What do women mean by poor sleep? A large population-based sample with polysomnographical indicators, in ﬂ ammation, fatigue, depression, and anxiety *

Survey studies indicate that reports of disturbed sleep are prevalent and may be prospectively linked to several major diseases. However, it is not clear what self-reported disturbed sleep represents, since the link with objective sleep measures (polysomnography; PSG) seems very weak. The purpose of the present study was to try to investigate what combination of variables (PSG, in ﬂ ammation, fatigue, anxiety, depression) that would characterize those who complain of disturbed sleep. This has never been done before. Participants were 319 women in a population-based sample, who gave ratings of sleep quality, fatigue, depression, and anxiety, then had their sleep recorded at home, and had blood drawn the following morning for analysis of immune parameters. Correlations and hierarchical multivariable regression analyses were applied to the data. For ratings of dif ﬁ culties initiating sleep, the associations in the ﬁ nal step were ß ¼ .22, (p < .001) for fatigue, ß ¼ 0.22 (p < .001) for anxiety, and ß ¼ 0.17 (p < .01) for sleep latency, with R2 ¼ 0.14. The rating of repeated awakenings was associated with fatigue (ß ¼ 0.35, p < .001) and C-reactive protein (CRP) (ß ¼ 0.12, p < .05), with R 2 ¼ 0.19. The rating of early morning awakenings was associated with fatigue (ß ¼ 0.31, p < .001), total sleep time (TST) (ß ¼ (cid:2) 0.20, p < .01), and CRP (ß ¼ 0.15, p < .05), with R 2 ¼ 0.17. Interleukin-6 and Tumour Necrosis Factor were not associated with ratings of sleep problems. The results indicate that subjective fatigue, rather than objective sleep variables, is central in the perception of poor sleep, together with CRP.


