneliness and brain rhythmic activity remains unclear. In this study, we conducted an exploratory resting-state electroencephalogram (EEG) study to investigate the relationship between loneliness and brain rhythmic activity.

Resting-state fMRI studies have demonstrated that loneliness is linked to alterations in brain activity. These changes have been observed in the control network involved in cognitive control processes, the dorsal attention network involved in goal-directed top-down attention processing, the ventral attention network involved in stimulus-directed bottom-top attention processing, the default network involved in social and self-related processes, the somatomotor network involved in motor preparation and control processes, and the visual network involved in visual processing [Mwilambwe-Tshilobo et al. 2019; see Lam et al. (2021) and Morr et al. (2022) for reviews].

First, loneliness was associated with the dorsolateral prefrontal cortex (Layden et al. 2017, Feng et al. 2019, Dong et al. 2021), which is a core node of the frontoparietal cognitive control network. Second, weaker causal flow from the dorsal attentional network to the ventral attentional network was associated with loneliness, resulting in decreased suppression of bottom-up attention (Tian et al. 2017). Increased parietal-cerebellar connectivity also supports improved bottom-up attention to interact with the environment (Shao et al. 2020). Third, lonely individuals demonstrate a greater magnitude of within-network coupling for the default mode network, which may imply enhanced internally oriented cognition to compensate for the missing social relationships that lonely individuals require (Spreng et al. 2020). Finally, loneliness positively correlated with spontaneous activity in the inferior temporal gyrus (Yi et al. 2018). This area is a key node for predicting loneliness (Feng et al. 2019) and plays a critical role in perception (Haxby et al. 2001, Barton et al. 2004). Taken together, resting-state fMRI studies have demonstrated that lonely individuals show changes in brain networks, which are known to be associated with multiple distinct but nonmutual functions, including cognitive control, attention, social cognition, and perceptual processing.

However, the blood oxygenation level–dependent (BOLD) signal measured using fMRI is an indirect measure of neuronal activity (Sirotin and Das 2009), and the outcomes of BOLD signals can be confounded by nonneural cerebrovascular alterations (Chen et al. 2022). Contrastingly, EEG directly measures neuronal population activity and allows for the assessment of brain rhythms, which reflect neuronal oscillations (Garcés et al. 2022). Rhythmic activities are categorized into different frequency bands: delta (1–3 Hz), theta (4–7 Hz), alpha (8–12 Hz), beta (13–29 Hz), and gamma (30–80 Hz). Delta oscillations are related to inhibitory control processes (Harmony 2013), while theta oscillations are associated with memory processes (Klimesch 1999). Alpha oscillations are the dominant intrinsic oscillatory activities of the human brain during eyes-closed rest (Palva and Palva 2007). Beta oscillations are associated with negative emotions, attention, reward, and sensorimotor processing (HajiHosseini et al. 2012, Mas-Herrero et al. 2015, Huebl et al. 2016, Chota et al. 2023, Roxburgh et al. 2023). Gamma oscillations are closely related to attentive processing of information, active maintenance of memory contents, and conscious perception (Singer 2001, Herrmann et al. 2004, Womelsdorf and Fries 2006).

Additionally, the EEG records electrophysiological signals primarily from postsynaptic potentials generated by cortical pyramidal cells (Baillet 2017). The summed postsynaptic potentials of all the aligned pyramidal cells in a patch of cerebral cortex are assumed to be equivalent to dipoles (Song et al. 2015). The current dipoles can be used to reconstruct the cortical source space and to localize the brain networks (Michel and Brunet 2019). The activation of these networks has been reliably assessed in the cortical source space to investigate the brain activity changes in psychiatry populations, such as autism spectrum disorder (Garcés et al. 2022), major depressive disorder (Rolle et al. 2020), and post-traumatic stress disorder (Toll et al. 2020).

Barros et al. (2022) have used resting-state EEG to assess the relationship between loneliness and frontal alpha asymmetry (i.e. alpha power over the right relative to the left frontal brain regions), and they found no significant correlation. However, previous studies suggest that it is possible to capture the relationship between loneliness and resting-state EEG signals. For instance, Tullett et al. (2015) found a significant correlation between resting-state EEG signals and nostalgia, a feeling often triggered by loneliness (Wildschut et al. 2006, Zhou et al. 2008). Research has shown a significant relationship between resting-state EEG signals and depressive symptoms among preclinical participants (e.g. Qin et al. 2022, Chen et al. 2024) and a significant relationship between resting-state EEG and anxiety (e.g. Roxburgh et al. 2023, Sidelinger et al. 2023), while depressive symptoms and anxiety have been found to be closely associated with loneliness (Grygiel et al. 2023, Wolters et al. 2023). Furthermore, brain network activity is rooted in electrophysiological processes (Sadaghiani et al. 2022). As mentioned earlier, previous fMRI studies have shown a correlation between loneliness and the activity of multiple brain networks.

