INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Excessive daytime sleepiness is a relatively common complaint among patients with sleep-disordered breathing (SDB). It is associated with several adverse effects, including impaired daytime performance, neuropsychological dysfunction, and an increased risk for motor vehicle accidents (1). The multiple sleep latency test (MSLT) is a widely used and standardized method for quantifying the degree of daytime sleepiness (2). It usually consists of four to five nap opportunities equally spaced at 2-h intervals during the day. With each nap, the latency between "lights out" and sleep onset is determined and the mean or median sleep latency from the naps is used to grade the severity of daytime sleepiness.

While several studies (3) have previously examined the relationship between the MSLT and the results of the nocturnal polysomnogram (PSG) in SDB, the etiology of daytime sleepiness in these patients is not well understood. Initial reports modeling the sleep latency as the outcome variable suggested that the fragmentation of sleep associated with the occurrence of apneas and hypopneas may be the primary cause of daytime sleepiness (5). Physiologic correlates of sleep fragmentation such as the number of arousals and the distribution of sleep stages were found to be significant predictors for the occurrence of daytime sleepiness in SDB. In these studies, parameters reflecting nocturnal hypoxemia provided little or no added predictive power. However, more recent data indicate that nocturnal hypoxemia may be the primary determinant for the occurrence of daytime sleepiness in SDB (3, 4, 8). This discrepancy regarding the etiology of daytime sleepiness in SDB can be explained by two factors. First, because sleep fragmentation and nocturnal hypoxemia often coexist in the presence of significant apnea, it is not surprising that the specific cause has been difficult to clarify. Second, the use of correlation and regression methods in the available studies may be an inefficient approach to the analysis of the MSLT. Because time to sleep onset as measured during the MSLT represents "failure-time" data, survival analysis is a more appropriate method for examining the determinants of sleep latency in SDB. Therefore, the present study was designed to examine the independent effects of apnea-hypopnea frequency, nocturnal hypoxemia, and sleep fragmentation on daytime sleepiness in SDB using the technique of survival analysis. We hypothesized that the apnea-hypopnea frequency, degree of nocturnal hypoxemia, and sleep fragmentation would independently contribute to an increased risk of daytime sleepiness in patients with SDB.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Patient Selection

Using a retrospective cohort design, we identified all patients (age  18 yr) who had nocturnal polysomnography performed at the Johns Hopkins Sleep Disorders Center from November 1995 through January 1998. From this group, we selected those patients who had SDB as one of their sleep-related diagnoses. SDB was diagnosed if the apnea- hypopnea index (AHI) on the overnight PSG was greater than or equal to 10 events per hour. Patients studied on continuous positive airway pressure (CPAP) or for clinical follow-up were excluded from enrollment into the study sample. Additional exclusionary criteria included split-night (< 6 h) or extended (> 9 h) polysomnographic recording, and participation in ongoing clinical research protocols. Moreover, patients requiring supplemental oxygen during any part of the nocturnal study were also considered ineligible, because associations of daytime hypersomnolence with oxyhemoglobin saturation were of primary interest in this study.

Of the 1,667 patients who had an overnight PSG in our laboratory from November 1995 through January 1998, 247 patients did not meet the inclusion criteria for the current investigation (Figure 1). Fifty-one patients were excluded because their PSGs were done for clinical follow-up, 128 had either a split-night or extended protocol, 37 required supplemental O₂, 18 were studied with CPAP, and 13 were participants in ongoing clinical research studies. Although a total of 1,420 patients were considered eligible, 338 did not have a MSLT and were therefore excluded. An additional 341 patients were also excluded from the analysis because of an AHI < 10 events/h, resulting in a final cohort of 741 patients with SDB as one of their sleep-related diagnoses (AHI  10 events/h).

View larger version (25K): [in this window] [in a new window]   Figure 1.   Flowchart showing the selection process for the study sample.

