Association between percentage stage 1 and age in men and women. Adjusted (least squares mean) levels of percentage stage 1 at each age quartile (≤54, >54 to ≤61, >61 to ≤70, and >70 years) for men (triangles) and women (circles).
Association between percentage stage 2 and age in men and women. Adjusted (least squares mean) levels of percentage stage 2 at each age quartile (≤54, >54 to ≤61, >61 to ≤70, and >70 years) for men (triangles) and women (circles).
Association between percentage stage 3-4 and age in men and women. Adjusted (least squares mean) levels of percentage stage 3-4 at each age quartile (≤54, >54 to ≤61, >61 to ≤70, and >70 years) for men (triangles) and women (circles).
Redline S, Kirchner HL, Quan SF, Gottlieb DJ, Kapur V, Newman A. The Effects of Age, Sex, Ethnicity, and Sleep-Disordered Breathing on Sleep Architecture. Arch Intern Med. 2004;164(4):406-418. doi:10.1001/archinte.164.4.406
Polysomnography is used to assess sleep quality and to gauge the functional effect of sleep disorders. Few population-based data are available to estimate the variation in sleep architecture across the population and the extent to which sleep-disordered breathing (SDB), a common health condition, contributes to poor sleep independent of other factors. The objective of this study was to describe the population variability in sleep quality and to quantify the independent associations with SDB.
Cross-sectional analyses were performed on data from 2685 participants, aged 37 to 92 years, in a community-based multicenter cohort study. Dependent measures included the percentage time in each sleep stage, the arousal index, and sleep efficiency. Independent measures were age, sex, ethnicity, comorbidity status, and the respiratory disturbance index.
Lighter sleep was found in men relative to women and in American Indians and blacks relative to other ethnic groups. Increasing age was associated with impaired sleep in men, with less consistent associations in women. Notably, women had, on average, 106% more slow wave sleep. Sleep-disordered breathing was associated with poorer sleep; however, these associations were generally smaller than associations with sex, ethnicity, and age. Current smokers had lighter sleep than ex-smokers or never smokers. Obesity had little effect on sleep.
Sleep architecture varies with sex, age, ethnicity, and SDB. Individual assessment of the effect of SDB on sleep quality needs to account for other host characteristics. Men, but not women, show evidence of poorer sleep with aging, suggesting important sex differences in sleep physiology.
Sleep is increasingly recognized as a biological function necessary for optimal daytime functioning. Insufficient or poor quality sleep has been linked to neurocognitive impairments,1- 5 end-organ dysfunction and chronic health conditions,6- 8 and increased mortality.9,10 The public health importance of sleep is underscored by a recent report that most of the US workforce experiences sleepiness that interferes with job performance11 and by reports that sleepiness contributes to work and vehicular accidents.12
Insufficient or poor-quality sleep may be caused by personal or societal influences, may occur as secondary effects of other health problems, or may be due to several primary sleep disorders. Of the primary sleep disorders, sleep-disordered breathing (SDB) is a common problem in primary care settings.13 This condition is estimated to affect 2% to 4% of middle-aged adults14 and more than 25% of older populations.15
Sleep quality is objectively measured using polysomnography (PSG), which is performed routinely as part of the clinical evaluation of patients presenting with sleep disorders. The absolute and percentage times in given sleep stages, as well as the pattern and timing of progression from one stage to another, provide information on overall sleep architecture and are used to quantify the degree of sleep fragmentation. Sleep characterized by frequent awakenings, arousals, and little slow wave (delta or stage 3-4) sleep is considered to be lighter or nonrestorative, resulting in daytime sleepiness and impaired daytime function.16 In the case of SDB, assessment of sleep architecture with PSG is often used to grade the overall severity of the underlying breathing disorder and to gauge response to therapeutic interventions. The disorder is considered to be more severe when the occurrence of respiratory disturbances disrupts sleep continuity. Evaluation of response to therapy often includes assessment of the extent to which sleep fragmentation improves, as measured by a reduction in the arousal index (ArI) or the relative time in stage 1 sleep or by an increase in the time in slow wave or rapid eye movement (REM) sleep.17- 19
Despite the frequency with which sleep stage data are used in the evaluation of patients with sleep disorders or daytime sleepiness, there are few normative data that describe the population heterogeneity in sleep architecture. Previous studies20- 28 have included small samples, representing a limited range of demographic and comorbid conditions. The extent to which the population variability in sleep architecture may be attributable to specific sleep disorders, to the effects of aging or other demographic factors, or to comorbidity is poorly understood. There are little data that specifically address the degree to which the common condition of SDB may alter sleep architecture independent of other effects, such as age or comorbidity. Therefore, clearly defined criteria for interpreting variations in sleep stage proportions across individuals who differ by age, sex, and various indices of comorbidity are lacking. Understanding the population heterogeneity of sleep architecture may be important in clarifying the distribution of daytime sleepiness in the population.
