Associations between insomnia symptoms and functional connectivity in the UK Biobank cohort (n = 29,423)

An increasing number of studies harness resting‐state fMRI functional connectivity analysis to investigate the neurobiological mechanisms of insomnia. The results to date are inconsistent and the detection of minor and widely distributed alterations in functional connectivity requires large sample sizes. The present study investigated associations between insomnia symptoms and resting‐state functional connectivity at the whole‐brain level in the largest sample to date. This cross‐sectional analysis used resting‐state imaging data from the UK Biobank, a large scale, population‐based biomedical database. The analysis included 29,423 participants (age: 63.1 ± 7.5 years, 54.3% female), comprising 9210 with frequent insomnia symptoms and 20,213 controls without. Linear models were adjusted for relevant clinical, imaging, and socio‐demographic variables. The Akaike information criterion was used for model selection. Multiple comparisons were corrected using the false discovery rate with a significance level of q < 0.05. Frequent insomnia symptoms were associated with increased connectivity within the default mode network and frontoparietal network, increased negative connectivity between the default mode network and the frontoparietal network, and decreased connectivity between the salience network and a node of the default mode network. Furthermore, frequent insomnia symptoms were associated with altered functional connectivity between nodes comprising sensory areas and the cerebellum. These functional alterations of brain networks may underlie dysfunctional affective and cognitive processing in insomnia and contribute to subjectively and objectively impaired sleep. However, it must be noted that the item that was used to assess frequent insomnia symptoms in this study did not assess all the characteristics of clinically diagnosed insomnia.


INTRODUCTION
Approximately 35% of the general population suffer from insomnia symptoms, whereas about 10% meet the criteria for insomnia as a disorder (Ohayon, 2002;Ohayon & Reynolds, 2009). Insomnia increases the risk of developing other mental disorders and somatic morbidities (Hertenstein et al., 2019;M. Li, Zhang, Hou, & Tang, 2014) and is associated with high costs for health-care systems (Daley, Morin, LeBlanc, Grégoire, & Savard, 2009;Thiart et al., 2016). Despite the socioeconomic and health impact of insomnia, its pathophysiology has not been fully identified. From a psychological perspective, dysfunctional affective and cognitive processes are considered core factors in the aetiology of insomnia (Schiel et al., 2020;Schmidt, Harvey, & Van der Linden, 2011). For example, worry and rumination at bedtime are thought to cause physiological arousal and, as a consequence, interfere with the onset and maintenance of sleep (Carney, Harris, Moss, & Edinger, 2010). Furthermore, impeded adaptation of emotional responses and heightened emotional reactivity to sleep-related stimuli are thought to cause emotional distress and contribute to disrupted sleep (Baglioni, Spiegelhalder, Lombardo, & Riemann, 2010;Van Someren, 2020). Still, little is known about the underlying neurobiological processes. In this regard, an increasing number of studies use functional imaging techniques to investigate the neural correlates and mechanisms that may contribute to the development and maintenance of insomnia . Several of these studies have used resting-state fMRI functional connectivity, a method that allows the analysis of temporal correlations between spatially distinct brain regions, shedding light on the brain's inherent organisation and functioning (Khazaie et al., 2017). Research on functional connectivity in insomnia has mainly focussed on the investigation of the default mode network (DMN), a large-scale brain network that is involved in self-generated thought (that is, internally directed and stimulus-independent), which is more active during a resting state compared with conditions requiring externally directed attention (Andrews-Hanna, Smallwood, & Spreng, 2014;Schiel et al., 2020). However, there are inconsistent findings, with studies reporting increased connectivity (Leerssen et al., 2019), unaltered connectivity (Regen et al., 2016), and reduced connectivity (Nie et al., 2015) between nodes of the DMN in insomnia patients compared with healthy good sleepers. In addition, studies have described altered connectivity in people with insomnia in the salience network (SN) and in the frontoparietal network (FPN), largescale brain networks involved in cognitive and affective processing (Chen, Chang, Glover, & Gotlib, 2014;S. Li et al., 2018;Wei et al., 2020).
