Introduction
Depression is the most common form of psychiatric pathology and affects approximately 350 million people and is one of the most common reasons for disability-adjusted life years worldwide (DALYs & Collaborators, 2018). Depression leads to several unfavorable medical and social consequences and decreases quality of life and adaptive ability (Kendler, Hettema, Butera, Gardner, & Prescott, Reference Kendler, Hettema, Butera, Gardner and Prescott2003; Mathers & Loncar, Reference Mathers and Loncar2006). Depressive symptoms include depressed mood, loss of pleasure, and loss of interest in most daily activities (DSM-5). Depression is also associated with sleep disturbances (Gebara et al., Reference Gebara, Kasckow, Smagula, DiNapoli, Karp, Lenze and Reynolds2018a; Staner, Reference Staner2010). Disturbed sleep is one of the diagnostic criteria for depression in multiple screening instruments. Compared with those without depression, individuals with depression have abnormal sleep characteristics on objective and subjective assessments, including delayed sleep onset, decreased sleep efficiency, more frequent arousals, and increased or decreased total sleeping time (Wiebe, Cassoff, & Gruber, Reference Wiebe, Cassoff and Gruber2012). Furthermore, there is evidence that insomnia-specific interventions, such as cognitive behavioral therapy for insomnia, may also lead to improvements in depression (Gebara et al., Reference Gebara, Siripong, DiNapoli, Maree, Germain, Reynolds and Karp2018b).
Sleep disturbances in depressed patients have been studied using polysomnography (PSG). Researchers have already for some time focused on changes in sleep architecture as a potential neurophysiological substrate of depression (Kupfer, Reference Kupfer1976). A large body of evidence suggests that adults with depression show clear reductions in slow wave sleep (SWS) and increased daytime dysfunction (Armitage, Reference Armitage2007; Baglioni et al., Reference Baglioni, Regen, Teghen, Spiegelhalder, Feige, Nissen and Riemann2013; Wiebe et al., Reference Wiebe, Cassoff and Gruber2012). Young adults with clinical depression also have lower levels of slow wave activity compared with healthy controls, particularly in the beginning of the night (Cheng, Goldschmied, Deldin, Hoffmann, & Armitage, Reference Cheng, Goldschmied, Deldin, Hoffmann and Armitage2015). There is also evidence that children and adolescents with internalizing disorders (Holm et al., Reference Holm, Forbes, Ryan, Phillips, Tarr and Dahl2009) and depressed adolescents (Rao et al., Reference Rao, Dahl, Ryan, Birmaher, Williamson, Rao and Kaufman2002) have significant reductions in SWS.
Another finding related to clinical depression is altered rapid eye movement (REM) sleep. Adolescents and adults with internalizing disorders, including depression, show decreased REM latency (the time between sleep onset and first REM period), increased REM length and density (Riemann, Berger, & Voderholzer, Reference Riemann, Berger and Voderholzer2001), and more recently, higher REM sleep discontinuity (Pesonen et al., Reference Pesonen, Gradisar, Kuula, Short, Merikanto, Tark and Lahti2018). A higher number of REM arousals is associated with insomnia, depression, and anxiety (Baglioni et al., Reference Baglioni, Nanovska, Regen, Spiegelhalder, Feige, Nissen and Riemann2016) and is suggested to hinder the overnight resolution of emotional distress (Wassing et al., Reference Wassing, Benjamins, Dekker, Moens, Spiegelhalder, Feige and Van Someren2016) in adults. REM sleep instability may promote chronic hyperarousal and dysfunction in emotional neural networks, increasing the risk for internalizing disorders.