Polysomnography and Daytime MSLT

The nocturnal PSG consisted of continuous polygraphic recordings (Model 78d; Grass Instruments, Quincy, MA) of a modified electrocardiographic (V₆) lead, right and left electro-oculographic leads, submental and bilateral anterior tibialis surface electromyograms, and two electroencephalographic leads (C3-A2, C3-O1). Respiration was monitored throughout the night with thermocouples (Protech, Woodinville, WA) at the nose and mouth, and with thoracic and abdominal strain gauges (Piezo-crystal; EPM Systems, Midlothian, VA). Continuous recording of the oxyhemoglobin saturation (Sa_O2) was obtained with an oximeter (model 3700; Ohmeda, Englewood, CO). Sleep-stage scoring was performed on 30-s epochs by a team of trained technicians according to the criteria of Rechtschaffen and Kales (9) and subsequently reviewed by one of the authors (A.R.S. or P.L.S.). Apnea was defined as complete cessation of airflow for at least 10 s. Hypopnea was defined as a reduction in the airflow which was 10 s in duration and was associated with an electroencephalographic arousal or a drop in the Sa_O2 ( 4%). The apnea-hypopnea index (AHI) was defined as the total number of apneas and hypopneas per hour of total sleep time.

When clinically indicated, an MSLT was undertaken on the day after the nocturnal PSG. The MSLT consisted of a series of four 20-min nap trials at 2-h intervals. Patients were instructed at 0900, 1100, 1300, and 1500 h to lie down on a bed in a quiet, darkened bedroom and were allowed to sleep. The recording montage during the MSLT was similar to the one used for the nocturnal study. Each nap trial lasted 20 min if sleep did not occur. If the patient did fall asleep within 20 min, the trial was terminated 15 min after sleep onset. Between naps the patients were instructed not to sleep and were monitored by trained technicians. The sleep latency for each nap trial was defined as the time to the first 30-s epoch composed of at least 15 s of sleep stage 1. If no sleep occurred during a nap trial, a value of 20 min was recorded. Given the truncated distribution of sleep latency from each nap, the median time was used to characterize the level of daytime sleepiness.

Description and Coding of Variables

The medical records of the final study cohort were reviewed and the following demographic information was collected on each patient: age, sex, and body mass index (BMI). Information abstracted from each patient's nocturnal PSG included: time in bed (TIB), total sleep time (TST), time in sleep stages 1 and 2 (Stg1 and Stg2), time in slow wave sleep (SWS), time in rapid eye movement (REM) sleep, and AHI. Baseline and average drop in Sa_O2 (Sa_O2) during the night were also noted for each patient. Sleep latency from each nap trial and the resultant median was extracted from the MSLT.

For the purpose of this analysis, age and BMI were studied as both continuous and categorical variables. The following age categories were used: < 40, 40 to 64.9, and  65 yr. BMI was classified into five categories as follows: < 25, 25 to 29.9, 30 to 34.9, 35 to 39.9, and  40 kg/m². Similarly, parameters from the nocturnal PSG were also modeled as continuous and categorical variables. AHI was grouped according to the underlying apnea severity: 10 to 29.9, 30 to 59.9, and  60 events/h. Baseline Sa_O2 was assessed in four categories: < 85, 85 to 89.9, 90 to 94.9, and  95%. Severity of oxyhemoglobin desaturation (Sa_O2) during the night was categorized as  5%, 5.1 to 10%, 10.1 to 15%, and > 15%. Sleep architecture variables (TST, Stg1, Stg2, SWS, REM) were grouped in tertiles based on the distribution of the respective variables in the final study cohort. Time in each sleep stage was also expressed as a percentage of TST (Stg1%, Stg2%, SWS%, and REM%). To avoid the rank order assumption within each categorical variable, indicator variables were used with the lowest strata of each variable as the reference category.

