Disordered Breathing - Semantic Scholar

Method for Detection of Respiratory Cycle-Related EEG Changes in SleepDisordered Breathing Ronald D. Chervin, MD, MS1; Joseph W. Burns, PhD2; Nikolas S. Subotic, PhD2; Christopher Roussi, MS2; Brian Thelen, PhD2; Deborah L. Ruzicka, RN, PhD1 1Sleep

Disorders Center and Department of Neurology, University of Michigan, Ann Arbor, Mich; 2Emerging Technologies Group, Altarum Institute, Ann Arbor, Mich

Study Objectives: In sleep-disordered breathing (SDB), visual or computerized analysis of electroencephalogram (EEG) signals shows that disruption of sleep architecture occurs in association with apneas and hypopneas. We developed a new signal analysis algorithm to investigate whether brief changes in cortical activity can also occur with individual respiratory cycles. Design: Retrospective. Setting: University sleep laboratory. Participants: A 6 year-old boy with SDB. Intervention: Polysomnography before and after clinically indicated adenotonsillectomy. Measurements: For the first 3 hours of nocturnal sleep, a computer algorithm divided nonapneic respiratory cycles into 4 segments and, for each, computed mean EEG powers within delta, theta, alpha, sigma, and beta frequency ranges. Differences between segment-specific EEG powers were tested by analysis of variance. Respiratory cycle-related EEG changes (RCREC) were quantified.

Results: Preoperative RCREC were statistically significant in delta (P < .0001), theta (P < .001), and sigma (P < .0001) but not alpha or beta (P > .01) ranges. One year after the operation, RCREC in all ranges showed statistical significance (P < .01), but delta, theta, and sigma RCREC had decreased, whereas alpha and beta RCREC had increased. Preoperative RCREC also were demonstrated in a sequence of 101 breaths that contained no apneas or hypopneas (P < .0001). Several tested variations in the signal-analysis approach, including analysis of the entire nocturnal polysomnogram, did not meaningfully improve the significance of RCREC. Conclusions: In this child with SDB, the EEG varied with respiratory cycles to a quantifiable extent that changed after adenotonsillectomy. We speculate that RCREC may reflect brief but extremely numerous microarousals. Citation: Chervin RD; Burns JW; Subotic NS et al. Method for detection of respiratory cycle-related EEG changes in sleep-disordered breathing. SLEEP 2004;27(1):110-5.

INTRODUCTION

manuscript describes tests among 10 children to determine whether the new algorithm may have clinical utility beyond that offered by standard polysomnographic measures.4

SLEEP-DISORDERED BREATHING (SDB) CAUSES SLEEPINESS IN ADULTS AND PROBABLY CONTRIBUTES TO HYPERACTIVE BEHAVIOR AND COGNITIVE DEFICITS IN CHILDREN.1 However, the mechanism that links the abnormal respiration to these outcomes is not completely understood. Experimental comparisons of hypercapnia, hypoxemia, and increased breathing effort suggest that the latter is likely to cause arousals at the termination of apneas.2 Arousals of sufficient duration and frequency, in turn, probably contribute to daytime neurobehavioral morbidity. The minimum duration that an arousal must have to contribute to daytime consequences is not known. Arousals too subtle or brief to cause visually scored, electroencephalogram (EEG)defined arousals can be identified by other means, such as autonomic measures, and can still cause daytime sleepiness.3 In this context, we hypothesized that transient “microarousals” invisible to the human eye could occur on a breath-to-breath basis during the sleep of patients with SDB. To test this hypothesis for the first time, we developed a new computerized algorithm and applied it to data recorded from a child with SDB. This manuscript describes the analytic approach, along with variations that were tested. An accompanying

METHODS Subject and Procedures A 6 year-old boy scheduled for adenotonsillectomy was recruited to participate in an ongoing Institutional Review Board-approved study of sleep and behavior in children. The indication for surgery, according to the child’s otolaryngologist, was obstructed breathing, both while awake and while asleep. The child had loud nightly snoring, observed apneas, and daytime mouth breathing. His tonsils were enlarged (“kissing”), his height was 48 inches (75th percentile), and his weight was 86 pounds (> 95th percentile). At the time of enrollment, the child had not received a preoperative polysomnogram. The past medical history was significant for a thyroglossal cyst, removed at age 3. The child attended kindergarten, and his parents reported some inattentive behavior but no excessive daytime sleepiness. Laboratory-based nocturnal polysomnography was performed 4 days before the operation and again 1 year later. The child took no medications at either time, except for an antibiotic at the first evaluation. Digital polysomnography included 4 EEG channels (C3-A2, C4-A1, O1-A2, O2-A1 of the 10-20 international electrode placement system), 2 electrooculogram channels (right and left outer canthi), chin and bilateral anterior tibialis electromyography, 2 electrocardiogram leads, nasal and oral airflow (thermocouples), thoracic and abdominal excursion (piezoelectric strain gauges), finger oximetry (SaO2, with viewable but not recorded pulse waveform), end-tidal CO2, and transcutaneous CO2. Sleep-stage scoring followed standard protocols.5 Apneas were scored when airflow was absent for at least 2 breath cycles. Hypopneas were scored when at least 2 breath cycles of diminished airflow, chest movement, or abdominal movement were followed by an arousal, an awakening, or a 4% oxygen desaturation.

