INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Excessive daytime sleepiness is the predominant symptom for which nasal continuous positive airway pressure (nCPAP) is prescribed in obstructive sleep apnea (OSA). This sleepiness appears to result from recurrent arousal from sleep triggered by repetitive episodes of partial or complete obstruction to the pharyngeal airway (1, 2), rather than intermittent hypoxia (3). The quantification of sleep fragmentation is therefore central to the assessment of these patients and a range of techniques for achieving this has been developed. Unfortunately, many of the methods used seem to correlate poorly with daytime sleepiness (4, 5) probably owing to limitations in both measurement of daytime sleepiness and in our understanding, and hence quantification, of recurrent arousals. To try to improve sleep fragmentation assessment, it has been suggested that computer-based signal analysis may be helpful (6) because it is entirely objective and can detect electroencephalograph (EEG) changes too subtle for visual scoring (7). Non-EEG indices of arousal, including autonomic activation (8) and body movement (9), have also been suggested as possible alternatives to manual EEG scoring. There are, however, no data comparing these techniques with traditional scoring.

One problem in comparing these methods for quantifying sleep fragmentation is the selection of an appropriate reference standard against which to judge them. Comparison of the new indices with manually scored EEG microarousals is less than ideal because these microarousals are themselves only poor correlates of objective daytime sleepiness (5), perhaps because daytime sleepiness is due to many factors other than recurrent arousals, such as sleep deprivation (10), drugs (11), xanthines (12), or cytokines (13). The improvement in daytime sleepiness in response to nCPAP may be the most clinically relevant reference standard for patients with OSA because this is the main endpoint of clinical significance. This study uses this as the reference standard against which to compare the various approaches to the quantification of sleep fragmentation in OSA. It investigates the relationship of both the initial daytime sleepiness and nCPAP responsive daytime sleepiness with EEG markers of sleep fragmentation (manual EEG scoring of American Sleep Disorders Association [ASDA] arousals [14], 1.5-s microarousals, and computerized markers of EEG disruption) as well as non-EEG markers of sleep fragmentation (autonomic arousals [15] and movement indices [9]) in a population of subjects referred to a sleep clinic for investigation for possible OSA.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Subjects

Forty-one subjects (36 men, 5 women) were randomly recruited from the Oxford Sleep Clinic where they had been referred for investigation for snoring and possible OSA. Presenting symptoms were either snoring or daytime sleepiness or both. After the first 30 subjects were recruited, the presence of snoring and > 4% oxygen saturation dip rate was reviewed and the final 11 subjects specifically chosen (on the basis of > 4% oxygen saturation dip rate) to ensure the population represented the full spectrum of disease, from nonsnoring, through simple snoring, to mild/moderate and severe OSA (Table 1). This study was approved by the Central Oxford Research Ethics Committee, and subjects gave consent in accordance with the Committee's requirements.

Protocol

Subjects spent one day in the laboratory prior to the study polysomnogram, for assessment of daytime sleepiness. The night following the polysomnogram, all subjects, regardless of the sleep study findings, were commenced on nCPAP. At the end of a 4-wk nCPAP treatment, subjects returned for a further day for the assessment of their post-treatment daytime sleepiness.

Assessment of daytime subjective and objective sleepiness prior to nCPAP. Subjects received written instructions to have a normal night of sleep before attending the laboratory for the test and to abstain from drinks containing alcohol or caffeine from midnight before and throughout the day of the test. Compliance with these instructions was confirmed on arrival at the sleep laboratory. The Oxford Sleep Resistance test (OSLER test) was used to objectively quantify daytime sleepiness. This is a behavioral test of sleepiness similar in structure to the 40-min EEG-based maintenance of wakefulness test (MWT) (16). It uses the response to a regularly flashing light-emitting diode (LED) to assess wakefulness. This test has the advantage that it avoids instrumentation of the subject and is objective (being computer-controlled). It has been shown to separate patients with OSA from normal subjects as well as the EEG-based MWT (17). During the OSLER test, four sleep resistance challenges were performed in a darkened room at two hourly intervals throughout the day and sleep was concluded to have occurred when there had been no response to the LED for 21 s. The results from these four tests were averaged to give a single figure. Subjective sleepiness was assessed using the Epworth Sleepiness Scale (ESS) (18).