This study assesses the variation of sleep architecture with age, ethnicity, sex, and obesity and quantifies the extent to which SDB affects sleep architecture independent of these effects. Analyses were performed for participants in the Sleep Heart Health Study (SHHS). This cohort, representing a large sample of subjects with a wide age and respiratory disturbance index (RDI) range, is derived from a multicenter, community-based study that has used standardized unattended (generally in-home) PSG and rigorously controlled scoring approaches.
The overall objectives and study design of the SHHS have been reported previously.29 Briefly, the SHHS is a prospective cohort study aimed at investigating the relationship between SDB and cardiovascular disease (CVD). Participants were recruited from 9 existing epidemiological studies in which data on cardiovascular risk factors had been collected previously. From these parent cohorts, a sample of participants who met the inclusion criteria (age ≥40 years, no history of treatment of sleep apnea, no tracheostomy, and no current home oxygen therapy) was invited to participate in the baseline examination of the SHHS. Several cohorts oversampled snorers to increase the study-wide prevalence of SDB. Between November 8, 1995, and January 31, 1998, the study enrolled 6443 individuals. The analysis was limited to 2685 subjects whose study data met the criteria described herein.
The baseline SHHS data collection session, generally conducted at the participants' homes, included a brief standardized health interview and questionnaire administration, assessment of current medication use, blood pressure and anthropometric measurements, and unattended PSG.29 A self-completed sleep habits questionnaire (administered before or at the time of the baseline examination) provided information on perceived sleep disturbances and sleep quality. A history of physician-diagnosed medical illnesses, previous surgical treatments, or medical procedures was obtained from the health interview or from data obtained from the parent cohort databases. The interviewer recorded prescribed and selected nonprescribed medications taken within the 2 weeks before the home visit. The interviewer transcribed information from the medication container onto the data collection form, including the medication name, strength, and quantity prescribed per period. Medications that were members of drug classes of interest (including antipsychotics, benzodiazepines, monoamine oxidase inhibitors, nontricyclic antidepressants, and tricyclic and tetracyclic antidepressants) were identified using a computer program that matched medications to their appropriate national drug code numbers and class code.30
Overnight, PSG was performed using the Compumedics Portable PS-2 System (Abbottsville, Australia), using methods previously detailed.31 Sensors were placed and equipment was calibrated during an evening visit by a certified technician. Data collection included 2 central electroencephalograms (EEGs); right and left electro-oculograms; a bipolar submental electromyogram; thoracic and abdominal excursions (inductive plethysmography bands); airflow (detected by a nasal-oral thermocouple [Protec, Woodinville, Wash]), oximetry (finger pulse oximetry [Nonin Medical, Inc, Minneapolis, Minn]), and electrocardiogram and heart rate (using a bipolar electrocardiogram lead); body position (using a mercury gauge sensor); and ambient light (on or off, by a light sensor secured to the recording garment). Following equipment retrieval, the data, stored in real time on personal computer cards, were downloaded to the computers at each respective clinical site, locally reviewed, and forwarded to a central reading center (Case Western Reserve University).
Sleep stages were scored according to the guidelines developed by Rechtschaffen and Kales.32 Stages 3 and 4 were combined (slow wave sleep). Arousals were identified according to American Sleep Disorders Association (American Academy of Sleep Medicine) criteria.33 An apnea was defined as a complete or almost complete cessation of airflow (<25% of baseline), as measured by the amplitude of the thermocouple signal, lasting 10 seconds or longer. Hypopneas were identified if the amplitude of a measure of flow or volume (detected by the thermocouple or thoracic or abdominal inductance band signals) decreased to less than 70% of the amplitude of baseline breathing for 10 seconds or longer, but did not meet the criteria for apnea. For this study, only apneas or hypopneas associated with at least a 4% oxyhemoglobin desaturation were considered in the calculation of the RDI. Manually marked events were tabulated by specialized software to provide the summary measures used in this study. The methods for assuring scorer reliability are detailed in a previous publication31; the results of an early formal scoring reliability study34 have also been published. Each data channel was assigned a quality code grade according to the duration and quality of signals collected, and each study was given an aggregate quality grade based on the overall interpretability and duration of artifact-free signals.31 After the study began, additional codes were developed to allow the scorer to note if problems were encountered in scoring or if the entire sleep period was not captured.