Given the dearth of systematic research on the relationship between resting-state EEG and loneliness, in this study, we conducted exploratory research to reveal the potential relationship between loneliness and resting-state EEG signals. We initially conducted a resting-state power spectrum density (PSD) analysis, which can quantify local synchronization within brain regions. Studies have found that resting-state power changes in different frequency bands are associated with several psychiatric disorders [see Newson and Thiagarajan (2019) for a review]. To reveal more spatial information and facilitate comparison with previous fMRI studies, a PSD analysis was also conducted in the source space.

Moreover, we calculated resting-state functional connectivity (FC) across different frequency bands. Resting-state FC analysis can reveal the circuit mechanisms underlying large-scale organization and information processing of brain activity (Van Den Heuvel and Pol 2010, Siegel et al. 2012), and the prediction of long-term symptoms is now preferentially based on brain disconnections (de Schotten and Forkel 2022). Similar to the PSD analysis mentioned earlier, an FC analysis was performed in both the electrode and source spaces. We calculated FC using the orthogonalized power envelope (PE) correlation method (Hipp et al. 2012), which has high reproducibility (Duan et al. 2021). This method orthogonalizes the signals to remove the zero-phase lag, which reflects the effect of volume conduction. Volume conduction is a potential confounder that can lead to high spatial autocorrelation and spurious connectivity results because EEG signals measured at the scalp reflect the summation of many neuronal sources (Toll et al. 2020). Furthermore, we used connectome-based predictive modeling (CPM) analysis to evaluate FC between whole-brain regions in different frequency bands. CPM analysis adopts a multivariate pattern analysis based on whole-brain resting-state FC without any prior bias (Shen et al. 2017). Thus, CPM analysis is appropriate for the exploratory study.

This study is an exploratory on the altered PSD and FC in loneliness during resting state using electrode- and source-level analyses of EEG data. The initial findings of the study are mostly explorative before future confirmatory investigations could be proposed. Therefore, no specific hypotheses are made in this exploratory study. According to the findings of fMRI studies, we predict that the alterations could be revealed across a broad range of brain regions, including control, attention, default, and perceptual networks.

Materials and methods

Participants

A total of 147 college student participants [99 women; mean age (M) = 19.97 years, s.d. = 1.78 years, age range = 17–25 years) were recruited from Southwest University (Chongqing, China) through online advertising. Eight participants were excluded due to excessive EEG artifacts (segment rejection rate > 25%), leaving 139 participants for the final analysis (94 women; M = 19.96 years, s.d. = 1.79 years, age range = 17–25 years). All participants were right-handed and had no history of neurological or psychiatric disease. Written informed consent was obtained from all participants in accordance with the seventh revision of the Declaration of Helsinki (World Medical Association 2013). This study was approved by the Ethics Committee of the Faculty of Psychology, Southwest University, Chongqing, China.

Assessment of loneliness

Loneliness was assessed using the revised University of California Los Angeles Loneliness Scale (Russell 1996) for all participants. The questionnaire had 20 self-reported items scored on a four-point scale, ranging from 1 to 4, and had high internal consistency (Cronbach’s α = .89–.94) and excellent reliability (test–retest r = 0.73) (Russell 1996). Higher total scores on the questionnaire indicate higher levels of loneliness.

EEG data acquisition and preprocessing

Raw resting-state EEG data were recorded using an EEG system (Neuroscan, Herndon, VA, USA) with 64 Ag/AgCl electrodes arranged according to the international 10–20 system. To detect eye movement, the left eye supra- and infra-orbital electrodes were used to record the vertical electrooculogram (EOG), and the left versus the right orbital rim electrodes were used to record the horizontal EOG. All the electrode impedances were maintained <5 kΩ. The recording was digitized at a sampling rate of 500 Hz with a band-pass filter of direct current to 100 Hz and referenced to an electrode between FCz and Cz. Once the system was properly fitted, participants were asked to sit comfortably facing the black screen. Then, participants were prompted to remain in this position with eye-close for 5 min, remain relaxed, minimize eye movements, and stay awake. To prevent participants from falling asleep while recording, the experiment will be interrupted when higher amplitude and lower frequency EEG waves occur frequently. Additionally, after the experiment, participants were asked whether they stayed awake during the experiment. Consequently, three participants were re-recorded to obtain the full 5 min of awake and uninterrupted recoding.