Whilst recent research contributes knowledge on atypical connectivity within and between resting-state brain networks, inconsistencies remain, which may be attributed to heterogeneous sampling, methodological differences, and/or small sample sizes Tahmasian et al., 2018). Moreover, it is possible that insomnia is characterised by minor and widely distributed connectivity alterations (Van Someren, 2020). However, whole-brain functional connectivity analyses with sufficient power to detect such small effects would require large sample sizes. To address these gaps, we carried out a study to investigate the association between wholebrain resting-state functional connectivity and insomnia symptoms in the UK Biobanka cohort that is 100-fold larger than any previous study. Given the inconsistencies in previous research, we tested the non-directional hypothesis that functional connectivity is independently associated with insomnia symptoms.

Participants
The UK Biobank is a prospective epidemiological study with over 500,000 participants aged 40-69 years at baseline. Data collection includes questionnaires, physical and cognitive measures, and biological samples (Sudlow et al., 2015). In 2016, data collection was extended to multimodal brain imaging and was scheduled to be completed in 2022 with scans of 100,000 subjects (Miller et al., 2016).
The present study used data from the first imaging visit, which was available for 37,848 subjects at the time of data analysis. Subjects were excluded if they self-reported a neurological condition (n = 897), sleep apnea syndrome (n = 166), or excessive daytime sleepiness (n = 0), or had missing data for any variable or covariate that was included in the analysis (n = 7362). This resulted in a sample size of 29,423 for the current analysis. All participants gave written informed consent, and all procedures were approved by the NHS National Research Ethics Service (Ref. 11/NW/0382). The present analysis was conducted as part of UK Biobank application 6818 and has been preregistered on the Open Science Framework (https://osf.io/8drce).

Imaging data
The present study made use of subject-specific network matrices generated by an image-processing pipeline developed and run on behalf of UK Biobank (Alfaro-Almagro et al., 2018). Image acquisition protocols and processing pipelines have been described with greater detail in a previous overview article (Miller et al., 2016). Comprehensive documentation is available online at https://biobank.ctsu.ox.ac.uk/ crystal/crystal/docs/brain_mri.pdf. During image acquisition, the participants were instructed to keep their eyes fixated on a crosshair, relax and think of nothing in particular. Pre-processing of raw data included motion correction, grand-mean intensity normalisation, highpass temporal filtering, echo-planar image unwarping, gradient distortion correction unwarping, and removal of structured artefacts by FMRIB Software Library's ICA-based X-noiseifier. Group independent component analysis (ICA) was performed using FSL Packages with 4100 datasets at two different dimensionalities (100 and 25). Group-ICA components that were not neuronally driven (i.e. artefacts) were identified and removed by experts of the UK Biobank team, resulting in neuronally driven sets of 55 and 21 components, respectively. Subsequently, ICA spatial maps were mapped onto each participant's fMRI timeseries to receive one representative node timeseries per ICA component for each subject, which were then used to estimate subject-specific network matrices. In this context, the term "node" describes the spatial map of an ICA component that is typically spatially non-contiguous (i.e. one node can encompass multiple spatially distinct brain areas across the whole brain). Network modelling was carried out using the FSLNets toolbox for both full and partial temporal correlation coefficients between each pair of node timeseries.
Pearson correlation scores were transformed into z-statistics. The partial correlation network matrices with a resolution of 55 Â 55 were used in the present analysis, because when the number of estimated components is higher, these are more likely to represent smaller regions or network nodes (Miller et al., 2016), and the majority of previous studies on functional connectivity in insomnia have reported functional connectivity differences within resting-state networks (Schiel et al., 2020). In addition, given the hierarchical nature of brain areas that can be explained at different levels of complexity, the examination of functional connectivity between nodes encompassing core parts of resting-state networks also provides information on between-network differences. The advantage of using partial correlation lies in the better estimation of direct connection strengths com-

Insomnia symptoms
Self-reported insomnia symptoms served as a predictor variable in the current analyses and were assessed by means of the question "Do you have trouble falling asleep at night or do you wake up in the middle of the night?". As such, our definition provides no information on daytime functioning, sleep opportunity, or duration of sleep complaints, which would be mandatory for diagnosing insomnia disorder.