For this study, subjects with exposures or conditions likely to have large effects on sleep architecture other than SDB were excluded from analysis. These exclusion criteria included: current use of medications known to affect sleep (antipsychotics, benzodiazepines, monoamine oxidase inhibitors, and antidepressants); reported alcohol consumption of more than 14 drinks per week; and a report of awakening often or almost always with leg cramps or leg jerks or of awakening almost always with coughing or wheezing, pain in the joints or muscles, or back pain. The first 516 SHHS studies, which were scored before the implementation of scoring reliability codes, were excluded. Analyses of stage distribution were restricted to studies in which no problems assigning sleep stage were noted. Of 5927 PSG studies performed after the implementation of scoring reliability codes, 2275 were excluded for one of the following reasons: 733 individuals (357 women and 376 men) met exclusion criteria for study quality, 860 individuals (456 women and 404 men) met exclusion criteria for medication or alcohol use, and 864 individuals (519 women and 345 men) met exclusion criteria for symptoms with known sleep disruption (the categories are not mutually exclusive). Analyses of sleep efficiency (SLE) were also restricted to those studies in which light calibration data enabled an accurate assessment of the time of "lights off" and the entire period of sleep was considered captured (ie, the recording time included "final morning awakening" and all study time was of acceptable quality to stage sleep).
The percentage time in each sleep stage was calculated based on the total time asleep. Sleep efficiency was defined as the percentage total time asleep, divided by the total time in bed after lights off to the time of final awakening. The ArI was defined as the total number of arousals in sleep, divided by the total sleep time. Cardiovascular disease was defined as affirmative answers to having physician-diagnosed heart failure, myocardial infarction or heart attack, or previous coronary artery bypass or angioplasty. Lung disease was based on an affirmative answer to having physician-diagnosed long-standing bronchitis, asthma, or emphysema. Hypertension and diabetes mellitus were present if the participant answered affirmatively to having been told by a physician that he or she had these conditions. Smoking status was categorized as "never smoker" (having smoked <20 packs of cigarettes in a lifetime), "ex-smoker," or "current smoker." Age and body mass index (BMI), calculated as weight in kilograms divided by the square of height in meters, were categorized into quartiles and RDI into categories that are commonly considered as identifying disease severity levels.
To stabilize the variance and correct for nonnormality, the log-log transformation was used for percentage stage 1, percentage stage 2, percentage stage 3-4, and SLE, and the log transformation for ArI. A transformation was not required for percentage REM. If we let P represent a proportion, then the log-log transformation is y = −log[−log(P)]. This transformation is undefined when P is zero. When this occurs, each value of P is increased by a constant of 0.001.
Analyses included unpaired t tests, contingency table analyses, analyses of variance, multiple linear regression models, and nonparametric techniques. For bivariate analyses of the sleep variables, the Wilcoxon rank sum test was used for binary independent variables, with the Kruskal-Wallis extension used for polychotomous independent variables. When significant effects were observed for polychotomous variables (ie, those examined by quartile), post hoc pairwise tests with a Bonferroni correction were used to test pairwise differences across multiple levels of the covariate. A multiple linear regression analysis was used to predict each stage distribution, SLE, and ArI (dependent variables), possibly transformed, as a function of age, ethnicity, BMI, sex, comorbidity, and RDI. Age, sex, and RDI with all 2-way and 3-way interactions were initially entered into a regression model. Interactions were tested and retained if significant (P<.05); otherwise, they were removed from the model. The main effects of age, sex, and RDI were retained throughout the modeling and reevaluated at the end. Next, the comorbidity variables were considered and were tested in a backward stepwise fashion. Last, BMI was entered into the model. Additional 2-way interactions were again tested. The likelihood ratio test was used to compare nested models, and the Akaike information criterion was used to compare nonnested models. The Akaike information criterion is defined as twice the negative log-likelihood plus twice the number of variables. Adjusted least squares means and 95% confidence intervals were estimated from each final model and transformed back to the original scale for presentation. Analyses were performed using SAS (version 8.1; SAS Institute, Cary, NC).
Characteristics of the study sample and the overall SHHS cohort are presented in Table 1. The sample included approximately equal proportions of men and women and included 22% ethnic minorities. An RDI greater than 5 (median, 4; range, 0-97) was found in 45% of the sample. Approximately 10% of the sample met the criteria for CVD and lung disease, and 34% had a history of hypertension (HTN). These sample characteristics, including the distribution of sleep stages, were comparable to those in the overall cohort, except that the overall cohort included a slightly higher proportion of subjects with chronic health conditions.
The distributions of sleep stages, SLE, and ArI, according to subject characteristics, are shown in Table 2. As summarized herein, consistent differences in sleep quality are especially apparent when comparing men and women and individuals of different ages. However, ethnicity and comorbidity also are associated with differences in several measures of sleep architecture.
Men had evidence of lighter sleep compared with women. Specifically, men had significantly more percentage stage 1 and percentage stage 2, had a higher ArI, and had lower percentage stage 3-4, percentage stage REM, and SLE.
All measures of sleep architecture varied significantly with age. Post hoc tests showed that percentage stage 1 and percentage stage 2 were significantly higher in individuals older than 61 years compared with those 54 or younger. Percentage stage 3-4 was significantly lower in those older than 54 compared with younger individuals. Percentage stage REM did not differ across the first 2 age quartiles, but was significantly lower among individuals older than 61 vs those 61 or younger. The ArI also did not differ between the first 2 age quartiles, but was significantly higher compared with any other combination of older vs younger age quartiles. Sleep efficiency decreased significantly with increasing age; all pairwise comparisons among age quartiles showed significant differences.