Offline EEG data analysis was performed using the EEGLAB toolbox (Delorme and Makeig 2004) and custom scripts in MATLAB R2019b (MathWorks Inc., Natick, MA, USA). First, after down-sampling to 250 Hz, EEG data were high- and low-pass filtered at 0.5 and 80 Hz, respectively (‘eegfilt’ function with finite impulse response default settings).

The bad channels identified via visual inspection were interpolated using spherical methods. Independent component analysis (ICA) was performed to correct the artifact components. Artifact components containing eye, muscle, heart, 50-Hz line noise, and channel noise were detected using the ICLabel plugin (Pion-Tonachini et al. 2019). High-frequency activity signals recorded by noninvasive EEG are often contaminated with artifact signals of muscle activity (Muthukumaraswamy 2013). To reduce the effects of muscle artifacts, independent components were removed based on a prior comparably loose criterion that the probability of the artifact component label was >0% and “brain” the component label was <10%. Notably, up to 20% of the total number of components, i.e. 10, were allowed to be removed. On average, the mean of 5.80 (s.d. = 2.7) components was removed for each participant. Next, the ICA-corrected EEG data were re-referenced to a common average and segmented into 4-s nonoverlapping segments. Finally, segments with absolute EEG voltages exceeding ±85 μV were marked and removed. If >25% of the segments were marked and removed, the entire dataset was excluded from subsequent analysis. The mean number of removed artificial segments is 0.78 (s.d. = 1.87) for remaining 139 participants.

Resting-state PSD in the electrode space

For each participant, a PSD analysis was performed on the time series in the electrode space using Welch’s (1967) method. The method requires dividing the time series into overlapping sliding windows and performing a fast Fourier transform on each time-series window. Windows were set to 512-point length with 50% overlap. To reduce the spectral leakage, the Hamming function was multiplied by each window. Consequently, the PSDs with ∼0.5 Hz frequency resolution were obtained for each electrode. To minimize the multiple comparisons for the statistical analysis of PSDs, the PSDs were averaged within the nine electrode regions of interest (ROIs). ROIs were defined as follows: frontal left (AF3, F3, and F5), frontal midline (F1, FZ, and F2), frontal right (AF4, F4, and F6), central left (FC1, C1, CP1, and C3), central midline (FCZ, CZ, and CPZ), central right (FC2, C2, CP2, and C4), posterior left (P3, P5, and PO3), posterior midline (P1, PZ, and P2), and posterior right (P4, P6, and PO4). Five canonical frequency bands were determined: delta (1–3 Hz), theta (4–7 Hz), alpha (8–12 Hz), beta (13–29 Hz), and gamma (30–80 Hz). To normalize the distribution, the absolute power of each frequency band was logarithm-transformed.

SPSS software (version 26.0; SPSS, Chicago, IL, USA) was used to analyze the association between PSDs and loneliness scores. To remove the effects of differences in sex and age, a partial correlation analysis was performed for each of the five bands and nine ROIs. Additionally, a false discovery rate (FDR) correction (Benjamini and Hochberg 1995), 140 tests in total, including the subsequent analyses, with a family size of 140, was used for multiple comparisons to reduce the risk of type I errors, and the corrected P-values were subsequently obtained. Corrected P-values were considered significant at P < .05.

Resting-state PSD in the source space

EEG source reconstruction was conducted to obtain the source space current time series. Reconstruction of EEG cortical surface source space data was performed using the Brainstorm toolbox (Tadel et al. 2011). First, the symmetric boundary element method (Kybic et al. 2005) was used to build a realistic head model based on the Montreal Neurological Institute brain template (ICBM152) in OpenEEG (Gramfort et al. 2010). The noise covariance matrix was then set as an identity matrix, assuming that the noise variance in all electrodes was equal and unit. Next, three orthogonal direction current dipoles were generated at each of the 15 002 vertices on the cortical surface to obtain a spatial imaging kernel using the minimum norm estimation approach with depth weighting and regularization. To reduce the computational cost for each vertex, a single principal direction dipole was subsequently reserved using principal component analysis. Finally, the source space current density time series was obtained by projecting the electrode space time series onto a cortical surface. Furthermore, the 15 002 vertices were segmented into 100 nodes based on the Schaefer atlas (Schaefer et al. 2018), corresponding to the 17 brain functional networks: the control network (A/B/C), default network (A/B/C), dorsal attention network (A/B), limbic network (A/B), salience/ventral attention network (A/B), somatomotor network (A/B), temporo-parietal network, and visual network (central/peripheral). As with the electrode space, the PSD analysis was performed in the source space. The PSDs were obtained for each node and averaged for the 17 networks listed earlier. Finally, a partial correlation analysis was performed to evaluate the correlation between loneliness scores and natural logarithm-transformed PSDs for each of the five bands and 17 networks. The FDR correction, 140 tests in total, including the other analyses in this study, with a family size of 140, was used for multiple comparisons to reduce the risk of type I errors, and the corrected P-values were subsequently obtained. Corrected P-values were considered significant at P < .05.