Statistical Analysis

The dependent variable of interest was the median time to sleep onset from the MSLT. We reviewed clinical and polysomnographic factors that previously have been shown to be associated with daytime hypersomnolence as assessed by the MSLT. These factors include: age, sex, BMI, AHI, oxyhemoglobin desaturation, amount of time in sleep stages 1 and 2, time in SWS, and TST. Bivariate associations between these factors and the median sleep latency were initially examined using the Kaplan-Meier product-limit analysis. Because the primary objective of our study was to examine the determinants of significant daytime sleepiness, a median sleep latency of 10 min was used for the purpose of censoring observations. The log-rank test and the generalized Wilcoxon test were used to determine the statistical significance of differences in survival curves across ordered categories of each risk factor. Cox's proportional hazards regression was then used to assess the unadjusted relative risks (RR) and the corresponding 95% confidence intervals (95% CI) for the association between hypersomnolence and each of the above mentioned factors.

To define the independent contribution of each risk factor, we performed multivariate analyses with the stepwise addition of covariates to the proportional hazards regression model. For the stepwise modeling, a forward selection procedure was employed in which the addition of covariates was guided by our primary hypotheses regarding the etiology of daytime hypersomnolence and also by the findings from our bivariate analyses. The likelihood ratio test was used at each stage to compare models with and without the particular covariate. The presence of interaction between the factors of interest was also explored by the inclusion of cross-product terms into the multivariate model.

To examine the independent effect of sleep stage distribution on hypersomnolence, each sleep stage (expressed as a percentage of TST) was included in the multivariate model with the exception of Stg1%. Because the sum of the individual sleep stage percentages is equal to unity, the inclusion of all sleep stages in the multivariate model would lead to an indeterminate solution. Moreover, exclusion of Stg1% from the multivariate model, which includes the other sleep stages, provides a reference category (sleep stage 1) for interpreting the model coefficients associated with the other sleep stages. For example, the model coefficient associated with Stg2% would be interpreted as the RR for hypersomnolence associated with a given increase in Stg2% (and therefore an equivalent decrease in Stg1%) after adjusting for all other variables in the model. Sleep efficiency (TST/TIB) was not used in the modeling process because TST and TIB were included as predictor variables.

The statistical significance of all RR was determined by the two- sided test of the beta coefficient with a p value of 0.05. Although continuous and categorical forms of each variable were used in the model building process, results from the categorical analyses are presented for ease in interpretation. The proportional hazards assumption was tested for each factor of interest and was not violated. All statistical analyses were conducted using the STATA statistical software package (STATA Inc., College Station, TX).

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Patient Characteristics

Clinical characteristics and PSG data for the final study sample (n = 741) are summarized in Table 1. There were 576 men (77.7%) and 165 women (22.3%) with mean age of 51.7 ± 13.4 yr. Because this was a clinic-based sample, it was not surprising to find that the group was generally obese with an average BMI of 35.3 ± 8.5 kg/m² and had evidence of moderate to severe SDB with an AHI of 47.6 ± 29.3 events/h. There were approximately equal numbers of patients within each category of apnea severity: 260 patients with an AHI of 10 to 29.9, 251 patients with an AHI of 30 to 59.9, and 230 patients with an AHI  60 events/h. The baseline Sa_O2 for the sample was 96.0 ± 2.2% and the average drop in Sa_O2 during the night was 5.7 ± 4.1%. Results of the MSLT showed evidence of significant daytime hypersomnolence in these patients with a mean sleep latency of 6.4 ± 4.9 min. Five hundred eighty-nine patients (79.5%) had a median sleep latency of less than 10 min.

A comparison of the characteristics of patients with (n = 1,082) and without (n = 338) a MSLT in the sample of eligible patients (n = 1420) showed no differences in the clinical characteristics of the two groups. Specifically, we found no significant differences in the mean age (49.9 versus 49.6 yr), BMI (34.2 versus 33.4 kg/m²), or gender distribution (69.6 versus 65.1% men) between patients with and without a MSLT. Moreover, the AHI (33.9 versus 31.9 events/hr) was similar between the two groups.