Percentage stage 1 and percentage stage 2 were significantly lower and percentage stage 3-4 was significantly higher among those in the lowest BMI quartile (<24.8) compared with individuals in any of the higher BMI quartiles. However, significant differences in these stage distributions were not seen across the 3 highest BMI quartiles. Percentage stage REM did not vary significantly by BMI quartile. Sleep efficiency and ArI were lower in those with a BMI greater than 30.7 compared with those with a BMI of 30.7 or lower, but did not differ among the lower 3 quartiles.
All measures of sleep architecture other than percentage stage REM varied significantly by ethnicity. In particular, American Indians had a higher percentage stage 1 than whites or blacks; a higher percentage stage 2 and lower percentage stage 3-4 than whites, blacks, Hispanics, or Asian Americans; and a lower percentage SLE than Asian Americans. Blacks had a higher percentage stage 2 than whites or Hispanics, a lower ArI than whites, and a lower SLE than Asian Americans.
All measures of sleep architecture other than SLE differed, with evidence of poorer sleep, in individuals with a history of CVD compared with those without this history. The sleep architecture variables that differed significantly in those with a history of stroke were percentage stage 1, SLE, and ArI. Individuals with a history of diabetes mellitus had more percentage stage 1 and percentage stage 2, less percentage stage 3-4 and percentage REM, a lower SLE, and a higher ArI than those without diabetes mellitus. Those with HTN had a higher ArI and lower percentage stage REM and SLE than subjects without hypertension. Sleep architecture variables did not vary in subjects with a history of lung disease.
Current smokers or ex-smokers had significantly more percentage stage 1 and percentage stage 2 than never smokers. Percentage stage 3-4 was highest in never smokers and lowest in current smokers (all smoking status groups differed statistically significantly from each other). Percentage stage REM was statistically higher in current smokers compared with ex-smokers. Sleep efficiency was significantly lower and ArI was higher in ex-smokers compared with never smokers.
Table 3 shows the variation of sleep architecture according to commonly used clinical cutoff values of RDI (≤5, >5 to ≤15, >15 to ≤30, >30 to ≤50, and >50).35 In men and women, increasing RDI was associated with large increases in the ArI. In men, increasing RDI was associated with increased percentage stage 1 and percentage stage 2, decreased percentage stage 3-4 and percentage stage REM, and decreased SLE. In women, increasing RDI was associated with a decrease in percentage stage REM and SLE, but changes overall were less than those observed in men.
For men and women, individuals in the highest category of RDI (>50; women, n = 11 and men, n = 36) tended to have a higher ArI than those with lower RDIs. Other measures of sleep architecture suggested poorer sleep quality among men with an RDI greater than 50 compared with men with lower RDIs. However, no differences were statistically different for subjects in the RDI categories greater than 30 to 50 compared with greater than 50 for men, women, or all subjects combined. Therefore, in further analyses, we combined the 2 highest levels of RDI categories.
Given the marked variation of sleep architecture variables with demographic factors and chronic conditions, many of which vary between each other and with SDB, multiple linear regression models were fit. The results of these analyses are summarized in Table 4, which shows the adjusted least squares means for significant effects for each measure of sleep architecture.
After adjusting for each of the other factors, significant effects were observed for ethnicity (with post hoc tests, American Indians had significantly higher percentage stage 1 than blacks [P = .01] or whites [P<.001]), smoking status (current smokers had higher percentage stage 1 than ex-smokers [P<.005] or never smokers [P<.001]), stroke (higher percentage stage 1 was found in those with a stroke history, P = .005), and RDI (those with RDIs >5 to ≤15 had more percentage stage 1 than those with RDIs ≤5 [P<.001]). A significant interaction was observed between age quartiles and sex (P = .004). Examination of the age × sex interaction revealed that within each quartile of age men had more percentage stage 1 than women. With increasing age, men had more percentage stage 1; no age effect was apparent for women (Figure 1). After adjusting for these factors, percentage stage 1 was not associated with BMI, CVD, diabetes mellitus, or HTN.
In the final model for percentage stage 2, the main effect for ethnicity was significant (P<.001), and interactions were found between age quartiles and sex (P<.001) and between BMI quartiles and RDI levels (P = .02) (Table 4). Post hoc tests showed that American Indians had significantly higher percentage stage 2 than Hispanics (P<.001) or whites (P<.001) and that blacks had higher percentage stage 2 than whites (P<.001) or Hispanics (P = .02). Within each age quartile, men had significantly higher percentage stage 2 than women; an increase in percentage stage 2 with advancing age was observed only among men (Figure 2). The overall effects of BMI showed that percentage stage 2 was lowest among individuals in the second BMI quartile (>24.8 to ≤27.4) compared with the lowest or higher BMI quartiles. Increasing percentage stage 2 was observed with increasing RDI. Because of a BMI and RDI interaction, further assessment of BMI and RDI effects showed that the relative difference in percentage stage 2 between the lowest and highest RDI levels was largest for subjects with BMIs in the lowest quartiles (data not shown).