Resting-state FC analysis in the electrode space

FC analysis was performed in the electrode space for each participant based on the PE correlation between orthogonalized signals. First, the EEG time-series signals on 60 electrodes were band-pass filtered into five frequency bands, as previously described. For each frequency band, the Hilbert transform was subsequently performed to yield analytical signals on the 60 electrodes (Fig. 1a). The analytical signal on one electrode was selected for orthogonalization concerning all other electrodes, resulting in 59 orthogonalized analytical signals. Then, the selected signal was multiplied by its conjugate signal to obtain plain PEs, and the 59 orthogonalized PEs were multiplied by their conjugate signals to obtain 59 orthogonalized PEs. The plain and orthogonalized PEs were natural logarithm-transformed to normalize the distribution. Next, 59 Pearson correlation coefficients (edges) between the plain and 59 orthogonalized PEs were calculated. Repeating the above process for the analytical signals on all other electrodes, we obtained a correlation coefficient matrix of 60 × 60 (excluding the main diagonal). Finally, the matrices were Fisher-transformed to obtain the final FC matrix (Fig. 1b). Consequently, we obtained FC matrices within each of the five frequency bands for each participant.

Figure 1.

The workflow of the FC analysis. (a) The calculation of the analytical signals in the electrode and source space. The electroencephalography time-series signals were recorded by 60 electrodes on the scalp. These signals were then band-pass filtered and Hilbert transformed to obtain the band-limited analytical signals of the electrode space. The spatial imaging kernel was estimated using the minimum norm estimation approach. The kernel was multiplied by the analytical signals of the electrode space to obtain the analytical signals of the source space. (b) The calculation of the orthogonalized PE correlations between the analytical signals for each frequency band. The analytical signal on one electrode or vertex was selected to orthogonalize with the analytical signals on all other electrodes or vertices. The orthogonalized analytical signals on all other electrodes or vertices were obtained. The Fisher-transformed correlation coefficients of these PEs were calculated. The procedure was iterated 60 times in the electrode space or 15 002 times in the source space to obtain the orthogonalized PE correlation coefficients between the analytical signal on all electrodes or vertices. The correlation coefficient matrix was the final FC matrix in the electrode space. In the source space, the correlation coefficient matrix was averaged within 100 brain nodes to obtain the final FC matrix.

The FC matrix for each frequency band was used to build a predictive model using the CPM analysis. Leave-one-participant-out cross-validation (LOOCV) was used to validate the model (Fig. 2). Furthermore, a permutation test was conducted to evaluate the statistical significance of the model performance.

Figure 2.

The procedure of leave-one-participant-out cross-validation in CPM analysis. (a) One participant was selected as a test sample, and all other participants as the training set. (b) A partial correlation was calculated between loneliness scores and edges of FC in the training set, and the significant edges were combined as features. (c) The features and loneliness scores of participants in the training set were submitted to SVR analysis to build a predictive model. (d) The features of the test sample were submitted to the predictive model, and a predicted score was obtained. (e) The (a)–(d) steps were iterated 139 times to obtain the predicted scores for all participants. (f) Spearman’s rank correlation between the predicted and actual loneliness scores was calculated.

Figure 2 shows the procedure of LOOCV. One participant was used as the test sample, and the remaining participants were used as the training set. The edges significantly (P < .05) associated with loneliness in the training set were selected as features. The features of all participants in the training set were combined as a feature set to train the model. A partial correlation analysis was performed to evaluate the association strength to remove the effects of differences in sex and age. Then, the feature set combined with the loneliness scores was submitted to support vector regression (SVR) analysis with a standard regularizing parameter C = 0.1 to build a prediction model using the LIBSVM library (Chang and Lin 2011). Once the predictive model was built, the features of participants in the test sample were input into the model to obtain a predicted loneliness score. This procedure was iterated 139 times until the predicted loneliness scores for all participants were obtained. Finally, the actual prediction result r-value was defined as the correlation coefficient between the predicted and actual scores, calculated using Spearman’s rank correlation to avoid the effects of skewed data and outliers.