Participants that responded "usually" were categorised as subjects with frequent insomnia symptoms, participants that responded "never/rarely" or "sometimes" were categorised as control subjects without frequent insomnia symptoms.

Covariates
Sleep duration was assessed by asking "About how many hours sleep do you get in every 24 hours? (please include naps) ". Participants were categorised as short sleepers (<7 h), normal sleepers (7-9 h), and long sleepers (>9 h) based on guidelines (Watson et al., 2015). Daytime sleepiness was assessed by means of the question "How likely are you to doze off or fall asleep during the daytime when you do not mean to? (e.g. when working, reading or driving) ". Participants were categorised as sleepy if they answered "sometimes" or "often" and as non-sleepy if they answered "never/rarely". Participants that answered "all of the time" were excluded because of excessive daytime sleepiness. Chronotype was assessed by means of asking "Do you consider yourself to be…?" with responses "definitely a 'morning' person", "more a 'morning' than an 'evening' person", "more an 'evening' than a 'morning' person", and "definitely an 'evening' person". Participants were categorised into three groups: early chronotype, late chronotype, and intermediate chronotype, which was obtained by merging the two middle responses. Sleep medication and psychotropic medication use were dichotomised into "sleep medication use" versus "no sleep medication use" and "psychotropic medication use" versus "no psychotropic medication use" (for a list of included medications, see Supplementary Tables S1 and S2). Socioeconomic status was assessed by the Townsend index, a measure of material deprivation, which was log-transformed with the equation "ln (x + 7) " due to skewed distribution. Educational qualifications were recorded, and were dichotomised into the categories "college/university degree" versus "no college/university degree". Depressive symptoms were assessed by means of the question "Over the past 2 weeks, how often have you felt down, depressed or hopeless?". Participants were assigned to the category "depressive symptoms" if they responded "several days", "more than half the days", or "nearly every day", whereas participants that responded "not at all" were assigned to the category "no depressive symptoms". Further covariates that were included in the analysis comprised intracranial volume (ICV; grey matter volume, white matter volume and volume of cerebrospinal fluid), average resting-state fMRI head motion across space and time points, age, sex, and body mass index (BMI). Selection and categorisation of non-imaging related variables is consistent with our previous analysis of cognitive performance data in the UK Biobank (Kyle et al., 2017).

Statistical analysis
Descriptive data are presented as mean values with standard deviations and the percentage of the sample reporting specific response options.
Associations between insomnia symptoms and functional connectivity were tested by three different linear models for each brain connection.
The term "connection" describes a functional relationship between a pair of nodes. The strength and direction of this relationship is indicated by a partial correlation coefficient (ranging between À1 and + 1). Partial correlation coefficients of all pairs of nodes derived from connectivity matrices were set as dependent variable and insomnia symptoms were set as independent variable in all models. Model specific assumptions of linearity, residual homoscedasticity, and residual normality were visually checked using scatter plots and normal probability plots.
In a first step (model 1), associations between insomnia symptoms and functional connectivity were analysed in a linear model adjusting for age, sex, ICV, average head motion, and BMI. In a second step (model 2) self-reported level of education, socioeconomic status, sleep medication, depressive symptoms, and psychotropic medication were included as covariates in addition to those described in model 1. In a final step (model 3) sleep duration, daytime sleepiness, and chronotype were included as further covariates in addition to those described in model 2. Group mean functional connectivity values were assessed to interpret the direction of significant associations. The Akaike information criterion (AIC) was used for model selection. The AIC measures the utility of a model and is derived from a model's maximum likelihood estimate corrected by model complexity (Akaike, 1974). The model with the lowest AIC was selected for each connection. The false discovery rate was used to correct for multiple comparisons with a significance level of q < 0.05 (Benjamini & Hochberg, 1995). All analyses were conducted using R version 3.6.3.