Percentage stage 3-4 findings varied significantly with ethnicity and smoking status (P<.001 for both). Post hoc tests showed that American Indians had significantly less percentage stage 3-4 than any other ethnic group. Blacks had significantly less percentage stage 3-4 than whites or Hispanics. Current smokers also had significantly less percentage stage 3-4 than ex-smokers or never smokers (P<.001 for both). In addition, significant interactions were observed between age quartiles and sex (P<.001) (Table 4) and between BMI quartiles and RDI levels (P = .01). Percentage stage 3-4 was significantly lower in men within each age stratum and decreased with increasing age in men but not in women (Figure 3). Percentage stage 3-4 was highest among individuals with an RDI of 5 or less compared with individuals with higher RDI levels. However, within BMI quartiles, consistent changes in percentage stage 3-4 across the 3 highest RDI quartiles were not evident (data not shown).
When initially modeling using data from men and women, a significant 3-way interaction between age, RDI, and sex (P = .02) was observed for percentage stage REM. Therefore, to improve interpretability, the results for the model for percentage stage REM are shown stratified by sex (Table 4). In the final model for percentage stage REM in men, the main effects for age (P = .003) and RDI (P<.001) were significant, with decreasing percentage stage REM observed with increasing age and increasing RDI. In the final model for percentage stage REM in women, the main effect for age (P<.001) was significant, with a decrease in percentage stage REM with increasing age. In women, a significant RDI and BMI interaction was observed such that a decrease in percentage stage REM with increasing RDI was only observed in women in the lowest BMI quartile (percentage stage REM, 20.8 for those with a BMI ≤24.8 and an RDI ≤5; and percentage stage REM, 17.3 for those with a BMI ≤24.8 and an RDI >30).
The ArI was significantly predicted by age (P<.001), sex (P<.001), RDI (P<.001), and BMI (P = .001) and nearly significantly by ethnicity (P = .052). No interactions were significant. Specifically, the effect of RDI level is similar in men and women. Post hoc tests showed that the ArI was significantly lower in those 54 years or younger compared with all older subjects. Individuals with a BMI of 24.8 or lower had a significantly higher ArI than those with a BMI greater than 27.4. Whites had a significantly higher ArI than blacks (P = .03). All pairwise comparisons between RDI categories were significant, with any given RDI category being significantly different from all lower RDI categories.
In the final model for SLE, the effects for age (P<.001), sex (P<.001), RDI (P<.003), ethnicity (P = .008), and HTN (P<.008) were significant. Sleep efficiency was lower in men than in women and in those with HTN. No interactions were significant. Post hoc comparisons showed that all age categories other than the oldest 2 were significantly different from each other, with lower SLEs found in older vs younger age groups. No pairwise difference between ethnic groups was significant. Higher SLE was observed in those with RDIs of 5 or less than in those with RDIs greater than 15 to 30.
Table 5 shows the partial r2 values for each variable in each model. Sex explained the highest proportion of the variance for each sleep stage variable. In particular, sex explained nearly 15% of the variance in percentage stage 3-4 and 11% of the variance in percentage stage 2. Other covariates (age, ethnicity, BMI, and smoking) explained less than 5% of the variance for percentage stages 1, 2, and 3-4 and REM. In general, the RDI explained 1% to 4% of the variance in each model for each sleep stage. In contrast, the RDI explained 17% of the variance in the ArI.
Analyses were repeated using the BMI as a continuous variable (log transformed) rather than quantified according to quartiles. No substantive differences in any of the models were noted.
In this study, we used a database of standardized PSG data, coupled with covariate data describing host and environmental factors that may affect sleep, to assess sleep architecture in a large, ethnically and geographically diverse sample of adults aged 37 to 92 years. Because of the high prevalence of SDB in the population, and the growing use of PSG to assess this condition, it was of specific interest to assess the extent to which sleep architecture is altered by SDB independent of the effects of age and sex. Analyses were limited to subjects free from exposures to large quantities of alcohol or to drugs known to affect sleep architecture and who did not report frequent sleep awakenings due to leg cramps, leg jerks, coughing, wheezing, or pain. Findings should thus be broadly generalizable to healthy middle-aged and older community residents. Unlike most other studies, the available covariate data and the large sample size permitted an analysis of the extent to which various host and other risk factors independently and interactively predicted sleep architecture.