Sleep recording. Sleep was recorded on an eight-channel tape recorder (MPA-2/9200; Oxford Medical Instruments, Abingdon, UK). Recording included frontal EEG (Fp2-A2, which in preliminary work has been shown to improve respiratory related arousal detection [19]), two channels of electro-occulograph (EOG), and submental electromyography (EMG). Respiration was recorded from oronasal airflow, ribcage and abdomen movements, and arterial pulse oximetry (3700 pulse oximeter; Ohmeda, Herts, UK). Apnea was defined as a fall in airflow to less than 20% of baseline for more than 10 s and hypopnea as a fall in airflow to less than 50% for the same period. Indirect arterial beat-to-beat blood pressure was monitored using pulse transit time measured between the ECG R-wave and the photoplethysmographically detected arrival of the pulse wave at the left ring finger (RM10; Parametric Recorders, London, UK) (20). Movement was detected by subtracting serial digitized video images gathered every 2 s (Visilab Sleep Monitoring System; Stowood Instruments, Oxford, UK).

nCPAP titration and initiation. nCPAP was commenced the following night in all subjects, who were all given the same information regarding nCPAP treatment regardless of disease severity. It was explained that nCPAP would correct snoring but that it was not possible, at present, to confidently predict from the sleep study whether they would gain any improvement in daytime sleepiness or vitality on nCPAP, therefore they would be given 4 wk of treatment with measurements before and after for comparison. Subjects viewed an educational video and skilled technical staff chose the most comfortable nasal mask and head-gear, ensuring the subjects were at ease with the system. The first night of treatment was in the laboratory using the Horizon Auto-adjust nCPAP system (7354I Series; Sunrise Medical, Devilbiss Division, West Midlands, UK), which adjusts nCPAP pressure according to the presence of snoring, apneas, and hypopneas. During the nCPAP titration night subjects were monitored using a multichannel system, monitoring > 4% oxygen saturation dip rate, snoring and arousal confirmed by movement (9), and brief heart rate rises (autonomic nervous system activation) (15) (Visilab Sleep Monitoring System; Stowood Instruments). nCPAP pressure was chosen as that which controlled all snoring and sleep fragmentation related to upper airway incompetence (in all subjects this was the 90th percentile of overnight nCPAP pressure or higher). In the nonsnoring subjects, where the auto-adjust nCPAP pressure did not rise above 3 cm H₂O, the minimum nCPAP pressure used was 5 cm H₂O to avoid patient hypoxemia. The following morning, subjects were reviewed and supplied with a nCPAP machine (7353 Series; Sunrise Medical, Devilbiss Division), which included a compliance counter based on actual machine use (by detecting changes in motor speed with breathing). They were also provided with our usual support information, including a contact telephone number that, it was emphasized, they should use for any problems with their treatment.

Follow-up assessment of sleepiness and nCPAP compliance. After 4 wk on nCPAP all subjects returned to the sleep laboratory where the OSLER test and the Epworth score were repeated and nCPAP compliance noted. The change in OSLER test and Epworth score were calculated by subtracting pretreatment from post-treatment values, therefore quantifying the change in objective and subjective sleepiness with nCPAP. Treatment indices were all expressed as the total number of events divided by the Rechtschaffen and Kales' sleep period time (the total time from initial sleep onset to final arousal from sleep [21]).

Analysis

EEG indices of sleep fragmentation. Sleep was staged manually according to standard criteria (21). ASDA arousals were scored using the predefined ASDA criteria (14) and 1.5-s arousals were scored as a return to alpha rhythm for at least 1.5 s with or without a rise in EMG frequency.