To evaluate the prediction performance of the built model, the statistical significance of the correlation between the predicted and actual scores was assessed using a permutation test. First, the previously described LOOCV procedure was conducted on randomly shuffled loneliness scores and iterated 1000 times to obtain a null distribution of chance predicted result r-value, defined as the correlation coefficient between the predicted and shuffled loneliness scores. Second, the number of null r-values in the distribution is greater than or equal to the actual r-value divided by 1000, providing an estimated P-value for the statistical significance of the predicted result.

Accordingly, a prediction model and statistical P-value were obtained for each of the five frequency bands. To control for multiple comparisons, 140 tests in total, including the PSD analyses, FDR correction with a family size of 140 was used, and an FDR-corrected P-value of <.05 was considered significant.

Resting-state FC analysis in the source space

We obtained a spatial imaging kernel of 15 002 diploe components across the 60 electrodes in the source reconstruction analysis mentioned earlier. The kernel was multiplied by the analytical signals of the electrode space to obtain the analytical signals of the source space (Fig. 1a). Similar to the electrode space, the analytical signals of the source space were used to calculate the FC within each of the five frequency bands for each participant. To reduce the computational burden, the edges of the vertices within each brain node were averaged; thus, the final FC matrices with a size of 100 × 100 (excluding the main diagonal) were determined (Fig. 1b). We also performed a CPM analysis on the FC matrices of all frequency bands, and five models were obtained.

Results

PSD analysis in the electrode space

Table 1 presents the results of the PSD analysis of the electrode space. Significant negative correlations were found between loneliness scores and beta-band powers across the frontal, central, and posterior midline ROIs (P < .019). Marginal significant correlations between loneliness scores and beta-band powers were found in the posterior left and right ROIs (P < .097). Gamma-band powers recorded marginally significant negative correlations with loneliness scores in the central right ROI (P = .094). However, no significant correlation was found between the other frequency bands and the ROIs (P > .103).

Table 1.

The partial correlation coefficients of the loneliness scores and powers of the five frequency bands across nine ROIs in the electrode space.

		Delta	Theta	Alpha	Beta	Gamma
Frontal	Left	−0.078	−0.157	−0.157	−0.294**	−0.147
	Midline	−0.124	−0.134	−0.167	−0.281*	−0.118
	Right	−0.083	−0.173	−0.170	−0.320**	−0.132
Central	Left	−0.109	−0.176	−0.171	−0.291**	−0.176
	Midline	−0.100	−0.132	−0.142	−0.257*	−0.083
	Right	−0.194	−0.185	−0.171	−0.314**	−0.203†
Posterior	Left	−0.069	−0.133	−0.127	−0.225†	−0.098
	Midline	−0.087	−0.141	−0.114	−0.237*	−0.176
	Right	−0.057	−0.133	−0.125	−0.197†	−0.036

The effects of sex and age differences were excluded. Significant or marginally significant correlation coefficients that survived correction for multiple comparisons are shown in bold.

^† P < .1.

^* P < .05.

^** P < .01.

PSD analysis in the source space

Table 2 presents the results of PSD analysis in the source space. A significant negative correlation between loneliness scores and beta-band powers across the control (A/B), default (A/B/C), dorsal attention (B), limbic (A/B), salience/ventral attention (A/B), and somatomotor (A/B) networks was found (P < .037). A marginally significant negative correlation between loneliness scores and beta-band powers in the dorsal attention (A), temporo-parietal, and visual (central/peripheral) networks was found (P < .094). For the other four frequency bands, we found marginally significant negative correlations between loneliness scores and powers of theta and alpha bands in the limbic A network (P < .094). No other significant correlations were observed (P > .103).

Table 2.

The partial correlation coefficients of the loneliness scores and powers of the five frequency bands across 17 networks in source space.