The summary measures of sleep quality varied with sex, ethnicity, and age. After adjusting for all other significant factors, poorer sleep architecture—as measured by increased percentage stages 1 and 2, decreased percentage stages 3-4 and REM, lower SLE, and a higher ArI—was found in men compared with women. In fact, for all measures other than the ArI, sex explained the largest proportion of the variance in each sleep architecture measure. In the overall sample, increasing age was associated with indices of poorer sleep architecture, as measured by decreasing percentage stage REM and SLE and increasing ArI. A reduction in percentage stage 3-4 and an increase in percentage stages 1 and 2 with increasing age, however, was observed in men but not in women. American Indians had higher percentage stages 1 and 2 and lower percentage stage 3-4 sleep than all other groups, while blacks had more percentage stage 2 and less percentage stage 3-4 than whites or Hispanics. In contrast, ArI was highest among whites.
Sex differences in sleep architecture have been suggested previously. Studies, generally of samples of less than 150 individuals, also have reported a relative increase in stage 1 sleep in men27,28,36 and an increase in stage 3-4 in women.26,27 Inconsistent effects of sex on REM sleep have been reported previously.36 These studies, however, did not control for SDB, which is increased in men and could have partly explained the sex differences noted. The present findings demonstrate sex differences in sleep architecture that are independent of RDI, age, ethnicity, BMI, and the comorbidity due to CVD, HTN, diabetes mellitus, and stroke. Specifically, female compared with male SHHS participants had, on average, approximately 23% less percentage stage 1, about 12% less percentage stage 2, approximately 5% more percentage stage REM, and, most striking, 106% more percentage stage 3-4. In addition to the observed sex differences in the percentage time in given sleep stages, women spent a longer period asleep than men (mean ± SD, 6.05 ± 1.00 vs 5.74 ± 1.00 hours, P<.001).
The finding of better sleep quality as assessed by objective measures contrasts with previous studies that have shown greater perceived sleep difficulties in women compared with men. Women report more problems initiating and maintaining sleep and have been reported to use hypnotics at a greater frequency than men.37,38 We may have overestimated sleep quality in women by excluding subjects with current hypnotic use or those with frequent sleep complaints. Although the sleep stage distribution in the sample analyzed was comparable to that of the larger SHHS cohort, a somewhat higher proportion of women than men met any of our exclusion criteria (61% of women vs 56% of men were excluded, P<.001). Among the records excluded because the participant had reported the use of medication, excessive alcohol consumption, or frequent awakenings, a slightly larger number were from female (n = 961) than male (n = 744) participants. We therefore assessed whether similar sex differences in sleep architecture were seen in subjects excluded from the analyses. These analyses showed similar sex differences; for example, the mean ± SD percentage stage 3-4 was 21.0 ± 11.8 vs 12.5 ± 10.5 among the women and men, respectively, who were excluded because of medication use, alcohol consumption, or symptom exclusionary criteria. This suggests that our findings were not biased by preferentially excluding women with poorer sleep from the analyses. We also assessed whether hormone therapy may have explained the sex differences in percentage stage 3-4. Although 29% of women reported use of progesterone, estrogen, or combination hormone medication, adjusting for this did not alter the findings (data not shown). The reasons for better sleep in women than men could have potentially related to unmeasured confounders that negatively affect sleep architecture and preferentially affect men. However, this seems unlikely because we adjusted for most key comorbidities. In addition, the smaller studies37,38 that reported sex differences were predominantly studies of healthy volunteers.
In the entire study sample, across the age range of 37 to 92 years, age was negatively and linearly associated with all measures of sleep architecture other than percentage stage 1. These associations were independent of potential confounders (which also vary with age) such as obesity, SDB, and comorbidity; however, marked differences in the effects of age were observed in men and women, with an effect of age on percentage stages 1, 2, and 3-4 observed for men but not for women. This is shown in Figure 1, Figure 2, and Figure 3, in which a progressive reduction in percentage stage 3-4 and a progressive increase in percentage stages 1 and 2 are seen with increasing age in men, with little age-related change in women. Among men, percentage stage 2 was 6% higher and percentage stage 3-4 was 52% lower in subjects in the highest compared with the lowest age quartiles. In contrast, no consistent variation of percentage stages 1, 2, or 3-4 with age was observed in women.
The change in sleep architecture with aging, most prominently the reduction of percentage stage 3-4 in men, has been reported previously23,27,39,40; however, similar to the studies of sex effects, the extent to which changes could have related to underlying comorbidity or to SDB has been unclear. Bliwise36 has summarized the literature on aging and sleep and has suggested that a reduction in percentage stage 3-4 may be a sensitive biomarker of aging. Possible reasons for reduced percentage stage 3-4 may include age-related reductions in cortical mass, cortical metabolism, or neurotransmitter levels; changes in circadian rhythm; or other neuroendocrinological or nervous system activity.23,36 The somatotropic axis (as assessed by secretory patterns of growth hormone and insulin-like growth factor 1) and control mechanisms that affect stage 3-4 sleep appear to be highly integrated systems that are interactive and affected by common neuroendocrinological control mechanisms.41- 43 Age-related changes in the somatotropic axis are well described.40,42 A recent study40 of 149 men reported a decrease in percentage stage 3-4 from 18.9% among male subjects aged 16 to 24 years to 3.4% among men aged 36 to 50 years, changes that paralleled decreases in growth hormone levels. Our data suggest the need to further investigate the interrelationships among age, neuroendocrine function, circadian rhythm, and sleep homeostasis, and how such systems may be differentially changed with age in men and women. Understanding the basis for the marked differences between men and women in sleep architecture with age could provide insight into how biological (eg, hormonal factors) vs environmental (eg, sleep habits) factors may modulate neurophysiologic systems.