Bivariate Analysis

The Kaplan-Meier survival curves by AHI category are presented in Figure 2. The cumulative probabilities of having a median sleep latency of less than 10 min for the three apnea categories were 71.1%, 78.9% and 89.6%, respectively, indicating that as the apnea-hypopnea frequency increases the likelihood of falling asleep within 10 min during the MSLT progressively increases (p < 0.001 log-rank test). There was an inverse association between average Sa_O2 recorded during the night and the severity of daytime sleepiness (Figure 3). Ninety-six percent of the patients with a Sa_O2 > 15% and 74.8% of the patients with a Sa_O2  5% had a median sleep latency of less than 10 min. As shown in Figure 3, patients with a Sa_O2 in the range of 5 to 15% had an intermediate risk of falling asleep compared with the other two groups.

View larger version (16K): [in this window] [in a new window]   Figure 2.   Kaplan-Meier survival curves by AHI category.

View larger version (17K): [in this window] [in a new window]   Figure 3.   Kaplan-Meier survival curves by SaO2 category.

To examine the effect of altered sleep architecture on daytime somnolence in SDB, Kaplan-Meier survival curves were also constructed for the amount of time spent in each sleep stage during the night (Figure 4). Because patients with significant apnea have repeated disruption in the continuity of their sleep, they often display an abundance of "light" sleep, evidenced by increased sleep stage 1. It was hypothesized that an increase in sleep stage 1 would be associated with an increased risk for daytime hypersomnolence, whereas an increase in stage 2 or SWS would be associated with a decreased risk. Moreover, it was hypothesized that there would be no risk associated with the amount of REM sleep. As shown in Figure 4A, patients in the highest tertile of Stg1% were significantly more sleepy compared with those patients in the lower two tertiles (p < 0.001 log-rank test). An increase in the amount of Stg2% or SWS% was inversely correlated with the risk of hypersomnolence (Figures 4B and 4C). No significant association was noted between the amount of REM sleep and the risk of daytime hypersomnolence. TST, however, was noted to be associated with an increased risk with patients in the highest tertile of TST being at the highest risk for having a median sleep latency of less than 10 min (Figure 4D).

View larger version (25K): [in this window] [in a new window]   View larger version (22K): [in this window] [in a new window]   Figure 4.   Kaplan-Meier survival curves by sleep stage: (A) stage 1, (B) stage 2, (C ) SWS, and (D) TST.

We also examined the association between clinical variables and the risk of hypersomnolence using Kaplan-Meier survival curves (not shown). Younger age (< 65 versus  65 yr) and obesity (BMI  30 versus < 30 kg/m²) were found to be associated with an excess risk. Eighty-one percent of patients in the younger age group (< 65 yr) had a median sleep latency of less than 10 min compared with 74.1% of the patients in the older group ( 65 yr). Similarly, patients with a BMI  30 kg/m² were observed to have an increased risk with 82.5% having a median sleep latency of less than 10 min. In contrast, 71.8% of patients with a BMI of < 30 kg/m² had a median sleep latency of less than 10 min. Male sex in this clinic- based sample was not associated with an added risk of hypersomnolence.

Cox proportional hazards regression was used to describe and quantify the influence of the above clinical and polysomnographic variables on the risk for hypersomnolence. The bivariate results from the proportional hazards modeling of the data are presented in Table 2. Patients with an age  65 yr had a reduced RR for hypersomnolence in comparison to patients with an age less than 40 yr (reference category). The reduction in risk associated with increasing age was only true for the older age group ( 65 yr) and achieved borderline statistical significance (RR: 0.78, p < 0.07). Gender did not appear to affect the risk of falling asleep in the first 10 min during the MSLT. BMI ( 30 kg/m²) was a significant predictor of hypersomnolence with RR of 1.44, 1.63, and 1.79 in the third, fourth, and fifth BMI categories, respectively.

As expected, AHI was related to sleepiness with an increasing trend in RR. In comparison to the reference category (AHI: 10 to 29.9 events/h), patients with an AHI of 30 to 59.9 events/h had a RR of 1.30, whereas patients with an AHI  60 events/h had a RR of 2.11 for hypersomnolence. There was no apparent relationship between the baseline Sa_O2 and propensity to fall asleep during the MSLT. However, the risk associated with the average Sa_O2 during the night was elevated and showed a significant dose-response relationship. Compared with the reference category of  5%, the RR for average Sa_O2 decreases of 5.1 to 10, 10.1 to 15, and > 15% were 1.43, 2.16, and 4.46, respectively.