Neural network EEG analysis. The EEG was digitized to 8-bit accuracy at 128 Hz and analyzed using a neural network EEG analysis system (22) (Questar; Oxford Medical, Abingdon, UK), which is based on autoregressive modeling of the EEG signal. Using this technique, the full range of the sleep/wake spectrum can be described on a second-by-second basis ranging from +100 (awake) to 100 (deep sleep). The neural network outputs are probability estimates and, unlike conventional sleep staging, the EEG is quantified on a continuous scale that is not linearly related to conventional sleep stages. A neural network output of +100 indicates wakefulness, a value of 0 indicates REM/light sleep, and a value of 100 stage 4 sleep. Based on our previous development work this neural network output was postprocessed in two different ways: (1) the sleep descent index (23), to identify the number of times the subject descended into sleep; (2) the standard deviation of sleep depth, to quantify overall variability in the EEG signal during sleep.

Sleep descent index. The neural network sleep descent index was calculated by smoothing the raw 1-s neural network output using a 3-s moving mean window, and the number of events where sleep descended by more than 50% of the sleep-wake spectrum over a period of 45 s or less was counted (Figure 1). The number of such events was then expressed per hour. As the neural network technique is not linearly related to traditional sleep staging and describes sleep depth on a continual spectrum second by second, it is difficult to equate these 50% sleep descents with traditional sleep stages. For example, a fall from +100 to 0 would be falling from wide awake to light sleep, while an equal numerical fall from 0 to 100 is a move from light sleep to stage 4. The absence of a direct scaling to traditional staging is due to the derivation of the neural network sleep spectrum, free from a priori principles (22), and is one of the reasons that this study uses a clinical reference standard rather than simply comparing to traditional indices.

View larger version (12K): [in this window] [in a new window]   Figure 1.   This graph shows the method of identifying events for descent index. Second-by-second changes of the smoothed neural network sleep depth spectrum are shown over 30 s. The points shown represent the start of descents that fulfill the criteria (a greater than 50% fall of the full sleep-wake spectrum).

Standard deviation (SD) of sleep depth. The standard deviation of the neural network sleep depth (neural network SD) was calculated over successive 60-s windows and averaged to give a single figure for the whole night. This provides a non-rule-based index of oscillations in EEG sleep depth, which is not influenced by the slower variations in sleep depth variability associated with the ultradian sleep cycle.

Both the neural network sleep descent index and the neural network SD were scored automatically by computer to avoid observer bias and ensure reproducibility.

Non-EEG-based Indices of Sleep Fragmentation

Autonomic arousal index (AAI). The pulse transit time (PTT) is measured with every heart beat, and is oversampled and stored at 5 Hz to ensure no values are missed. In keeping with our previous development work (24), a fall in PTT of greater than 15 ms occurring over a 4-s to 45-s period was defined as an autonomic arousal event (20). The start and end times for PTT analysis were taken from the Rechtschaffen and Kales' sleep period time (SPT) and the total number of PTT arousals were expressed per hour to give the AAI.

Movement event index (MEI). Movements were detected by digital subtraction of video images every 2 s (9). A moving mean window of the image changes across 10 serial images (20 s) was used as the baseline against which to detect arousal-related movement. A movement event was then defined as a change of at least two standard deviations from this baseline on two sequential 2-s windows. When this occurred another event could not be identified for at least 20 s. These events were identified automatically by computer to ensure reproducibility and the total number expressed per hour to give the MEI.

Statistical Analysis

Associations between individual markers of sleep fragmentation and measures of daytime sleepiness were explored using Spearman's correlation analysis (SAS statistical software package; SAS Institute, Cary, NC). In order to assess which sleep fragmentation indices independently contributed to the variance in pretreatment sleepiness, all the indices of sleep fragmentation and OSA severity were entered into a forward stepwise multiple regression analysis with the OSLER test or ESS score as the dependent variable. Similarly, to investigate predictors of sleepiness responsive to nCPAP, all the studied indices were entered into a forward stepwise multiple regression analysis, with the change in OSLER test or ESS on nCPAP as the dependent variable. The criteria for entry into the models was p < 0.05 using Spearman's correlation analysis, with a significance level of p < 0.15 required for retention in the models.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Subjects in the studied population included the full range of pharyngeal collapsibility during sleep, ranging from nonsnoring to severe OSA on the diagnostic study, and a wide range of both pretreatment subjective and objective sleepiness (Table 1) and measured sleep fragmentation severity (Table 2). Post nCPAP measures of sleepiness (OSLER test and Epworth score) were gathered in 40 subjects (one subject refused to attend for follow-up assessment). Both subjective and objective sleepiness improved on nCPAP. The median Epworth score fell from 14 to 7 (p < 0.0001, paired t test) and the OSLER test increased from 29.6 min to 39.5 min (p < 0.0001, paired t test) as shown in Figure 2.