Introduction
Subjective sleep problems have a high prevalence in the population, 17.5% for insomnia and 26.6% for "subsyndromal insomnia" [31], and are linked to a number of diseases, including cardiovascular disease, Alzheimer's disease, and diabetes [24].However, the link between reports of disturbed sleep do not correspond well with polysomnographic (PSG) indicators of poor sleep [6,9,16] and a systematic study to search for traditional PSG indicators of insomnia found no significant associations with reports of poor sleep [15].In addition, a large group of insomnia patients have been identified as suffering from "sleep state misperception" or "paradoxical" insomnia because of the lack of correspondence between the insomnia diagnosis and objective sleep data [11].Still, a metaanalysis has found a modest difference in sleep architecture between patients with insomnia and controls [5].In addition, there are several developments that find microarchitectural indicators, such as sleep spindles, high or low frequency spectral power to be possible indicators of reported sleep problems, including clinical insomnia [4,16], but clear and simple criteria are still lacking.
The observations above raise the question of what respondents mean when they report "disturbed sleep", if it is not objectively observed long bouts of wake, low sleep efficiency, many awakenings, etc.Interestingly, Harvey et al. [20] have shown that individuals' perception of good or poor sleep seems less based on the perception of self-reported sleep continuity variables (as the number of awakenings, sleep efficiency, etc.), than on states like fatigue, or anxiety, or mood.Both anxiety [13] and depression [35] are common occurrences in disturbed sleep in general, as is fatigue [10,18].Thus, there is a possibility that psychological states may drive a perception of disturbed sleep, rather than, or in addition to, subjective (e.g.remembered awakenings) or objective (e.g.sleep efficiency) quantitative aspects of sleep [19].
Other non-PSG variables such as inflammation markers may also be involved in the perception of disturbed sleep.For example, we recently found that levels of C-reactive protein (CRP) were related to reported difficulties in sleep maintenance and early awakenings, but not with PSG indices of sleep continuity or slumber stages (extracted as principal components) [17].Furthermore, levels of CRP and Interleukin 6 (IL-6) are increased in individuals with subjectively disturbed or short sleep [26], fatigue [14], or depression [32], while the link to anxiety is less well established [32].The immune system has also been seen as a sleep regulator, promoting sleep, and high level immune system activation interferes with sleep [7].Key cytokines seem to be IL-6 and TNFalpha.
No work seems to exist on CRP in this respect.Taken together, several immune parameters may be associated with the perception of poor sleep (as well as with fatigue, depression or objective sleep).The chain of causation may, apparently, be bi-directional.
A new development in the subjective/objective sleep issue is that it seems that insomnia patients suffer from a fragmented REM sleep, labeled "Restless REM, which may prevent emotional resetting through sleep [38], possibly due to interruption of the suppression of noradrenaline activation from the locus coeruleus that normally occurs during REM sleep.Experimental interference with this suppression prevents emotional resetting [39].One of the variables in the restless REM concept, REM episode interruption, is highly correlated with the Insomnia Severity Index (ISI) [39].Thus, restless REM may be another interesting variable in understanding the meaning of self-reported sleep quality.
Considering the observations above, it seems an interesting idea to try to understand the meaning of subjectively poor sleep by looking at PSG, immune parameters, fatigue, depression, and anxiety in the same analysis, rather than one at a time.This would inform us which of the variables that have the strongest association with subjectively poor sleep, including which variables that do not add any independent explanatory power.Based on the strong presence of fatigue in insomnia, we would expect fatigue to be a key predictor that modifies the association between subjective sleep problems and depression, immune parameters, and anxiety.PSG variables are unlikely to be modified since their association with poor sleep is expected to be weak.This type of analysis has not been carried out before, and the results would increase our understanding of perceived disturbed sleep.With this approach we build on the previous study [17] with PSG and CRP, but instead of combining variables into factors (as in the previous study), we will use individual PSG variables and add other explanatory variables such as depression, anxiety and fatigue, as suggested by Harvey et al.For subjective sleep quality, such symptoms may be the three types of sleep disturbances used to diagnose insomnia disorder, namely, difficulties initiating sleep, repeated awakenings with difficulties going back to sleep, and early morning awakenings [30].
The present study focuses on women, mainly since women are understudied and also have a pattern of a high frequency of sleep complaints [27], compared to men, while objectively they seem to sleep quite well, compared to men, for example, amounts of stage 1 are lower, and stage 3 are higher than in men [1,8].
The purpose of the present study was to investigate what combination of variables (PSG variables, inflammatory markers (CRP, IL-6 and TNFa), as well as self-reported fatigue, depression, anxiety), would be associated with self-reported sleep problems in a population-based female cohort, using a multivariable approach.
The hypothesis was that fatigue would be the key variable associated with subjective sleep, modifying the associations of the other variables, except PSG variables, which were not expected to show significant associations as stand-alone predictors.The results may have implications for understanding sleep complaints in the population, as well as in groups with insomnia disorder.

Background and design
The present study is part of a study of sleep in women, using the community-based cohort "Sleep and health in women" (SHE) at Uppsala University [36].The purpose of the SHE study was to study obstructive sleep apnea and a number of metabolic health parameters in women.A representative sample of 10,000 women in the Uppsala region (Sweden) responded to a sleep/health questionnaire (response rate 71.6%).A random sample (n ¼ 400) was drawn from the cohort for PSG recording and blood samples (snorers were oversampled).These participants had their sleep recorded during one weekday and responded to a questionnaire on sleep and health [36].The sample used in the present study comprised 319 participants, since 81 had a technical error in the start of the recording, which affected mainly sleep latency measures.The study was approved by the Ethical committee of the University of Uppsala (Dnr 01e238).