We postulated that obesity could negatively affect sleep architecture via effects on SDB, general effects on overall health and fitness, or a direct effect of obesity on sleep quality. In unadjusted analyses, percentage stages 1, 2, and 3-4 varied with BMI quartile, with evidence of poorer sleep among heavier individuals. However, after adjusting for the RDI, age, and sex, BMI level explained less than 1% of variance in any sleep summary measure. Unexpectedly, adjusted analyses suggested a lower ArI in more obese individuals. It is possible that the subjects in this sample who had the lowest BMI had other health conditions or other factors that caused frequent arousals that were not identified in this study. Our data, however, do not suggest that moderate levels of obesity, independent of the effects of RDI, adversely affect sleep architecture.
This is the first large-scale study, to our knowledge, that explores the variation of sleep architecture across ethnic groups. One recent study,44 which included 17 blacks, reported that this group had more stages 1 and 2 and less stage 3-4 than members of other ethnic groups. In the present study, after adjusting for other factors, blacks and American Indians had lighter sleep than whites, Asian Americans, or Hispanics. However, differences among ethnic groups, although statistically significant, were small and of unclear clinical significance. Furthermore, in all models, most of the variance in sleep architecture variables was unexplained by measured covariates. Therefore, it is unclear whether ethnic differences were due to innate genetic differences in sleep mechanisms or to unmeasured environmental effects, including psychosocial disruption, environmental noise, long-standing sleep deprivation, alcohol or other drug use, or unmeasured comorbid illness. Our analyses were restricted to individuals who reported drinking less than 14 alcoholic drinks per week. To further assess whether lesser quantities of alcohol could have affected the study findings, we repeated all final models with an additional term for self-reported alcohol use (available for n = 2500). These analyses showed significant effects for alcohol use on percentage stages 1 and 3-4. Adjustment for alcohol use did not, however, appreciably affect the estimates for ethnic effects.
Nicotine is a known stimulant, with hypothesized negative effects on sleep quality.45 Consistent with such an effect was the approximately 26% reduction in stage 3-4 in current smokers compared with ex-smokers or never smokers. This finding supports efforts to improve sleep hygiene that call for decreased smoking before bedtime.
After considering these host and environmental factors, we were able to quantify the independent associations of RDI level with sleep architecture. Our analyses showed that sleep stage distributions varied with RDI level, with increased fragmentation (an increased ArI) and increased percentage stages 1 and 2 and decreased SLE and percentage stage 3-4. These associations were independent of the effects of age, BMI, sex, ethnicity, and comorbidity on sleep architecture. Other than the ArI, the variation of sleep architecture with RDI was generally stronger in men than in women. For percentage stages 2 and 3-4, RDI and BMI interactions were observed, with larger effects of RDI observed for individuals in the lowest BMI quartile. Estimated differences in percentage stages 1, 2, 3-4, or REM for subjects in higher vs lower RDI categories were less than 25% (and often negligible). In contrast, a 126% higher ArI was observed in subjects with RDIs greater than 30 compared with those with RDIs of 5 or less.
Except for ArI, the overall associations of RDI on sleep stage distribution are comparable to or smaller than the observed effects of sex or age (explaining 1%-4% of the variance in the measures). In assessing the potential effect of SDB on sleep architecture, it is therefore important to consider the potential for confounding by age and sex. Overall, the lesser effect of RDI on sleep architecture as quantified in this study vs common clinical impressions may be due to differences in studying participants in community-based research compared with patients referred to a clinical laboratory. Not only are the latter likely to have more severe SDB but, because they may be specifically referred for evaluation for complaints of poor sleep, they may represent a group that is particularly sensitive to the sleep-disrupting effects of SDB. Patients, especially those with a greater RDI, may have a higher level of susceptibility to the sleep-disrupting effects of SDB than nonclinical populations with SDB.
It has been reported that measurements of ArI are less reliable than those of sleep stages.34,46 Indeed, in our sample, approximately 15% of records had EEG signals that were of insufficient quality to reliably score arousals. Notwithstanding this potential limitation, the ArI varied substantially with RDI independent of the associations of other demographic or health variables. This suggests that the ArI may be the most sensitive measure of the sleep-disrupting effects of SDB. Ongoing research aimed at automating arousal detection or using alternative approaches for identifying changes in EEG spectra47 may provide new tools to improve the use of arousal measurement in assessing the effects of SDB and other host factors on sleep quality.