In addition to analyzing differences in various individual sleep stages using conventional sleep parameters, some recent studies have focused on the overall dynamics of sleep stage architecture. For instance, Wei et al. (Reference Wei, Colombo, Ramautar, Blanken, van der Werf, Spiegelhalder and Van Someren2017) used Markov chain analyses to demonstrate that sleep stage dynamics derived from whole-night PSG characterizes insomnia better and independently of any conventional sleep parameter. Yetton, McDevitt, Cellini, Shelton, and Mednick (Reference Yetton, McDevitt, Cellini, Shelton and Mednick2018) further developed the analytic method by examining the static (e.g. minutes in stage, sleep efficiency) and dynamic measures of sleep architecture (e.g. transition probabilities and stage duration distributions) using a large dataset of 3202 nights from a non-clinical population. They observed that age, gender, time of day, and sleep time were associated with dynamic sleep architecture. Older participants and men had shorter bouts (more fragmentation) of Stage N2 and N3 (SWS) and transitioned less frequently to these stages.
In the current study, we examined whether sleep dynamics and architecture [defined as the cyclical pattern of sleep as it shifts between the different sleep stages, including non-rapid eye movement (NREM, stages N1, N2 and N3) and REM sleep], the amount of various sleep stages (specifically SWS and REM), as well as interrelations between different sleep stages are associated with depressive symptoms at a subclinical level in an adolescent population. We focused on investigating the associations and transition probabilities between the four sleep stages and studying the potential differences in those probabilities according to the level of depressive symptoms in adolescents. We hypothesized to find more disrupted sleep, i.e. more sleep stage transitions in depressed v. non-depressed individuals. However, in the absence of prior research with this methodological approach, the questions related to the specific dynamics of these transitions (from which to which stage) remain explorative.
Methods
Sample
Participants were derived from a community-based cohort including 1049 healthy singletons born between March and November 1998 in Helsinki, Finland (Strandberg, Jarvenpaa, Vanhanen, & McKeigue, Reference Strandberg, Jarvenpaa, Vanhanen and McKeigue2001). In years 2014–2015, all the cohort members who participated in the previous follow-up at age 12 (Pesonen et al., Reference Pesonen, Martikainen, Heinonen, Wehkalampi, Lahti, Kajantie and Raikkonen2014), and who lived within the 30 km radius from Helsinki (N = 279, 77.1% of the participants of the previous follow-up) were invited to a late adolescence follow-up (Kuula et al., Reference Kuula, Pesonen, Heinonen, Kajantie, Eriksson, Andersson and Raikkonen2018a; Merikanto et al., Reference Merikanto, Lahti, Kuula, Heinonen, Raikkonen, Andersson and Pesonen2018). Of them, 197 (70.6%) participated, and the analytic sample of the current study comprised 173 (mean age 16.9 years, s.d. = 0.3; 59% girls) adolescents who had complete records of an overnight sleep EEG measurement and depressive symptoms.
There were no reported diagnoses of schizophrenia or psychotic symptoms. The participants included in this study (n = 173) were more often girls than those in the original cohort (n = 873) (p-value for difference 0.004). There were no differences when compared with the original cohorts' age (Strandberg et al., Reference Strandberg, Jarvenpaa, Vanhanen and McKeigue2001), mother's body mass index, mother's age at the time of birth, gestational age, maternal alcohol consumption, length at birth, or birthweight (all p-values were not significant).
The Ethics Committees of the City of Helsinki Health Department and Children's Hospital in Helsinki University Central Hospital approved the study protocols. Informed written child and parent consents were obtained in early adolescence and only adolescent consent in late adolescence.
Polysomnography
Participants underwent overnight ambulatory PSG arranged according to the participants' schedules. Each participant completed a PSG measurement in their own homes, and in their own beds and they were asked to abstain from any stimulants, such as caffeine products after 4 p.m. in the measurement night. According to the study protocol, actigraphy measurements were initiated a week before the scheduled PSG night in order to investigate potential differences between the measurement night and other nights [see Kuula et al. (Reference Kuula, Pesonen, Merikanto, Gradisar, Lahti, Heinonen and Raikkonen2018b) for a detailed description of the actigraphy measurements]. All received monetary compensation (€50) for their effort. The measurements were performed throughout the entire school year excluding the summer holiday. Ninety percent of all PSG recordings were completed on school nights. All recordings were performed using SOMNOscreen plus (SOMNOmedics GmbH, Randersacker, Germany). A trained research nurse attached gold cup electrodes at six electroencephalogram locations [frontal (F) hemispheres, F3, F4; central (C) hemispheres, C3, C4; occipital (O) hemispheres, O1, O2] and at two locations for the mastoids (A1, A2). Overall frontal and central measurements were calculated as the means from both hemispheres. The electro-oculogram (EOG) and the electromyogram (EMG) were performed using disposable adhesive electrodes (Ambu Neuroline 715, Ambu A/S, Ballerup, Denmark), with two locations for the EOG and three locations for the EMG. In addition, an online reference Cz and a ground electrode in the middle of forehead were used. The sampling rate was 256 Hz (the hardware filters for SOMNOscreen plus are 0.2–35 Hz).