PSG recording
The Embla (Flaga, Iceland) solid state, portable, sleep recorder was used to record unsupervised sleep in the homes of the participants.Standard electrode (silver/silver chloride) montage was used (C3, C4) referenced to contralateral mastoids.In addition, two sub-mental electrodes, as well as electrodes at the outer canthi of the eyes were used.To adapt to AASM scoring, F4 was interpolated.Further sensors used, but not reported on here in detail, were bilateral anterior tibialis muscles, airflow with a three-port oronasal thermistor and a nasal flow pressure sensor, respiratory effort from piezo-electric belts (Resp-EZ, EPM Systems Midlothian, VA, USA), finger pulse oximetry (Embla A10 flex Sensor), electrocardiograms (V5), a piezo vibration sensor for snoring and a body position sensor.
One research nurse applied the electrodes, connected the equipment, and gave instructions in the early evening.Next morning, the equipment was retrieved by an experimenter.Data were lost for 1.5% (6 of 400) of the participants but the recording was then repeated within a short period of time.Diary information was used to establish lights out and lights on.
Sleep scoring (stages, respiration, arousals) were performed according to the classification criteria of the American Academy of Sleep Medicine [23] using the computer-assisted sleep classification system Somnolyzer 24 Â 7 [2,3].Here the terminology N1, N2 and N3 is used for sleep stages 1e3.Wake after sleep onset (WASO) represents time awake between sleep onset and offset in minutes, and is expressed as percent of the total sleep period (TSP).Stage shifts refers to changes between any of stages N1, N2, N3, Wake, or REM and is expressed per hour.Shifts from any of the sleep stages to wake is expressed as awakenings per hour.An apnea was defined as a cessation of airflow for at least 10 s, while a hypopnea was defined as at least 10 s of 50% reduced respiratory volume, together with at least 3% desaturation.The apnea-hypopnea index (AHI) was defined as the mean number of apneas and hypopneas per hour of sleep.REM episode interruption density was computed as number of changes from REM to wake, N1, N2, N3 per hour of REM sleep [39].

Blood samples and anthropometric measures
Participants returned in the morning after the PSG and (fasting) blood samples were drawn (between 7:00 and 9:00) for plasma CRP, IL-6 and TNFa.According to the instructions of the manufacturer, the minimum detectable levels for CRP were <0.2 mg/L, and for IL-6 <0.5 ng/L while TNFa was detectable across the entire range of measurements.In the statistical analyses, a value of 0.1 mg/L and 0.25 ng/L were used for CRP and IL-6, respectively, when the minimum detectable level was found.Height and weight were measured by a research nurse to calculate body mass index (BMI; kg/m 2 ).

Self-ratings
The questionnaire used included the Uppsala Sleep Inventory [22] and the Hospital Anxiety and Depression scale e HAD [28,41].They were filled out on the day of the sleep recording.Many of the questions were the same as was filled out for the 10,000 participants of the survey study.The questionnaire also contained questions on use of medication and presence of major diseases.We combined these into a variable that represented the presence of major disease (stroke, myocardial infarction, angina pectoris, or diabetes) and/or use of hypnotics, sedatives, antidepressants.
The three main dependent variables were obtained from the Uppsala Sleep Inventory.Participants were asked to state how much difficulty they have 1) "falling asleep in the evening," 2) "waking several times during the night," and 3) "waking too early and having difficulty falling asleep again."The responses did not refer to any particular time span.Ratings of fatigue were obtained from the same inventory.All items are scored from 1 ¼ not at all to 5 ¼ a great problems, and with 3 ¼ "neither".These values were used for correlation and regression analyses.A combined score of 4e5 was used to indicate sleep disturbance for each variable in order to provide a prevalence estimate.Note that the response focuses on whether the disturbance is a problem, not its frequency (which is a common approach).
HAD contains seven questions associated for depression and seven associated for anxiety (for example, "I look forward to things with joy", or "I feel tense").The score ranges from 0 to 3 (never e very often), yielding a maximum score of 21 for each scale.A score of !11 is usually taken to indicate anxiety or depression, respectively.In the analyses we use the continuous scale (0-21).

Statistics
In order to select individual PSG variables for the multiple regression analyses, correlations were computed between PSG variables and sleep problem ratings as well as between blood parameters and sleep problem ratings.All variables that were significantly correlated with at least one sleep rating were then entered into multiple regression analyses.These analyses were organized in four steps.CRP seems in prior work to both induce and interfere with sleep, as well as to be linked to anxiety/depression and fatigue.It was therefore of interest to put this variable first in the sequence to study how it might be modified by the subsequent steps.PSG variables were entered in model 2 since they might modify the associations between immune parameters and poor sleep.Depression and anxiety were added in model 3 since they were expected to be linked to poor subjective, as well as to objective sleep, and to immune activation.In model 5 fatigue was entered since we expected that variable to modify most of the associations in the previous steps.The inflammatory parameters showed a skewed distribution and were transformed using the natural logarithm (ln).Since total sleep time (TST) and sleep efficiency were highly correlated (r ¼ .78)and collinear, the latter variable was not entered into the regression analysis, but substituted for TST in a sensitivity analysis.
Finally, we selected the individuals with <80% and !90% sleep efficiency and compared them in terms of ratings of sleep problems, using a repeated measures analysis of variance.