	Delta	Theta	Alpha	Beta	Gamma
Control A	−0.095	−0.164	−0.158	−0.295**	−0.125
Control B	−0.108	−0.172	−0.158	−0.300**	−0.104
Control C	−0.071	−0.141	−0.132	−0.262*	−0.135
DN A	−0.097	−0.139	−0.137	−0.274*	−0.122
DN B	−0.093	−0.155	−0.165	−0.307**	−0.141
DN C	−0.109	−0.158	−0.150	−0.238*	−0.013
DAN A	−0.073	−0.144	−0.123	−0.213†	−0.020
DAN B	−0.117	−0.159	−0.160	−0.274*	−0.146
LB A	−0.145	−0.201†	−0.199†	−0.292**	−0.079
LB B	−0.039	−0.133	−0.182	−0.304**	−0.106
SAL/VAN A	−0.112	−0.156	−0.169	−0.282*	−0.071
SAL/VAN B	−0.114	−0.153	−0.175	−0.280*	−0.129
SM A	−0.141	−0.172	−0.166	−0.294**	−0.156
SM B	−0.132	−0.194	−0.158	−0.270*	−0.065
TP	−0.084	−0.157	−0.138	−0.217†	0.026
VIS Central	−0.101	−0.161	−0.149	−0.202†	−0.011
VIS Peripheral	−0.076	−0.143	−0.147	−0.226†	−0.093

The effects of sex and age differences were excluded.

^† p < 0.1,

^* p <0.05,

^** p <0.01.

Significant or marginally significant correlation coefficients that survived correction for multiple comparisons are shown in bold. Abbreviations: DN = default network; DAN = dorsal attention network; LB = limbic network; SAL/VAN = salience/ventral attention network; SM = somatomotor network; TP = temporo-parietal network; VIS = visual network.

FC analysis in the electrode space

Figure 3a shows the correlations between actual loneliness scores and predicted scores based on FC for the five frequency bands in the electrode space and the permutation test results. No significant or marginally significant (P > .419) correlations were found between actual loneliness scores and predicted scores of the models for all five frequency bands.

Figure 3.

Performance of the prediction model based on FC for the five frequency bands in the electrode and source spaces. (a) Spearman’s rank correlations between the actual and predicted loneliness scores for the five frequency bands in the electrode space, and the null distributions of correlation coefficient obtained by permutation test with 1000 iterations. The dashed line indicates the correlation coefficients obtained by the actual prediction. (b) Spearman’s rank correlations between the actual and predicted loneliness scores for the five frequency bands in the source space, and the null distributions of correlation coefficient obtained by permutation test with 1000 iterations. The dashed line indicates the correlation coefficients obtained by the actual prediction. (c) The 227 edges of nodes showed significant positive correlations with the loneliness scores in every iteration, and the total number of these edges within and between 17 networks. †P < .1.

FC analysis in the source space

Figure 3b shows the correlations between actual loneliness scores and predicted scores based on FC for the five frequency bands in the source space and the permutation test results. The alpha-band model shows a trend, but not statistically significant (P = .094), in predicting individual loneliness scores. Across the 139 iterations of LOOCV, the number of edges between each pair of nodes selected as features to build the model ranged from 324 to 819. Notably, all these edges showed significant positive correlations with the loneliness score, and 227 of them appeared in every iteration. Figure 3c shows these 227 edges and the number distribution between and within 17 networks. The highest total number of these edges was shown in four pairs of inter-network connections and one intra-network connection: somatomotor (B)-control (A), somatomotor (B)-dorsal attention (A), somatomotor (B)-salience/ventral attention (A), default (B)-salience/ventral attention (A), and somatomotor (B)-somatomotor (B). Therefore, these connections were the most important for the model to predict the individual loneliness scores. No significant or marginally significant correlations between the predicted and actual loneliness scores were found in the other four models, including delta, theta, beta, and gamma bands.

Discussion

This exploratory study measured resting-state EEG to systematically examine the patterns of whole-brain spectral density (PSD) and FC in five frequency bands among lonely individuals. The study found a correlation between loneliness and decreased beta-band power in the whole-brain region, as observed in both the electrode and source spaces. Additionally, there was a trend, although not statistically significant, indicating that loneliness scores may potentially associated to alpha-band FC patterns in the source space, specifically those from somatomotor, control, default, dorsal attention, and salience/ventral attention networks. These findings provide systematic insights into the neural mechanisms of loneliness based on resting-state EEG. They also provide new information to identify high-risk populations and prevent multiple mental disorders.