The strength of association between variables of sleep architecture and the risk of hypersomnolence was also explored using the proportional hazards model. For a 30-min increase in TST, the unadjusted RR was 1.16, indicating that patients with a higher TST were more likely to have a median sleep latency of less than 10 min. However, because TST is dependent on the time allowed for sleep, the unadjusted RR must be interpreted with caution. There was also a statistically significant association between the amount of time spent in each sleep stage expressed as a percentage of TST. A 10% increase in the amount of Stg1% was associated with a RR of 1.07, whereas a 10% increase of Stg2% or SWS% was associated with a RR of 0.92 and 0.80, respectively. Consistent with the findings from the Kaplan-Meier analysis, the bivariate proportional hazards model did not reveal any significant association between the amount of REM sleep and the degree of hypersomnolence.

Multivariate Analysis

To define the independent effects of nocturnal hypoxemia and sleep fragmentation on daytime hypersomnolence, a variety of multivariate analyses were performed. Each clinical and polysomnographic variable was tested sequentially in a multivariate proportional hazards model. Only those variables that were statistically significant or could potentially confound the relationships of interest were included in the multivariate regression models. Table 3 lists the factors that were significantly predictive with the corresponding RR from the final proportional hazards regression. As in the bivariate analysis, no significant association was observed between gender and the occurrence of hypersomnolence in this clinic-based sample. Although age was not a significant risk factor for hypersomnolence in the multivariate analysis, it was still included in the final model given the potential for confounding influence. When the effect of BMI was examined after adjusting for other variables, the previously observed association with hypersomnolence was virtually eliminated in all BMI categories except for those with morbid obesity. Patients with a BMI  40 kg/m² had an adjusted RR of 1.56 compared with the patients with a BMI < than 25 kg/m².

Of the polysomnographic variables examined in the multivariate analysis, the following factors remained significantly predictive: AHI, Sa_O2, and sleep stage distribution. Compared with the reference group (AHI: 10 to 29.9 events/h), the adjusted RR for categories of AHI were 1.30 for an AHI of 30 to 59.9 events/h and 1.65 for an AHI  60 events/h. Although Sa_O2 also remained inversely associated with the risk for hypersomnolence in the final regression model, the risk was somewhat attenuated. The adjusted RR for categories of Sa_O2, controlling for other factors, were 1.18 for Sa_O2 of 5.1 to 10%; 1.43 for Sa_O2 of 10.1 to 15%; and 1.94 for Sa_O2 > 15%.

With regard to sleep architecture, TST was again observed to be significantly associated with an increased risk of hypersomnolence with a RR of 1.31 for a 30-min increase in TST after adjusting for TIB and all other covariates. In contrast, increases in Stg2% and SWS% were protective from hypersomnolence. No association was noted between hypersomnolence and the amount of REM sleep.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The purpose of this study was to identify clinical and polysomnographic risk factors for the occurrence of daytime sleepiness among patients with SDB. In a group of 741 patients with an AHI  10 events/h, we used survival analysis techniques to assess the determinants of the MSLT, a standard measurement for quantifying the degree of daytime sleepiness. Using a multivariate proportional hazards model, we found several clinical and polysomnographic variables that were predictive of the MSLT median sleep latency. First, our data indicate that indices of nocturnal hypoxemia and sleep fragmentation independently contribute to an increased risk for hypersomnolence in patients with SDB. Second, disease severity, as evaluated by the AHI, was also predictive of an increased risk after controlling for the effects of nocturnal hypoxemia and sleep fragmentation. Third, age and gender were observed not to have a significant influence on the degree of hypersomnolence in this clinic-based sample. Interestingly, patients with morbid obesity (BMI  40 kg/m²) were found to be at increased risk for hypersomnolence even after adjusting for the confounding effects of age, gender, and AHI. The results of this study illustrate that the technique of survival analysis can be used to analyze the results of the MSLT, which measures the time to an event (falling asleep).