View larger version (15K): [in this window] [in a new window]   Figure 2.   Box and whisker plot to show the improvement in objective sleepiness on nCPAP. Each box shows the interquartile range, and the horizontal line within the box represents the median. The whiskers extend from the 10th to 90th percentiles and the individual dots represent the outlying data points.

Spearman's correlation analysis showed that all markers of sleep fragmentation (manual EEG scoring, automated EEG analysis, and non-EEG-based indices) had significant associations with daytime sleepiness and its improvement with nCPAP (Table 3). These relationships with sleepiness and its response to nCPAP were generally of similar magnitude for all the indices. The MEI and the > 4% oxygen saturation dip rate had the closest correlations with the changes in both subjective sleepiness (Epworth score) and objective sleepiness (OSLER test) on nCPAP (Figure 3). There was also a particularly close relationship between the changes in OSLER test result and the neural network SD (r = 0.61, p < 0.0001, Figure 4). The apnea/hypopnea index (AHI) also correlated significantly with sleepiness and its response to treatment. The strength of its relationship was similar to the sleep fragmentation indices but not as good as the MEI or the > 4% oxygen saturation dip rate.

View larger version (18K): [in this window] [in a new window]   Figure 3.   Scatter plots showing (left panel ) the relationship between improvement in subjective sleepiness (Epworth) on nCPAP and the number of movements detected per hour of sleep (movement event index), (right panel ) the relationship between improvement in objective sleepiness (OSLER) on nCPAP and the number of movements detected per hour of sleep (movement event index).

View larger version (17K): [in this window] [in a new window]   Figure 4.   Scatter plot showing the relationship between change in objective sleepiness on nCPAP (OSLER) and neural network SD.

Compliance correlated weakly with AHI, MEI, and > 4% oxygen saturation dip rate and rather better with the neural network variables. It did not correlate with baseline sleepiness or change in objective sleepiness on nCPAP. In keeping with this finding, some subjects improved on nCPAP despite using it less than 4 h per night. The population was divided into two groups on the basis of a compliance less than or greater than 4 h (Table 4). Although the median values of the change in ESS and OSLER show that overall there was no improvement in sleepiness in the low nCPAP users, the range shows that there were some subjects who achieved large improvements in sleepiness (ESS fall of 12 and improvements in OSLER of 15 min) despite usage of less than 4 h per night. Using Wilcoxon rank sum analysis there was no significant difference in either baseline sleepiness or improvement in objective sleepiness between these two groups, although the difference in improvement in ESS was significant (Table 4).

As might have been expected, there were no significant correlations between any of the sleep fragmentation indices and absolute measurements of post-nCPAP objective sleepiness. The MEI correlated weakly with post-treatment subjective sleepiness (r = 0.35, p = 0.03) but no other sleep fragmentation index showed a significant correlation, including AHI.

Subgroup Analysis

Paired t tests were performed on two subgroups, subjects with an AHI < 5 and subjects with an AHI < 10 to look for possible deterioration in sleepiness on nCPAP that may have produced artifactual correlations. There was a small improvement in subjective sleepiness that was not statistically significant in those subjects with an AHI of < 5 (Epworth, 3, p = 0.1; OSLER test, +1.43, p = 0.6) and just statistically significant where the AHI was < 10 (Epworth, 2.85, p < 0.01; OSLER test, 0.04, p = 0.9).