The relation between loneliness and beta-band power reaches significance

Significant negative correlations were found between loneliness scores and beta-band powers in both the electrode and source spaces across whole-brain regions, especially in the frontal–parietal brain regions. Negative mood, attention, reward system, and/or sensorimotor processes are speculated to contribute to these correlations. Negative mood may contribute to the correlation between loneliness and decreased beta-band power. Negative stimuli can lead to a decrease in beta-band power (Huebl et al. 2016, Merkl and Sterner 2016, Roxburgh et al. 2023, Chen et al. 2024). For instance, Chen et al. (2024) reported a significant negative correlation between beta-band power and depression scores in preclinical participants, indicating that decreased beta-band power was sensitive to depressed mood. Similarly, Roxburgh et al. (2023) found that a sustained reduction in beta-band activity was associated with anxiety induced by threat. Du et al. (2023) and Jiang et al. (2019) suggest that decreased beta-band power reflects a disruption of multimodal integration for emotional processing. Loneliness is a negative emotional state (Hawkley 2022, Hayes et al. 2022, Piejka et al. 2023). The decrease in beta-band power associated with loneliness may reflect a disruption of emotional processing (Finley and Schaefer 2022, Wong et al. 2022).

Attention-related processing may also contribute to the significant correlations between loneliness and beta-band power. Beta-band activity in the frontal and parietal regions is associated with attentional control processing (Engel and Fries 2010, Gola et al. 2013, Li et al. 2017). Loneliness has been found to correlate with decreased beta activity, which may be associated with increased stimulus-driven, bottom-up attentional processing (Engel and Fries 2010, de Carvalho et al. 2015, Rosero Pahi et al. 2020, Palacios-García et al. 2021). Lonely individuals tend to be hypervigilant to socially threatening information (Vanhalst et al. 2015, Cacioppo et al. 2016, Spithoven et al. 2017, Smith et al. 2022). This may cause them to pay more attention to their surroundings to detect threatening information and to use enhanced cognitive resources to decode such information. Consequently, increased bottom-up attention may contribute to the correlation between loneliness and decreased beta-band power. Alternatively, defocused attention may contribute to the correlation between loneliness and beta-band power. Defocused attention is an open, unfocused, unselective, and effortless mode of attention that allows irrelevant information to be perceived and processed (Von Hecker and Meiser 2005, Von Hecker et al. 2013). According to Chen et al. (2024), decreased beta-band power was closely related to defocused attention during depressive mood. Similar to a depressive mood, the decrease in beta-band power during loneliness is related to defocused attention, indicating a unique adaptive value that facilitates individuals to move away from negative outcomes (Von Hecker and Meiser 2005, Von Hecker et al. 2013, Chen et al. 2024).

In addition, the reward process may contribute to the significant correlations between loneliness and beta-band power. Studies have found that increased beta-band power is sensitive to positive feedback (HajiHosseini et al. 2012, Mas-Herrero et al. 2015, Alicart et al. 2020, Meyer et al. 2021). Lonely individuals show attenuated sensitivity to positive feedback, exhibiting diminished responsiveness to positive social interactions (Lieberz et al. 2021), and hyposensitivity to social inclusion (Vanhalst et al. 2015), and view positive social interactions as less rewarding (Qualter et al. 2015). Thus, the reduced beta-band power in lonely individuals may reflect the attenuated sensitivity to positive rewards.

Finally, sensorimotor processes may contribute to the significant correlations between loneliness and beta-band power. Beta oscillations are related to sensorimotor functions, and decreased beta-band power suggests increased activity in the somatosensory system (Engel and Fries 2010, Kilavik et al. 2013, Chota et al. 2023). According to Roxburgh et al. (2023), threat-related reductions in beta-band power may reflect increased readiness to act as a protective mechanism. Lonely individuals are hypervigilant to social threats (Vanhalst et al. 2015, Cacioppo et al. 2016, Spithoven et al. 2017) and regulate their emotions through active rejection or social withdrawal (Preece et al. 2021, Delgado et al. 2023), and they may increase action readiness to cope with potential threats as a protective mechanism, resulting in reduced beta-band power. Alternatively, reduced beta-band power in lonely individuals may reflect an enhanced sensory process to compensate for impaired perceptual discrimination, as Mąka et al. (2023) found that loneliness reduced efficiency in perceptual decision-making.

The relation between loneliness and alpha-band FC shows a trend but does not reach significance

There was a trend, although not statistically significant, indicating that loneliness scores were related to increased alpha-band FC patterns in the source space. Specifically, these patterns were observed in the somatomotor, control, default, dorsal attention, and salience/ventral attention networks. According to a systematic review by Miljevic et al. (2023), higher alpha FC in depression is associated with a lack of ability to regulate internal neural processes. Considering the close association between depression and loneliness (Grygiel et al. 2023), we speculate that the observed trend of increased alpha-band FC with higher loneliness scores may suggest a deficiency in the ability to regulate internal processes in lonely individuals, potentially due to their heightened external attention to the environment.