Our findings seem to be in contrast with a number of other studies that have previously examined the relationship between the MSLT and results of the overnight PSG. In one of the earliest reports of daytime sleepiness in SDB, Guilleminault and coworkers (6) noted no significant relationship between the severity of underlying apnea (AHI) and the results of the MSLT in a sample of 100 patients with SDB (AHI > 10 events/h). Moreover, indices of nocturnal hypoxemia, such as the frequency of Sa_O2 drops below 80% and the amount of time with a Sa_O2 less than 90%, did not show any significant association with the results of the MSLT. In contrast, variables of sleep architecture were highly correlated with the daytime sleep latency. Patients with shorter sleep latencies were observed on their nocturnal PSG to have an increase in the amount of sleep stage 1 and a decrease in the amount of stage 2, SWS, and REM sleep. These findings led the investigators to conclude that the disturbance in sleep continuity was the primary determinant of excessive daytime somnolence in SDB. Subsequently, Roehrs and colleagues (7) reached a similar conclusion and showed that in a group of 466 patients with SDB, parameters of sleep fragmentation were independently predictive of the daytime sleep latency, whereas measures of nocturnal hypoxemia provided no added predictive power. Additional evidence for the association between sleep fragmentation and daytime sleepiness was provided by a randomized, crossover study (5) in which SDB patients (AHI: 73 ± 26 events/h) treated with CPAP were subjected to periods of experimentally induced nocturnal hypoxemia. Using this approach, Colt and associates showed no significant difference between the MSLT after CPAP treatment under hypoxemic and nonhypoxemic conditions. These data supported the notion that sleep fragmentation may have a dominant role in the pathogenic link between SDB and daytime sleepiness.

More recently, several investigators (3, 4, 8) have suggested that nocturnal hypoxemia may be the primary cause of hypersomnolence in SDB. In two separate reports, Mendelson (3, 4) has shown that measures of oxyhemoglobin desaturation were the most significant predictors of the daytime sleep latency measured during the MSLT. Whereas indices of sleep fragmentation were also observed to predict the results of the MSLT, the addition of variables reflecting hypoxemia in the prediction models eliminated the observed association between sleep fragmentation and daytime sleepiness. In support of these findings, there are further studies (8) suggesting the importance of nocturnal hypoxemia in the causal pathway of hypersomnolence in SDB.

In contrast to previously published work, the results of the current study indicate that sleep fragmentation and nocturnal hypoxemia independently contribute to an increased risk of hypersomnolence in SDB. There is inconsistency in the current literature on the relative importance of sleep disruption and nocturnal hypoxemia in producing the adverse clinical consequences of SDB. Attempts to distinguish the effects of these two factors have been inconclusive. Experimental sleep disruption, without associated hypoxemia, has been shown to produce daytime sleepiness in normal subjects (10, 11). Conversely, hypoxemia in normal subjects (12) has been related to poor sleep quality and subsequent daytime sleepiness. Collectively, these studies would suggest that both sleep fragmentation and hypoxemia can independently contribute to daytime sleepiness and may act synergistically in patients with SDB.

In the present analysis, AHI remained a significant predictor for hypersomnolence after adjusting for the degree of nocturnal hypoxemia and sleep stage distribution. Because both hypoxemia and sleep stage disruption are consequences of disordered breathing events, no association would be expected between AHI and sleep latency after adjusting for these factors. The residual effect of AHI observed in the multivariate proportional hazards model suggests that sleep stage distribution is an imprecise surrogate for the degree of sleep fragmentation that occurs with an apnea or hypopnea. The use of more precise measures of sleep fragmentation, such as the frequency of sleep stage transition, might have eliminated this residual effect of AHI.

Our analysis of the predictors of hypersomnolence in SDB failed to identify age or gender as potential risk factors. These findings are in agreement with several other studies (4, 7, 8) that have examined the relationships between these clinical variables and the MSLT. Age and gender consistently have been shown to have comparatively weak, if any, associations with the daytime sleep latency. This lack of effect of age and gender may be due to the bias introduced by examining a symptomatic sleep clinic population. Studies (13) that have shown significant associations between these variables and either subjective or objective measures of daytime sleepiness have failed to account for potential confounders (i.e., AHI).