RESULTS
The sample comprised 15,968 females and 13,455 males with a mean age of 63.1 years (SD = 7.5 years). Frequent insomnia symptoms were reported by 9210 participants, whilst 20,213 participants were classified as controls. Socio-economic and clinical characteristics for the study sample are presented in Table 1. Compared with participants without frequent insomnia symptoms, those with frequent insomnia symptoms were older (p = 0.04), more likely to be female (p < 0.001), had a higher BMI (p = 0.008), reported shorter sleep duration (p < 0.001), were more likely to report daytime sleepiness (p < 0.001), were more likely to have early or late chronotype (p = 0.001), were less likely to hold a university or college degree (p = 0.038), were from a similar socioeconomic background (p = 0.820), were more likely to report taking sleep medication (p < 0.001), were more likely to report recent depressive symptoms (p < 0.001), and were more likely to report using psychotropic medication (p < 0.001).
In total, the analysis revealed 14 connections between 23 nodes significantly associated with insomnia symptoms across all three models. Overall, taking into account model fit and complexity, there was evidence for including sleep duration, daytime sleepiness, and chronotype as covariates in addition to age, sex, ICV, average head motion, BMI, self-reported level of education, socioeconomic status, sleep medication, depressive symptoms, and psychotropic medication (see Supplementary Tables S3 and S4). For three connections that only remained significant when including all covariates, the best-fit model only included age, sex, ICV, average head motion, and BMI.
Hence, these connections were disregarded, leaving 11 connections between 18 nodes significantly associated with insomnia symptoms (all p corrected < 0.05, all p uncorrected < 3.43 Â 10 À4 , see Figure 1). Therefore, the results presented in the following paragraphs were obtained using the final model with all covariates included (model 3).

DISCUSSION
The present study investigated associations between insomnia symptoms and whole-brain functional connectivity in a large population-   Li et al., 2017;Regen et al., 2016). Taken together, although speculatively, it is conceivable that increased connectivity within the DMN may be associated with dysfunctional forms of cognitive control in insomnia (Carney et al., 2010;Schmidt et al., 2011). Of note, our findings contrast with those of sleep deprivation studies, which describe decreased within-network connectivity of the DMN in the sleep-deprived state relative to the well-rested state (Chee & Zhou, 2019). These observations are in line with the notion that the pathophysiology of insomnia is characterised by cognitive-affective factors rather than by sleep loss (Van Someren, 2020). Interestingly, increased connectivity within the DMN is consistently found in depression (Mulders, van Eijndhoven, Schene, Beckmann, & Tendolkar, 2015); and the dorsal medial subsystem is activated during rumination in depressed patients (Zhou et al., 2020).
Furthermore, increased connectivity within the DMN is associated with lower levels of happiness and higher levels of rumination even in healthy adults (Luo, Kong, Qi, You, & Huang, 2016). Given that insomnia predicts the onset of depression (Baglioni et al., 2011;Hertenstein et al., 2019), and sleep problems are a common symptom of major depressive disorder (Soehner, Kaplan, & Harvey, 2014), increased within-network connectivity involving the dorsal medial subsystem and the DMN as a whole may indicate a common mechanism partially underlying the strong correlation between depression and insomnia.
Notably, the present analysis adjusted for differences in depressive symptoms. Still, the question remains whether connectivity changes emerge as a consequence of insomnia and/or constitute a F I G U R E 2 Associations between insomnia symptoms and functional connectivity between 55 nodes. Top left (above the diagonal): ß values indicate group differences between those with frequent insomnia symptoms and those without frequent insomnia symptoms. Larger squares correspond to higher absolute ß values (ranging from ß < 0.01 to ß ≥ 0.07). Red, increased positive connectivity in those with frequent insomnia symptoms compared with those without frequent insomnia symptoms; orange, decreased positive connectivity; blue, increased negative connectivity; cyan, decreased negative connectivity. Bottom right (below the diagonal): Corresponding p-values after false discovery rate correction. Larger squares correspond to smaller p-values (ranging from p corrected < 0.001 to p corrected ≥ 0.1). Green squares indicate significance at the 5% level vulnerability to psychopathology. In the case of depression, variability in DMN activity has been linked to both trait and state rumination (Hamilton et al., 2011;Zhou et al., 2020). Hence, it may be possible that DMN dysfunction underlying rumination constitutes both a predisposing and perpetuating factor to insomnia. Future studies with longitudinal designs would be able to examine whether pre-existing variability in DMN connectivity contributes to the risk of developing insomnia. In addition, imaging studies using therapeutic interventions could investigate whether altered DMN connectivity in people with insomnia is reversible and resembles rather state-like properties.