Multiple Regression Analysis

The multiple regression analysis was performed twice, initially with all sleep fragmentation indices except for the > 4% oxygen saturation dip rate, and then repeated with the inclusion of this variable (see DISCUSSION). In the first analysis, the best predictor of pretreatment objective sleepiness, nCPAP responsive subjective sleepiness, and nCPAP responsive objective sleepiness was the MEI. The neural network SD also independently contributed to the models describing nCPAP subjective and objective sleepiness (Table 5). The models of improvement in subjective and objective sleepiness were able to explain 44% and 51% of the variance, respectively. In the second analysis, there was no change in the models predicting pretreatment subjective and objective sleepiness. The > 4% oxygen saturation dip rate was the best predictor of the change in both subjective and objective sleepiness (Table 6) with little change in the total model variance.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

This study is the first to systematically compare traditional EEG, computerized EEG signal analysis, and non-EEG markers of sleep fragmentation, with two clinically helpful reference standardsobjective and subjective daytime sleepiness and their response to treatment with nCPAP. The results are of interest because they show that EEG indices of sleep fragmentation are more closely associated with subjective and objective daytime sleepiness than has been previously reported (4, 5, 25); in addition, despite these closer correlations, the non-EEG markers are at least as good as the traditional EEG indices at predicting these endpoints. In the multiple regression modeling, most of the explainable variance in sleepiness was associated with the non-EEG index, movement, and the variability in EEG sleep depth (the neural network SD). These two indices alone explained 44% and 51% of the improvement in subjective and objective sleepiness.

Toward the end of the study, the > 4% oxygen saturation dip rate was used to guide subject selection, to ensure a full representation of the severity of OSA in the study group (particularly for identifying patients with moderate/severe OSA). It is possible that this active use of the > 4% oxygen saturation dip rate in subject selection may have resulted in this signal including less random noise, and hence being a better correlate of the other endpoints (Table 3). For this reason, to be sure to exclude such bias in our analysis, we omitted the > 4% oxygen saturation dip rate from our main multiple linear regression analyses. When > 4% oxygen saturation dip rate was included in the multiple regression models it was the best predictor of the change in both subjective and objective sleepiness with little change in the total model variance. It did not change the results for baseline subjective and objective sleepiness.

The ability to identify from the sleep study those subjects whose daytime sleepiness will substantially improve on nCPAP is clinically useful. No previous studies have specifically examined this endpoint. In order to be able to identify predictors of improvement in sleepiness, we needed to generate a large spectrum of response to nCPAP. To achieve this it was necessary to treat subjects who, on the basis of the sleep study, would not normally have been considered for nCPAP treatment, such as nonsnorers. To ensure that this approach had not produced artifactual correlations by making these nonsnoring subjects worse (by fragmenting their sleep with nCPAP) we performed a subgroup analysis on subjects at the mild end of the spectrum (AHI < 5 and AHI < 10) and found there was no deterioration in either subjective or objective measures of daytime sleepiness. There were subtle improvements in the ESS in patients with mild disease that are probably due to the correction of snoring-related sleep fragmentation, but may be a placebo effect.

Our study shows closer associations with EEG indices of sleep fragmentation and daytime sleepiness than have previously been reported (4, 5, 25), which may be because of a number of factors. Our study population incorporates subjects evenly distributed across the disease severity spectrum, from nonsnorers to severe OSA, and consequently includes the full variance in sleepiness. Others have not studied such a wide range of severity of disease (2, 5), which may have weakened their observed correlations. We have also used frontal EEG to score arousals, which provisional work suggests is more sensitive at detecting arousals associated with respiratory events (19). Another reason for our closer correlations may be the methods we employed in objectively quantifying daytime sleepiness. From the perspective of the patient with OSA, sleepiness manifests itself as the inability to maintain vigilance and thus complete monotonous tasks. The OSLER test was designed specifically to quantify this problem since the endpoint of this test is the failure of function, which may not necessarily be the same as the endpoint of the conventional MWT (usually one epoch of stage 1 sleep). The OSLER test is similar in structure to the MWT, which others have shown to be a more appropriate test for investigating sleep-disordered breathing than the multiple sleep latency test (MSLT) (4), and more sensitive to the effects of nCPAP (26).