Three experienced researchers scored the PSG data manually using the DOMINO program (v2.7; SOMNOmedics GmbH, Randersacker, Germany) in 30 s epochs and categorized into N1, N2, N3, and REM according to AASM guidelines (The AASM Manual for the Scoring of Sleep and Associated Events). The relative amount of each stage was calculated based on total sleep time.
Depressive symptoms
Depressive symptoms were measured using the 21-item Beck's Depression Inventory II (BDI-II) (Beck, Steer, & Garbin, Reference Beck, Steer and Garbin1988) on the same night as the PSG measurement. In the BDI-II, each item is rated on a four-point scale (0–3) in terms of symptom intensity during the last 2 weeks, yielding a total score ranging from 0 to 63. In this study, internal consistency (Cronbach's α) was 0.92. To create a binary variable of depression risk and to correct for a potential gender and cultural bias, we used ⩾90th percentile score of BDI-II defined separately for boys and girls as a cutoff for moderate depression (Whisman & Richardson, Reference Whisman and Richardson2015).
Statistical analyses
Statistical analyses were conducted in the following order:
First, we calculated the total number of each 30 s epoch for each participant and tested the association between depression status and total duration of each epoch type using ANOVA. Shapiro–Wilk test results suggested that all sleep stage variables (W range from 0.51 to 0.98, all p-values <0.001) were normally distributed except N1 (p-value = 0.10). No sign of homoscedasticity in the association between sleep stages and moderate depression status (p-value range from 1.00 to 0.10) based on Breusch–Pagan test were found.
Second, we calculated the number and mean duration of each epoch type before each transition and tested the potential differences between depression status groups using ANOVA. We also investigated overall sleep v. wake bouts as consecutive epochs scored as stages N1, N2, N3, or REM and wake bouts as consecutive epochs scored as being at wake stage. We analyzed bout lengths of all individual sleep stages (staying in the same stage before the next, any other stage). For all five bout categories and being asleep v. wakefulness, survival analysis was used to estimate the probability for each observed duration in both depression status groups. We repeated the bout duration distribution survival analyses using the recurrent event multilevel (mixed-effects) survival analyses, taking into account the fact that multiple bouts were observed from each participant for both depression status groups. These analyses were performed to compare group-level sleep bout duration distributions with similar survival-analytical techniques as in previous studies on sleep–wake transition patterns (Lim et al., Reference Lim, Yu, Costa, Buchman, Bennett, Leurgans and Saper2011).
Third, we analyzed the independent associations (regularized partial correlations) between different epoch types in both depression status groups separately using a graphical Gaussian model (GGM) with the graphical least absolute shrinkage and selection operator (gLasso) algorithm. The gLasso eliminates spurious associations (edges) attributable to the influence of other nodes in the network and shrinks trivially small associations to zero (removes ‘false-positive’ edges) (Witten, Friedman, & Simon, Reference Witten, Friedman and Simon2011). In the final network, edges represent conditional independence relationships between nodes controlling for the effects of all other nodes (Epskamp, Borsboom, & Fried, Reference Epskamp, Borsboom and Fried2017). We set the value of the hyper-parameter γ to 0.2 to minimize the probability of false-positive edges (Epskamp et al., Reference Epskamp, Borsboom and Fried2017). The centrality of the individual stages within networks was calculated using the node strength centrality (number of connections weighted by the absolute strength of the connections). GGM illustrates the associations between various sleep bouts and whether these associations were different between depression status groups.