Strengths, limitations, and future directions

This study has several strengths. First, we are the first to systematically explore resting-state PSD and FC alterations related to loneliness. Second, our study provided unbiased and objective results than studies that rely on a priori knowledge of specific frequency bands or brain regions. Finally, the current results provide new insights into identifying high-risk populations to prevent other multiple mental disorders and developing neurobiologically informed interventions for loneliness. For example, beta-band activity can serve as an objective indicator to screen out lonely populations, who are at high risk of developing multiple mental illnesses. Loneliness can potentially be reduced by increasing brain beta activity using brain stimulation techniques, such as rhythmic sensory stimulation, repetitive transcranial magnetic stimulation, and transcranial alternating current stimulation (Herrmann et al. 2016). Virtual reality video is also a promising technique for enhancing beta activity, which was shown to significantly increase beta activity in the frontal lobe (Kweon et al. 2018).

However, this study has several limitations. First, this study is an exploratory investigation of resting-state EEG in relation to loneliness. The findings in this study are mostly explorative. We did not use any external sample to validate or replicate our findings. All the initial results are foundational before future confirmatory investigations could be proposed. Future studies are required to replicate these findings.

Second, we cannot conclude a deterministic causal direction of relationships among loneliness, cognitive–social–emotional processing, and neurobiological mechanisms. Our explanations of the correlation between loneliness and resting EEG signals are speculative and based mainly on studies that demonstrate how loneliness reveals cognitive–social–emotional biases (Morr et al. 2022, Delgado et al. 2023). In contrast, some studies have shown that cognitive–social–emotional processing can affect loneliness (Kearns and Creaven 2017, Hayes et al. 2022, Barjaková et al. 2023). In addition, other studies did not find the hypothesized cognitive, social, and emotional biases (D’Agostino et al. 2019, Lieberz et al. 2022, Mąka et al. 2023). The relation between loneliness, cognitive–social–emotional processing, and neurobiological mechanisms requires further investigation. For example, future studies could investigate whether increasing brain beta-band activity through brain stimulation (Herrmann et al. 2016) and virtual reality video techniques (Kweon et al. 2018) can reduce cognitive–social–emotional biases.

Third, this study focuses on trait-like, chronic loneliness, which can be measured using the University of California Los Angeles Loneliness Scale. Studies have shown that trait-like, chronic loneliness may have distinctive mechanisms compared to acute loneliness (Tomova et al. 2020, Saporta et al. 2021, Delgado et al. 2023). Future studies can explore the unique neurocognitive mechanisms between acute and chronic loneliness.

Fourth, a greater electrode density can improve the source localization accuracy (Song et al. 2015) although 64 electrodes provide acceptable spatial sampling for EEG source reconstruction. Furthermore, the use of template brain models, but not individual head models derived from participants’ MRIs, may reduce source localization accuracy. Therefore, future studies are needed for source reconstruction using high-density (128 or 256 electrodes) EEG data and participants’ head models.

Fifth, we only focused on the powers of the EEG signal, and the phases were unexplored. Neural dynamics measured through the phase can provide different information than those measured by power (Cohen 2014). Future studies are needed to evaluate brain rhythms based on phase.

Finally, FC analysis between the different frequency bands was not analyzed in our study, and our study only evaluated power-based connectivity. Future studies could evaluate phase-based connectivity and conduct cross-frequency coupling analysis.

Conclusion

This exploratory study is the first to use resting-state EEG to examine alterations in the dynamic pattern of brain activity among lonely individuals. We used a data-driven approach to analyze the PSD and FC in the electrode and source spaces. The results showed significant correlations between loneliness scores and decreased beta-band powers, and a trend of alpha-band FC is associated with individuals’ loneliness scores. These findings provide systematic insights into the neural mechanisms of loneliness and new information to identify high-risk populations and prevent multiple mental disorders.

Author contributions

Xin Hu (Data curation, Formal analysis, Visualization, Methodology, Software, Writing—original draft), Xufang Wang (Data curation, Data preprocessing), Changquan Long (Formal analysis, Methodology, Validation, Conceptualization, Resources, Supervision), and Xu Lei (Resources, Supervision).

Conflict of interest

None declared.

Funding

This work was supported by the Fundamental Research Funds for the Central Universities (SWU2209240), the National Natural Science Foundation of China (31971028), and the National Key Research and Development Program of China (2021YFC2501500).

Data availability

The datasets and scripts for the study are available at https://osf.io/xt3pk/.