Of all the indices we examined in our sample, body movement proved to be the best predictor of sleepiness and of improvement in sleepiness on nCPAP when the possibly biased > 4% oxygen saturation dip rate was omitted. Relatively few studies have previously examined movement as a sleep fragmentation index in OSA. Collard and coworkers (27) investigated movement arousals derived from EEG and EMG tracings in sleep-disordered breathing where they correlated with the AHI (r = 0.76) and showed that the movement arousal index decreased to the normal range with nCPAP treatment. They did not however look at the relationship between movement arousals and daytime sleepiness. Cheshire and coworkers (5) identified EEG arousals associated with an increase in EMG tone that correlated well with AHI (r = 0.88) but did not correlate with MSLT. Others have looked at non-EEG-based methods to detect movement (28) and have compared these with ASDA arousals but have not explored the relationship with daytime sleepiness. Our results suggest that movement, when quantified from whole body movement and not just EMG activation, can be a helpful marker of sleep fragmentation caused by sleep-disordered breathing.

Autonomic arousal ("subcortical" arousal) has also attracted attention as a potential index to quantify sleep disturbance in OSA (20). This technique identifies even very minor disturbing events, as a rise in blood pressure can occur without visible EEG arousal (7). Repeated, induced, autonomic arousals in the absence of visible EEG arousal cause increased objective daytime sleepiness as measured by the MWT in normal subjects (29), but prior to this study it was not known whether these "subcortical" events were independently important as predictors of daytime sleepiness in OSA. In this study we have found significant correlations between the autonomic arousal index and pretreatment objective sleepiness (r = 0.49) and between the autonomic arousal index and nCPAP responsive objective sleepiness (r = 0.44). These correlations were very similar to those between the ASDA arousal index and sleepiness. This suggests that detection of autonomic activation is as good as the ASDA arousal index for predicting daytime sleepiness in OSA, but it does not convey any extra information beyond that available from other indices.

Traditional methods of arousal detection have focused on the most direct index of electrical brain activity, the cortical EEG. Traditional Rechtschaffen and Kales' epoch-based sleep scoring (21) ignores brief arousals from sleep that are physiologically important in OSA and this led to the development of microarousal scoring, which includes rules for the identification of brief, 3-s arousals (14) and some even more sensitive, "in-house" definitions that require as little as 1.5 s of EEG activation (5, 30). Unfortunately, although these definitions are more useful than epoch-based sleep staging, their relationship with measured daytime sleepiness remains disappointing (5). Further improvements in the sensitivity of these manually detected EEG arousals seem difficult because this requires progressively more subtle EEG changes to be scored reliably, and as progressively shorter definitions are used, apparently "normal" sleep becomes very fragmented (31). Automated EEG analysis has been advanced as a possible way of improving this situation (6, 29), as it is entirely objective, capable of analysis with second-to-second resolution, and complex algorithms are possible that take into account several aspects of the EEG waveform. It is less time-consuming than manual EEG analysis and avoids the considerable interscorer variation in interpretation of ASDA arousals (32). However, although other studies have explored the use of automated analysis of EEG to identify arousals (29, 33), previous work has not assessed the relative usefulness of automated EEG analysis for identifying nCPAP responsive sleepiness. The neural network EEG analysis system used in this study uses autoregressive modeling to analyze EEG frequency spectrum changes. This is an acceptable method of analysis when studying transient changes as it is better suited to analysis over short time periods compared with the fast Fourier transform (FFT) whose ability to resolve EEG information is limited by the minimum size of the sample window. In previous development work (34) we have shown that postprocessing the neural network output (second-by-second sleep depth) enabled detection of 97% of ASDA events in subjects with severe sleep apnea. The neural network EEG analysis system has also provided a tool with which to quantify sleep depth variability that was calculated using the average of the neural network standard deviation across the night (using 60-s sequential windows). This measure does not require a denominator for calculation (so there is no requirement for traditional Rechtschaffen and Kales' analysis) and no arbitrary thresholds have been used in its definition. The correlations with sleepiness and these neural network indices are comparable with the other indices studied, suggesting these automated indices are as clinically useful as the traditional EEG sleep fragmentation indices in OSA. Interestingly, in the models excluding the > 4% oxygen saturation dip rate, this was the only index to explain any additional variance in nCPAP responsive sleepiness after considering body movement. This suggests that there is some variability in the EEG that is important in sleepiness which is not quantified by any index based on "event counting." Further developments in EEG signal analysis may clarify this further.