Fourth, we repeated the network analyses using relative importance regressions to detect the potential directions and direction differences between depression status groups. Fifth, we estimated using a Markov algorithm the transition probabilities between each state and these probabilities were further compared between depression status groups. A Markov algorithm defines a discrete sequence of states, this time sleep bouts, and thus models the probability of shifting from a specific bout to any other bouts to model the most probable sequence of various sleep bouts. Thus, we wanted to specifically analyze the differences in sequential sleep bout dynamics between those with minor depression and others.
We used the R package ‘qgraph’ (Epskamp, Cramer, Waldorp, Schmittmann, & Borsboom, Reference Epskamp, Cramer, Waldorp, Schmittmann and Borsboom2012) to plot and illustrate the final networks and the transition (probability) matrices. Solid lines represent positive and dashed lines represent negative associations (a thicker edge denotes a larger association between two stages). When illustrating transition probabilities, the solid lines represent similar and dashed lines represent significantly different transition probabilities between the depression status groups. Data analyses were performed with R (3.5.1) and a complete list of packages used is reported in the online Supplementary Material.
Results
Initial and sleep bout analyses
The sample consisted of 12% mild (⩾14 points) and 7% moderate (⩾20 points) depression scores. Of the participants, 21 were classified as having moderate depression. Table 1 presents descriptive characteristics of the participants according to depression status. Those with moderate depression smoked more often (p < 0.001) and reported lower physical activity (p = 0.049) than others. Most of the frequencies time asleep were spent in the N2 and N3 stages among all participants (Fig. 1). There were only small differences in the total times in each stage between depression status groups; none of the differences was statistically significant (p-value range from 0.28 to 0.87). There were no significant differences in the total duration of any sleep stage before the next stage between those with moderate depression compared to those without depression. Figure 2 shows the estimated bout survival functions for sleep, wakefulness, and individual sleep stages in those with moderate depression and others with associated 95% confidence intervals. All sleep bout survival curves took convex shapes. The wake bout survival curve took a concave shape. The group differences were all non-significant (p-value range from 0.21 to 0.95). Mixed survival analyses also revealed that none of the between-group differences was significant. The largest differences between the groups were in the N2 and N3 stages. Regarding actigraphy measurements of sleep duration or quality, the PSG night did not differ from other nights significantly (data not shown).
Fig. 1. Means and s.d.s of the various sleep stages in those with moderate depression and others.
Fig. 2. Survival curves of all sleep bouts in those with moderate depression and others.
Table 1. Sample characteristics (n = 173)
Sleep stage network analyses
GGM analyses suggested some differences in the network structure of the associations of the number of times spent in each sleep stage between those with moderate depression and those without (Fig. 3). The associations between N1 and N3 were significantly stronger in both directions of the association (interaction p-values 0.012 and 0.006) in those with moderate depression. Similarly, a stronger association was found from N1 to wake stage in those with moderate depression (p-value for interaction 0.002). These associations can also be detected in Fig. 4, which presents the relative importance networks in those with moderate depression and others. N1 was the most central in the network among those with moderate depression, whereas N2 had this position among others (online Supplementary Fig. S1).
Fig. 3. Gaussian graphical network of various sleep stages in those with moderate depression and others.
Fig. 4. Relative importance network of sleep stages in those with moderate depression and others.
Markov chain (transition probabilities) analyses within sleep stages
The transition probabilities between individual sleep stages in all participants are reported in online Supplementary Fig. S2. Most transitions are measured naturally within all stages and also within NREM stages (within N1, N2, and N3). The sleep stage transition probabilities in participants with moderate depression and those without depression are presented in Fig. 5 (significantly different transitions are marked with dashed lines). In participants with moderate depression, it was more probable to transit from N2 to N3 and from REM to N2 compared to those without depression (all tests presented in online Supplementary Fig. S3).
Fig. 5. Sleep stage transition probabilities (Markov chains) in those with moderate depression and others (statistically significantly different transitions in dashed lines).