Results
Table 1 shows mean ± SD for all variables.The mean age was 50.2 years (range: 22e72 years), and 56.9% were overweight or obese.The sleep disturbance with the highest prevalence was difficulties maintaining sleep.Note that PSG mean sleep efficiency was 83.8%, with a standard deviation of 12.3%, which means that the ±1SD range is 71.5e96.1%., that is, a wide span.Other sleep continuity variables showed a corresponding large distribution.In addition to Table 1 and it is of interest that 20.1% of the women report difficulties initiating sleep as belonging to the categories "a big" or "very big problem".Corresponding values for difficulties maintaining sleep was 15.2% and for early morning awakenings 27.9%.A total of 12.3% of the sample had a score above the cut-off (!11) for anxiety, and 3.3% above the cut-off for depression (!11).
Supplemental Tables S1aeb shows the correlations between all variables.IL-6 and TNFa lacked significant associations with sleep ratings, and were left out of further analyses.CRP showed a significant association with difficulties maintaining sleep and early morning awakenings.Depression, anxiety, and fatigue showed significant correlations with all three sleep problem ratings.TST, sleep efficiency, N2%, REM latency, REM%, sleep latency and WASO, showed significant associations with at least 1 rating (Table S1).The rest of the PSG variables did not show any significant association with the sleep problem ratings.The significant variables were entered into a multiple regression analysis against the dependent variables, adjusted for age, BMI, and major disease plus medication.However, collinearity was considerable between TST and sleep efficiency as well as with REM minutes.The latter two were, therefore left out of the regression.Also, WASO was left out of the regression because of strong collinearity with sleep efficiency.Table 2 shows results from the hierarchical multiple regression analysis with difficulties initiating sleep as dependent variable.Model 1 includes only lnCRP, which was not significantly associated with difficulties initiating sleep.When adding PSG parameters (model 2), sleep latency showed a significant association.When adding anxiety and depression (model 3), anxiety (but not depression) showed a significant association, and sleep latency remained significant.In model 4, fatigue was added, and showed a significant association, together with sleep latency and anxiety.The results mean, in terms of unstandardized regression coefficients (B) corresponding to the beta coefficients in Table 2, that an increase in fatigue by one unit (scale; 1e5) was associated with an increase in problems with difficulties initiating sleep by 0.21 units (scale: 1e5).That an increase of one unit in anxiety (scale 0e21), was associated with an increase in 0.07 units of difficulties falling asleep, and that an increase of 1 min of sleep latency was associated with an increase by 0.01 units of difficulties falling asleep.
Table 3 shows a significant association for lnCRP with difficulties maintaining sleep throughout the four models.None of the PSG variables showed a significant association when added in model 2. Anxiety, but not depression, showed a significant association with difficulties maintaining sleep when entered in model 3. Fatigue showed a significant association when entered in model 4, with CRP remaining significant.The B coefficients show that the results mean that an increase of one unit in fatigue was associated with an increase of 0.21 units of difficulties maintaining sleep (scale 1e5), and that an increase of one unit in lnCRP was associated with an increase of 0.06 units (or 0.013/unit of CRP) in difficulties maintaining sleep (scale 1e5).
In Table 4, lnCRP showed a significant association with early morning awakenings throughout the four models.TST showed a significant (negative) association with early morning awakenings when entered in model 2 and throughout models 3 and 4. Fatigue showed a significant association with early morning awakenings when entered in model 4, with lnCRP and TST remaining significant.The results mean that an increase in fatigue was associated with an increase of 0.29 units of problems with early morning awakenings (scale 1e5), that an increase in sleep duration of 1 h was associated with a decrease of 0.18 units of problems with early morning awakenings, and that an increase of one unit of lnCRP was associated with an increase of 0.18 units (or 0.039/unit of CRP) of problems with early morning awakenings.
We made several sensitivity analyses.In one, we excluded those participants who had reported disease or had taken sleep medication, sedatives or antidepressants (n ¼ 66).The associations were slightly attenuated, but all results remained significant.In another sensitivity analysis, sleep efficiency, WASO, and REM minutes were substituted, one at a time, for TST, in the multiple regression analyses (all variables were significantly correlated).They all showed significant regression coefficients (not shown) with early morning awakenings as the dependent variable, but not with the other two dependent variables.In a third set of sensitivity analyses (not shown), we added AHI (not significantly correlated with the sleep ratings) to the regressions, since we had over-sampled for individuals with frequent snoring.In no analysis did the significant regression coefficients change more than 0.01 units (not shown).In a fourth sensitivity analysis, we repeated the main study using ratings of sleepiness and awakening unrefreshed instead of fatigue (using the same formulation and response format as for fatigue).The results were very similar to those in which fatigue was used (not shown).