There was increased connectivity within the FPN, which is involved in a set of cognitive processes, including control of attention, executive function, and goal-directed cognition (Cole et al., 2013).
Although insomnia is associated with subjective cognitive complaints there is limited evidence for consistent objective cognitive impairment (Shekleton, Rogers, & Rajaratnam, 2010 Insomnia symptoms were also associated with altered betweennetwork connectivity. Participants who reported frequent insomnia symptoms showed increased negative connectivity between nodes of the FPN and DMN. The FPN has a modulatory role in the advancement and suppression of DMN activity, dependent on the need to produce an internal train of thought (Smallwood, Brown, Baird, & Schooler, 2012). For example, increased positive connectivity between the DMN and FPN can be observed during autobiographical planning tasks (Spreng, Stevens, Chamberlain, Gilmore, & Schacter, 2010). One interpretation of this result may be that increased negative as opposed to positive connectivity between the FPN and the DMN disrupts goal-directed thought and, in turn, facilitates dysfunctional forms of self-generated thought present in insomnia. Participants with frequent insomnia symptoms also showed decreased connectivity between the SN and a region of the DMN, which is in line with the findings of another study (C. Li et al., 2017).
Amongst other functions, the SN supports the coordination of brain network dynamics and is involved in the suppression of DMN activity (Menon & Uddin, 2010;Sridharan, Levitin, & Menon, 2008).
Decreased connectivity between the SN and DMN may result in diminished DMN deactivation and contribute to the earlier reported result of increased within-network connectivity of the DMN.
In comparison with controls, participants with frequent insomnia symptoms showed increased connectivity between the sensorimotor cortex and the SN, a network that plays a central role in the detection of salient stimuli through the integration of sensory, emotional, and cognitive information (Uddin, 2015). This connectivity pattern has recently been interpreted as a state of elevated "threat assessment" of bodily signals (Hegarty, Yani, Albishi, Michener, & Kutch, 2020).
Psychological models of insomnia assume that arousal and distress trigger selective attention towards and increase monitoring of internal sensations and external sleep-related stimuli (A. G. Harvey, 2002;Riemann et al., 2010). As a consequence, bodily information (e.g. fast heart rate or muscle tension) and environmental cues (e.g. noises outside the house) inconsistent with falling asleep are perceived as sleeprelated threats, leading to an underestimation of the amount of nighttime sleep and to an overestimation of the extent of daytime impairment (Harris et al., 2015;Harvey, 2002). Increased connectivity between the SN and the sensorimotor cortex, although speculatively, may underlie this aberrant detection and evaluation of internal and external stimuli that fuel sleep-related dysfunctional cognitive activity.
Insomnia symptoms were also related to decreased negative connectivity between sensory areas, mainly those involved in the processing of visual information, as well as between sensory visual areas and the dorsolateral prefrontal cortex. Aberrant connectivity between sensory regions may promote abnormal levels of sensory processing, which may increase an individual's susceptibility to perturbation by external or internal stimuli and interfere with sleep initiation and maintenance Riemann et al., 2015). Furthermore, altered sensory processing during sleep may also contribute to the phenomenon of experienced wakefulness during polysomnography measured sleep that is described by many people with insomnia (Harvey & Tang, 2012).