Discussion
Depressive symptoms and sleep characteristics are closely associated (Pesonen et al., Reference Pesonen, Gradisar, Kuula, Short, Merikanto, Tark and Lahti2018; Riemann et al., Reference Riemann, Berger and Voderholzer2001). However, as most studies have focused on quantitative mean indicators of different sleep stages in relation to depression status, there is a lack of understanding on whether the dynamic organization of sleep stages (i.e. the transition probability between the stages) is driven by depressive symptoms at a subclinical level. We report a series of findings that advance the knowledge of the potential association of moderate levels of depressive symptoms and sleep stage dynamics. The dynamic flow of sleep stage transitions is of interest in trying to understand how exactly the sleep is organized in a temporal mode in normal v. depressed individuals.
The physiology underlying the temporal regulation of the ultradian rhythm of REM/NREM alternation is not fully understood, but there is some evidence pointing to mechanisms that can be disturbed in depression. While the switch between the two brain states occurs rapidly within seconds, the recurrence of the NREM/REM cycle follows usually a much slower ultradian rhythm, ranging from few minutes in rodents to hours in humans (Weber et al., Reference Weber, Hoang Do, Chung, Beier, Bikov, Saffari Doost and Dan2018). The cyclicity of NREM and REM sleep depends on the signaling of GABAergic neurons in the ventral periaqueductal gray (vPAG) (Weber et al., Reference Weber, Hoang Do, Chung, Beier, Bikov, Saffari Doost and Dan2018), modulated via projections from medial prefrontal cortex (mPFC) (Chang, Chen, Qiu, & Lu, Reference Chang, Chen, Qiu and Lu2014). Depression alters the activity of mPFC (Drevets, Price, & Furey, Reference Drevets, Price and Furey2008), and consequently, vPAG, which is suggested to underlie REM sleep abnormalities in depressed people (Chang et al., Reference Chang, Chen, Qiu and Lu2014). Additionally regulating NREM/REM pattern is the antagonist interaction of cholinergic and aminergic brainstem cells (O'Malley & Datta, Reference O'Malley and Datta2013), which is destabilized in depression (Palagini, Baglioni, Ciapparelli, Gemignani, & Riemann, Reference Palagini, Baglioni, Ciapparelli, Gemignani and Riemann2013).
First, we observed some differences in the association between various sleep stages when analyzed as a sleep stage network using a GGM. This method has been previously used in network analytics, such as in mental health research. In psycho(pathological) networks, the focus is on individual symptoms or psychological states and how such symptoms or states influence each other. Symptom networks have been inferred and analyzed for several psychopathological domains, including depression (Borsboom, Reference Borsboom2017; Boschloo, Schoevers, van Borkulo, Borsboom, & Oldehinkel, Reference Boschloo, Schoevers, van Borkulo, Borsboom and Oldehinkel2016a; Boschloo, van Borkulo, Borsboom, & Schoevers, Reference Boschloo, van Borkulo, Borsboom and Schoevers2016b; Fried et al., Reference Fried, van Borkulo, Cramer, Boschloo, Schoevers and Borsboom2017; van Borkulo et al., Reference van Borkulo, Boschloo, Borsboom, Penninx, Waldorp and Schoevers2015). We applied this concept to analyze the dynamics of sleep stages and our results suggest that stage N1 seems to be an influential node stage, as it associates with the amount of other stages most strongly in individuals with moderate depression. As N1 is closest to wake and follows the short arousals that are not scored as true wake, this might indicate that homeostatic sleep regulation is impaired, both in terms of the sleep–wake switch and in terms of reaching deeper sleep from N1.
Second, we used Markov chain analyses to examine the transition probabilities by focusing more directly on the differences in the sequential structure of the sleep stages between those with moderate depression and others. We found that transitions from REM to N2 and also from N2 to N3 were more probable in those with moderate depression compared to those without depression, which suggests more instability in sleep stages among those with moderate depression. However, we did not find differences in survival probabilities in any of the sleep stage bouts.