Discussion
The main result in this large cohort study analyzing several possible predictors of reporting sleep problems simultaneously, was that fatigue turned out to be associated with the ratings of all three sleep problems, that lnCRP was significantly associated with rated difficulties maintaining sleep and early morning awakenings, that anxiety had a rather strong link with difficulties initiating sleep, that objectively measured sleep latency (PSG) was associated with rated difficulties initiating sleep, and that TST (PSG) was associated (negatively) with early morning awakenings.Difficulties maintaining sleep had the highest level of complaints of the three ratings, as well as the highest model fit in the fully adjusted analysis.
Overall, fatigue appears to have strongly modified the association of depression with all three sleep problems, and the association of anxiety with both repeated awakenings and too early awakenings.However, the CRP association with the latter two outcomes was not modified, nor was the association of TST with early awakenings, or for sleep latency with difficulties falling asleep."Thus, the initial hypothesis was only partially verified.With regard to the PSG variables, the weak association with ratings of sleep problems are compatible with previous work [6,9].The significant (r ¼ .17)association between sleep latency and difficulties initiating sleep seems logical, and agrees with prior work [9].Also, the significant (negative) association (r ¼ À.15), between TST and early awakenings seems logical since such awakenings in many cases would mean a shortening of sleep.However, we have failed to find other studies to compare with.Rated difficulties maintaining sleep was expected to be associated with PSG awakenings or arousals, but was not.Again, no previous studies of this association seem available.The weak objective/ subjective association was also reflected in the lack of difference in problems with repeated or early awakenings between very high and very low groups on sleep efficiency.Finally, REM interruption density lacked significant association with the three ratings of sleep symptoms which is in contrast to previous findings [39].However, the present study only used ratings of sleep, whereas previous work also included items on consequences of disturbed sleep (using the ISI), including worries about disturbed sleep.The latter might yield higher association with REM interruption density.It seems as if subjective sleep problems have an almost negligeable association with objective sleep.
One reason for the lack of association between reported sleep disturbances and PSG variables could be that one single recorded sleep may not be representative for several recorded sleeps across a period that correspond to, for example, a one to three-month period which ratings of sleep problems often refer to.There is a scarcity of such studies, but one large study (N ¼ 285) found relatively high intra-class correlation coefficients across three recordings [40], which argues against a large variability between recordings.Another possibility is that objective sleep quality (reduced sleep continuity) is not perceived by the individual, unless it is quite large.The comparison of the two subgroups with low and high sleep efficiency, seems to support such a notion as only very small differences in ratings of sleep disturbances were seen, even though the difference in mean sleep efficiency was very large.Possibly, new developments using AI and a multitude of derived microscopic PSG measures may identify new variables [4].
There is also the possibility that more subtle aspects of sleep have stronger links to subjectively disturbed sleep as, for example, high beta frequencies in the EEG [34], fragmented ("restless") REM sleep [38], or wake activity in the EEG during deep N3 sleep in insomniacs [12].However, the effects seem marginal, and the paradox remains, large segments of objective time awake during sleep are not perceived as difficulties sleeping.Nor are many awakenings.