Lastly, there was decreased connectivity within the cerebellum in association with insomnia symptoms, which is in line with a previous study that found reduced connectivity in the anterior lobe of the cerebellum (Huang et al., 2017), and with a recent meta-analysis that reported decreased activity in the right cerebellum in people with insomnia (Jiang, He, Guo, & Gao, 2020). Moreover, these functional alterations may be accompanied by structural cerebellar changes (Joo et al., 2013). Although the exact relationship between cerebellar processing and sleep is not fully understood, the cerebellum is thought to fine-tune neocortical forms of sleep-related activity, and cerebellar malfunction has been linked to the formation of altered sleep-wake cycles and various sleep disorders (Canto, Onuki, Bruinsma, van der Werf, & De Zeeuw, 2017). Furthermore, genome-wide association studies report enrichment of insomnia-symptom-associated genes expressed in multiple brain regions, including the cerebellum, frontal cortex, and anterior cingulate cortex, highlighting the role of these regions in the pathophysiology of insomnia (Lane et al., 2019).
Remarkably, the majority of connections were significantly associated with insomnia symptoms in the final model. Adjusting for sleep duration, daytime sleepiness and chronotype, although speculatively, may have created a more psychophysiological or stress-related phenotype, which is less characterised by sleep loss and sleepiness.
Indeed, people with insomnia commonly experience fatigue, i.e. a lack of energy, but do not differ highly from controls in terms of sleepiness and sleep duration Blanken et al., 2019).
Our study has several limitations. First, the UK Biobank sample consisted mainly of individuals in middle and older age. In addition, a low response rate during recruitment (5.5%) may have led to selection bias (Swanson, 2012). Second, our study was cross-sectional in nature and can only imply association but not causation. Third, due to the use of a single self-report item for assessing insomnia, we could not investigate associations between functional connectivity and insomnia subtypes. Fourth, it cannot be excluded that effects were also driven by sleep disorders that were not fully accounted for by exclusion of cases or model adjustment. Fifth, there was no information available on the time of day of image acquisition, no measurement of wakefulness during the scan, and no assessment of alertness, fatigue, or sleepiness immediately prior to or after the scan. As participants with frequent insomnia symptoms were more likely to report daytime sleepiness, it is conceivable that they were more prone to falling asleep during the scan compared with those without frequent insomnia symptoms, which could have influenced observed differences in functional connectivity.
Sixth, as the subset that was used for generating the group spatial maps did not consist equally of subjects with and without insomnia symptoms, our analysis might have been less sensitive to detect differences in functional connectivity between both groups. Lastly, the use of a single self-report item for assessing insomnia also raises the question whether this method correctly differentiated between participants with insomnia and those without. However, the predictive utility of this question was validated with a similar item in an independent sample from the Netherlands Sleep Registry; revealing a sensitivity of 98%, and a specificity of 96% in discriminating questionnaire-defined insomnia disorder cases from controls (Hammerschlag et al., 2017). Furthermore, the respective item correctly classified 91% of all cases with the diagnosis of insomnia disorder determined by a structured interview (Hammerschlag et al., 2017). Nevertheless, especially considering the older age of our sample, it is conceivable that participants with agerelated longer sleep onset latency and/or early awakenings do not perceive this as troublesome and/or are not affected by deficits in daytime functioning. Furthermore, it is possible that sleep problems in the current sample are reflected by a lack of sleep opportunity in many participants, which we did not control for.
In summary, we found multiple associations between insomnia symptoms and functional connectivity comprising altered within-and between-network connectivity involving nodes of the DMN, FPN, and SN in a large population-based sample. These findings support recent recommendations in the insomnia literature to study cognitive and affective processes and their underlying brain mechanisms, and strengthen the notion of insomnia as a disorder of cognitive-affective dysregulation (Schiel et al., 2020;Van Someren, 2020). Future studies are needed to understand the functional significance of these network differences and their relationship to cognitive-affective processes; whilst longitudinal designs and intervention studies will help investigate causal ordering.

AUTHOR CONTRIBUTORS
FH, KS, and SDK conceived and designed the analysis plan.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study were obtained from UK Biobank upon application, and legal constraints prohibit public