Our results are consistent with the previous study of Wei et al. (Reference Wei, Colombo, Ramautar, Blanken, van der Werf, Spiegelhalder and Van Someren2017) on sleep stage transitions in insomnia, suggesting that SWS (and especially N2) may be the central stages that warrant further attention in future studies. Studies on sleep-disordered breathing have also previously reported instability of stage N2 and REM sleep (Bianchi, Cash, Mietus, Peng, & Thomas, Reference Bianchi, Cash, Mietus, Peng and Thomas2010). Thus, it is possible that the microstructure and phasic components of stage N2 (spindles and K-complexes) may be particularly useful in examining the brain mechanisms related to moderate depression and other mental disorders.
There were no statistically significant differences in the survival functions of sleep stages between the depression status groups. The largest differences between the groups were for N2 and N3. We calculated the survival function probabilities using both means of each sleep bout and also using the recurrent event (mixed-effects) survival analyses, which produced similar results. Previous studies have used various parametrizations of sleep bout survival probabilities and thus we repeated our analyses using the semiparametric coxph function (flexsurv package). The only difference was that the survival function in the N3 bout was shorter in those with moderate depression compared to others (estimate −0.11, p = 0024).
Strengths and limitations
The sleep measures in this study were based on the state-of-the-art PSG in the participants' own homes, which increases measurement comfort and improves sleep quality. However, PSG measurements covered only one night for each participant, and repeated measurements would have provided more information and enabled more detailed analyses (Wei et al., Reference Wei, Colombo, Ramautar, Blanken, van der Werf, Spiegelhalder and Van Someren2017). However, actigraphy measurement comparisons between the PSG night and measurements covering a week before the PSG night suggested no differences regarding sleep duration or quality. Thus, the PSG measurement can be considered to represent a typical night's sleep. Although the specific timing of the PSG measurement in relation to the circadian phase of the participants may influence sleep staging, we were not able to define the circadian phase with sufficient specificity. Again, knowing that the measurements were conducted according to the participants' usual sleep schedules, measured with actigraphy for 1 week, lowers the probability of bias. As a similar methodological limitation, although BDI-II is a widely used and reliable instrument for measuring depressive symptoms, clinical interviews may have provided more reliable cutoff points for moderate depression. Using community-based samples with no clinical population also means that our cutoff point may have included mild cases in the case group. Related to that there was an unavoidable imbalance of group sizes derived from the sample and the depressed group is of modest size compared to the rest of the sample and that may be problematic as unequal sample size makes it more important that homogeneity of variance assumption would be met although we did not find this to be exceptionally problematic in group differences in the transition probabilities. Furthermore, our participants were all young and sleep stage architecture is known to be dependent on age (Tarokh, Van Reen, LeBourgeois, Seifer, & Carskadon, Reference Tarokh, Van Reen, LeBourgeois, Seifer and Carskadon2011; Yetton et al., Reference Yetton, McDevitt, Cellini, Shelton and Mednick2018).
Conclusions
Our results suggest that there are some differences in the dynamic sleep structure between those with moderate depression and others. These differences have previously been observed in those with diagnosed insomnia disorder or severe clinical depression. Although the results obtained in our study should be replicated, they suggest that differences in these transitions may be detected already in earlier and sub-clinical stages of depression or depressive symptoms. The network model of sleep stages offers new insights to the temporal organization of sleep in humans in relation to mood. While animal models provide strong candidate mechanisms for the physiological regulation of the rapid REM/NREM alterations (Weber et al., Reference Weber, Hoang Do, Chung, Beier, Bikov, Saffari Doost and Dan2018), there is a need for translational research to advance the clinical implications in humans.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291719003453.
Author contributions
KR designed the original GLAKU study. A-KP designed the sleep collection follow-up. ME and A-KP designed the original hypothesis and ME performed all the analyses in close collaboration with JL and A-KP. ME wrote the first draft of the manuscript. All the authors interpreted the results, revised the text, and approved the final draft of the report.
Financial support
Academy of Finland. The sponsors of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. Elovainio takes final responsibility for the decision to submit for publication.
Conflict of interest
We declare that we have no conflicts of interest.