An alternative idea is that the report of disturbed sleep is largely based on other influences than PSG indicators.A major finding of the present study was that fatigue was significantly associated with all three sleep problems, after adjustment for PSG and other variables.Similar multivariable studies are lacking, but fatigue is a common observation in insomnia disorder [10,18].As mentioned earlier, both insomniacs and non-insomniacs seem to associate  disturbed sleep mainly with fatigue and anxiety or depression, rather than, for example, with self-reported number of awakenings, sleep duration, sleep efficiency, or other quantitative measures [20].The design of the present study does not permit conclusions on causality, but it seems possible that a state of fatigue (due to non-sleep factors like immune activation, thyroid dysfunction, adrenocortical dysfunction, etc [33]) could make fatigued individuals look for causes of their fatigue among rather normal sleep perturbations [21].This possibility needs investigating in longitudinal studies.Depression, which is normally associated with sleep problems [35], appears to lose its significant association with sleep problems when fatigue was added, and fatigue is a basic characteristic of depression.Anxiety, however, was present in the final model in the analysis with difficulties of initiating sleep.
The findings on CRP agree with meta-analytic evidence of CRP being associated (positively) with self-reported sleep disturbances and short sleep [7,26].However, the second major finding was that CRP predicted problems with repeated and early awakenings, also after adjustment for PSG variables, fatigue, and other variables.Again, one may speculate that sleep problems may cause a chronic low grade inflammation since both CRP and IL-6 are linked to reported disturbed sleep in observational studies [26], and cognitive behavioral therapy for insomnia reduces CRP levels [25], and inflammation causes fatigue [14].Alternatively, if fatigue is present because of non-sleep related causes, the individual might attribute the causes of fatigue to their sleep [21].Possibly, also the malaise and anxiety associated with inflammation may induce a perception of poor sleep.To the best of our knowledge there exist no studies of the effects of CRP alone on objective or subjective sleep.This should be a topic for future studies.With regard to difficulties of initiating asleep, CRP is unlikely to be involved in the link with fatigue, since the association was not significant.Other causes of fatigue could be a number of diseases and states like anemia, as well as thyroid and adrenocortical dysfunction [33].The lack of association between ratings of disturbed sleep and IL-6 and TNFa could be due to circadian variation [37], while CRP is more stable across the 24 h span [29].
The strength of the study is its large sample and the consideration of multiple predictors of sleep quality ratings, including PSG, immune parameters and fatigue ratings.The multivariable approach is probably the most important contribution of the present study.A weakness is the cross-sectional design, which prevents conclusions on causality.The cross-sectional approach also involves a risk of common method variance.Another possible problem is that a single recorded sleep may be influenced by the context of that particular sleep, and thus not be representative of sleep quality subjectively integrated across a longer time period.While the argument is valid, it is still a reasonable expectation to find clear evidence of association in the present, as well as in other studies.One might also see the focus on only women as a weakness, or at least, we don't know if men would show a similar pattern of associations.One might on the other hand, see this focus as a strength since women tend to be under-represented in research studies.However, child birth, contraceptives, and menopausal state are likely to have affected sleep, although it is not clear if it would have affected the association between sleep problems and and its associated variables.
In conclusion, we found that subjectively poor sleep in a large population-based sample of women, was characterized mainly by fatigue and, for repeated awakenings and early awakenings, also by increased CRP levels.PSG sleep latency and TST had modest links to difficulties falling asleep and early morning awakenings, respectively, while anxiety was linked to difficulties falling asleep.It is suggested that the psychological state of the individual (like fatigue) may influence the individual to attribute their state to poor sleep (despite a lack of objective sleep disturbance).This needs to be confirmed in longitudinal studies, however.

Table 1
Mean ± SD, or %, and N for all variables.

Table 2
Results from hierarchical multiple regression with difficulties initating sleep as dependent variable.Beta coefficients (95% CI) and N for models.Adjusted for age, BMI, and illness/medication.

Table 4
Results from hierarchical multiple regression analysis with early morning awakenings as the dependent variable.Beta coefficients (95% CI).Only complete data, adjusted for age, BMI, and illness/medication.