We limited analyses to subjects with reliable EEG and electromyographic data and to subjects free of exposures expected to have a major effect on sleep architecture. The remaining sample is the largest community-based sample in which standardized PSG has been performed and contained abundant statistical power to detect even small effects and subtle interactions. The generally similar characteristics of the included and entire SHHS sample suggest that effects due to subject exclusion are probably small. Furthermore, previous analyses of SHHS data suggest there is little bias due to excluding studies of less than optimal quality.48 While medication and symptom-based exclusions may have slightly exaggerated the sex differences between men and women, the same pattern of sex differences in sleep architecture was seen in excluded subjects, as discussed herein. It should be emphasized that subjects were studied at a time when they were considered stable medically, generally at least 4 weeks after an acute illness or hospitalization. Furthermore, the SHHS cohort did not include individuals receiving home oxygen therapy, a tracheostomy, or nasal continuous positive airway pressure. Therefore, there may be a larger effect of RDI or of chronic health conditions, and less of a sex difference, in other populations that include more symptomatic or sicker individuals.
A limitation of this study was the lack of objective data regarding leg movements. Although we excluded individuals with reports of frequent leg movements, it is possible that unrecognized periodic limb movements may have partly explained some of the effects attributed to age or sex. Also, data were based on studies performed in the home rather than in a laboratory. Sleep architecture assessed with in-home studies may include slightly more REM and less stage 1 sleep than studies performed in a laboratory (Conrad Iber, MD, unpublished SHHS data, 2003).
The overall effect of chronic health conditions on sleep quality has been debated. Although this article was not designed to critically look at the effects of comorbidity on sleep architecture, we assessed the extent to which major common comorbidities postulated as being associated with SDB may have explained the effects of demographic and SDB variables on sleep architecture. Although sleep architecture varied with chronic illnesses in our unadjusted analyses, these associations decreased after the effects of age, SDB, and sex were considered. Therefore, in this generally healthy population, only a prior history of stroke (relative to effects on stage 1) and a history of HTN (relative to effects on SLE) significantly predicted sleep architecture. Other than for stroke and HTN, our data do not suggest a significant impairment of sleep due to chronic illnesses among community residents studied at a time when their health was at a stable level. These analyses, however, were restricted to subjects without frequent nocturnal awakenings and, thus, might underestimate effects of chronic health conditions of sufficient severity to overtly affect sleep continuity. Antihypertensive drug use was not an exclusionary criterion. Therefore, we cannot ascertain the extent to which the HTN effects could have been secondary to medication use.
In summary, this study demonstrates significant effects of age, sex, ethnicity, and SDB on sleep architecture. Overall, sex accounted for larger differences in sleep measures than other covariates. Similarly, an adverse effect of increasing age on sleep architecture was more evident in men than in women. We also report for the first time small variations of sleep architecture with ethnicity, with evidence of lighter sleep in American Indians and blacks compared with other ethnic groups. Some evidence for a deleterious effect of current smoking on sleep architecture also was found. Chronic illnesses not severe enough to cause obvious nocturnal awakenings did not appreciably affect sleep architecture after accounting for other associations. After adjusting for these various host and environmental factors, RDI level adversely affected all measures of sleep architecture; however, with the exception of the ArI, these associations were of a comparable or lesser magnitude than age or sex.
Corresponding author and reprints: Susan Redline, MD, MPH, Division of Clinical Epidemiology, Department of Pediatrics, Rainbow Babies and Children's Hospital, 11100 Euclid Ave, Cleveland, OH 44106 (e-mail: email@example.com).
Accepted for publication March 25, 2003.
This study was supported by grants U01 HL63463 (Case Western Reserve University), U01 HL53940 (University of Washington), U01 HL53941 (Boston University), U01 HL53938 (University of Arizona), U01 HL53916 (University of California, Davis), U01 HL53934 (University of Minnesota), U01 HL53931 (New York University), and U01 HL53937 (The Johns Hopkins University) from the National Heart, Lung, and Blood Institute, National Institutes of Health, Bethesda, Md.
This study was presented in part in abstract form at the American Thoracic Society; April 25, 1999; San Diego, Calif; and the American Professional Sleep Association; June 20, 1998; New Orleans, La.
We acknowledge the assistance and support of the SHHS participating institutions, investigators, and research assistants. A full list of participating institutions and investigators can be found at http://www.jhucct.com/shhs/. The SHHS acknowledges the Atherosclerosis Risk in Communities Study, the Cardiovascular Health Study, the Framingham Heart Study, the Cornell Worksite and Hypertension Studies, the Strong Heart Study, the Tucson Epidemiology Study of Airways Obstructive Diseases, and the Tucson Health and Environment Study, for allowing their cohort members to be part of the SHHS and for permitting data acquired by them to be used in the study.