In order to compare the relationship between sleepiness measures and sleep fragmentation indices, the same time denominator was used to calculate all the sleep fragmentation indices. Traditionally, EEG microarousals are expressed as the number of arousals per total sleep time (TST). In contrast, non-EEG-based indices are usually described as the number per hour of sleep period (as they are usually quantified without a parallel EEG recording). In order to describe the non-EEG indices in a comparative manner, this study has used SPT as the denominator for calculating all the sleep fragmentation indices. To confirm that the use of this denominator did not weaken the relationships with variables usually expressed per hour of TST, we repeated the correlation analysis for ASDA arousals, 1.5-s microarousals, and AHI, using the traditional methods for calculating these indices. This did not change the Spearman's correlation analysis results to any statistically or clinically significant extent (ASDA versus pretreatment ESS and OSLER, change in ESS, and change in OSLER, 0.49, 0.49, 0.41, and 0.42, respectively; AHI versus pretreatment ESS and OSLER, change in ESS, and change in OSLER, 0.43, 0.42, 0.55, and 0.53, respectively; 1.5-s microarousals versus pretreatment ESS and OSLER, change in ESS, and change in OSLER, 0.49, 0.47, 0.40, and 0.41, respectively) and did not alter the multiple regression analyses.

Compliance correlated weakly with the MEI (r = 0.39) and slightly closer with the neural network variables (neural network standard deviation, r = 0.47; neural network descent index, r = 0.48). Interestingly, it did not correlate with baseline sleepiness or change in objective sleepiness on nCPAP. In addition there were some subjects who used nCPAP for less than 4 h who had large improvements in sleepiness (ESS fall of 12 and improvement in OSLER of 15 min) and conversely subjects who, despite nCPAP usage of more than 4 h per night, had deterioration in sleepiness measurements. This implies that those subjects who benefit from nCPAP are not always those who use it the most. One reason for this is probably that there is considerable intersubject variation in the amount of nCPAP usage per night required for symptom improvement in the same way that there is intersubject variation in the amount of normal sleep required per night to prevent sleepiness (35).

As might have been expected, there were no significant correlations between absolute post-treatment levels of sleepiness and sleep fragmentation indices at baseline. The aim of this study was to identify patients who improved on nCPAP, not to predict the actual levels of post-treatment sleepiness. In this study population, subjects with low levels of post-treatment sleepiness included both nonsnoring, nonsleepy subjects and subjects with severe sleep apnea who had improved with treatment, therefore the lack of correlation is not surprising. This lack of relationship after nCPAP, despite positive correlations before treatment, indirectly supports the idea that measuring actual change in sleepiness on nCPAP filters out sleepiness due to other causes.

In conclusion, we have studied several new methods of quantifying sleep fragmentation and explored the ability of these indices alongside traditional EEG indices to predict sleepiness as well as identify those subjects whose symptoms of sleepiness will improve on nCPAP. In a study population assembled to be well suited to this analysis, significant and similar relationships between all the sleep fragmentation indices and sleepiness were found. The non-EEG-based sleep fragmentation indices and the neural network-based EEG analysis compared favorably with traditional EEG microarousals. Excluding the > 4% oxygen saturation dip rate (which may be falsely favored in our analysis), the best predictor of both pretreatment objective sleepiness and improvement in sleepiness with nCPAP, proved to be the number of body movements per hour, and only a non-rule-based index of EEG variability (neural network SD) added to this index in a multiple linear regression model. These two indices explained 51% of the variance in the improvement in objective sleepiness with nCPAP. This perhaps suggests that the arousals contributing most to daytime sleepiness in OSA are those causing the greatest oscillations in sleep depth or are large enough to cause body movement. We conclude that the automated indices of sleep fragmentation are at least as useful as traditional indices for quantifying OSA-related sleep fragmentation. In addition, the automated indices provide objective and cost-effective methods of identifying those OSA patients whose sleepiness is likely to respond to nCPAP.