An unexpected finding of this study was that morbid obesity (BMI  40 kg/m²) independently predicted an increased risk for hypersomnolence after adjusting for the severity of underlying apnea (AHI, hypoxemia, and sleep fragmentation). The explanation for this relationship is not entirely clear. However, this finding is consistent with a recent report (14) that noted a higher degree of objective daytime somnolence in morbidly obese (BMI 45.4 ± 1.3 kg/m²) patients without any evidence of SDB (AHI < 5 events/h) compared with healthy control subjects without obesity. It has been hypothesized (15) that this increase in sleep propensity in the absence of apnea may be secondary to alterations in the neuroendocrine axis in these individuals, perhaps because of a relative decrease in leptin, or possibly because of elevated levels of plasma cytokines (interleukin-6 [IL-6], tumor necrosis factor alpha [TNF-]). Alternatively, it is possible that the observed association between morbid obesity and hypersomnolence is caused by residual confounding.

Several limitations should be considered in the interpretation of our results. First, because this study was based on review of a polysomnographic database, specific clinical diagnoses for each patient were not known. The inclusion of patients with a primary sleep disorder other than SDB in the study sample introduces a degree of selection bias. However, because clinic-based samples tend to have high prevalence of daytime sleepiness due to a variety of sleep disorders, this bias would drive our results toward the null. Second, although one might suspect that modeling the MSLT using information from the overnight PSG may identify the determinants of daytime sleepiness, this is not necessarily the case. In addition to the continuity of prior sleep and the duration of prior wakefulness, the degree of sleepiness is influenced by several other factors. Circadian rhythms and neurobehavioral and motivational factors, such as the desire to fall asleep or remain awake, can have a significant impact on the propensity of sleep onset (16). These individual variations are not fully captured by the parameters of the nocturnal PSG and thus limit our ability to fully explain the significant variability in the MSLT results.

Nevertheless, this study has several advantages compared with previous studies in this area. To our knowledge, this is the first study that has employed the method of survival analysis to examine the determinants of the MSLT. The use of survival analysis is a significant departure from earlier work that has examined the relationships between the MSLT and clinical and polysomnographic characteristics. The main advantage of survival analysis over regression or correlation methods is its handling of censored data. Because the MSLT measures time to an event (falling asleep), patients who have not fallen asleep at the end of the nap trial (20 min) yield censored observations. Regression methods must truncate, eliminate, or randomize censored observations and can therefore bias or introduce error into the analysis. Furthermore, these methods do not allow for modeling the time-dependent nature of the outcome or predictor variables.

The results of our study further highlight the clinical significance of the MSLT in the evaluation of patients with SDB. There is considerable debate on the overall utility of MSLT in the management of patients with disorders of excessive daytime sleepiness. Within the context of SDB, our results would suggest that there is a direct relationship between the severity of underlying apnea and the level of daytime impairment as judged by the MSLT. In the clinical scenario where there is discordance between the degree of apnea and sleep latency, other potential causes of hypersomnolence should be investigated. For example, in a patient with mild apnea but severely decreased sleep latency, the degree of hypersomnolence may be attributable to other causes such as chronic insufficient sleep, poor sleep hygiene, or even narcolepsy. Although there does exist a relationship between apnea severity and sleep propensity, a small number of patients with severe apnea will not have a decrease in their sleep latency. In this situation, treatment options for apnea should be guided by other clinical implications, such as increased cardiovascular morbidity and mortality.

In conclusion, our results show for the first time that apnea-hypopnea frequency, degree of nocturnal hypoxemia, and sleep fragmentation independently contribute to an increased risk of daytime sleepiness in patients with SDB. In addition, we also show that morbid obesity is an independent risk for hypersomnolence after adjusting for potentially confounding variables. Finally, survival analysis is an alternative, and perhaps more efficient, approach for the analysis of the MSLT.