Disclosure Statement The University of Michigan and Altarum Institute have filed a provisional patent entitled, "System and method for analysis of respiratory cycle-related EEG changes in sleep-disordered breathing," Drs. Chervin, Burns, Subotic, and Roussi. Drs. Thelen, and Ruzicka, nothing to disclose. Submitted for publication April 2003 Accepted for publication September 2003 Address correspondence to: Ronald D. Chervin, MD, MS, Michael S. Aldrich Sleep Disorders Laboratory, 8D8702 University Hospital, 1500 E. Medical Center Dr, Ann Arbor, MI 48109-0117; Tel: 734-647-9064; FAX: 734-647-9065; E-mail: [email protected] SLEEP, Vol. 27, No. 1, 2004

110

Respiratory Cycle-Related EEG Changes in SDB—Chervin et al

Data Processing

Algorithm

The first 360 thirty-second epochs (3 hours) of scored sleep in each polysomnogram were analyzed. The early portion of the night was used because it is most likely to demonstrate high amounts of slow-wave sleep6,7 and is most likely to have an impact on daytime sleepiness.8 Respiratory cycles that were below the fifth percentile or above the 95th percentile for amplitude or length were rejected to avoid inclusion of apneas or spurious deviations in a typically sinusoidal airflow pattern. A computer program divided each respiratory cycle into 4 segments based on oral-nasal airflow (thermocouple) maxima and minima and the midpoints between them (Figure 1). For occasional clipped regions of respiratory cycles, midpoints were used to approximate locations of maxima and minima. Then the corresponding EEG power in the C3-A2 derivation—for delta (1-4 Hz), theta (5-7 Hz), alpha (8-12 Hz), sigma (13-15 Hz), and beta (16-30 Hz) frequencies—was calculated for each respiratory cycle segment in an algorithm described below. The frequency-specific mean power of each segment was normalized (divided by the mean power in that frequency for the entire breath cycle). Resulting values were averaged across approximately 2500 breaths to obtain mean relative energies for each segment. The mean energies for the 4 respiratory cycle segments were then compared and tested for significant differences using analysis of variance (ANOVA). The significance level was set at P < .01 for the ANOVA tests because control simulations (see results) indicated that the data may not fit a normal distribution perfectly; a more conservative critical value was necessary to ensure an overall 5% level of significance. The magnitude of respiratory cycle-related EEG changes (RCREC) for a given polysomnogram was calculated as the difference between the highest mean segment-specific EEG power and the lowest. The RCREC before and after the adenotonsillectomy were then compared.

In short, the frequency-specific EEG power in any given respiratory cycle segment (Figure 1) was standardized by expression as a proportion (mean power for the segment divided by mean power for the entire cycle) minus 1.00. Thus, if no difference in EEG power occurred across the respiratory cycle, all segments had a mean relative power of 0.00. More formally, to determine RCREC, a time-evolved representation of the EEG power spectrum was generated for each respiratory cycle segment. Since the EEG signal is inherently nonstationary, a time-frequency decomposition was used to estimate spectral power as a function of time. The EEG signal had been sampled at 100 Hz. The time-frequency decomposition of the EEG was made using the short-time Fourier transform, which generates a local estimate of the spectral power as a function of time.9 The Fourier transform of a 1-second window of data was repeatedly calculated as the window slid over the signal. A window of this size allowed an unweighted resolution of 1 Hz. Data were detrended and weighted with a Hanning window before the Fourier transform was applied.9 At each step, the window was moved 2 samples (2/100th of a second), essentially starting at the left-sided border of a respiratorycycle segment and proceeding until the start of the sliding window reached the right-sided border. Because many windows captured some activity in the subsequent segment, EEG powers assigned by this algorithm to a specific segment were in reality delayed by one half of a window, or 0.5 seconds. The EEG power for all the windows assigned to each segment were averaged, and the means were then normalized with respect to the mean power during the entire cycle. Mathematically, for a specific frequency or frequency band, computation of the ci ratio of EEG power during the breath segment to total power during the breath cycle was accomplished by the formula:

where S is short-time Fourier transform of EEG data, F is frequency band, T is respiratory cycle-defined time region, and Ti is the segmentdefined time region. RESULTS Respiratory-Cycle-Related EEG Changes Polysomnography demonstrated significant obstructive sleep apnea before but not after adenotonsillectomy, as shown in Table 1. Before the operation, the spectral power within the delta frequency band differed significantly between respiratory cycle segments (P < .0001; Figure 2A, at left). The lowest power appeared to occur during late expiration. However, taking into account the half-second shift that results from the sliding-window algorithm, the lowest delta power was probably located primarily during inspiration. Similar patterns were displayed by the other frequency bands (Figures 2B-2E, left-side panels), though the differences between segments did not reach significance for the alpha and beta frequencies. One year after the operation, the differences in EEG powers of all frequency bands during specific respiratory-cycle segments achieved statistical significance, though some had diminished in magnitude (Figures 2A-2E, right-sided panels). The magnitude of RCREC, as displayed in Figure 3 for each frequency band within each of the 2 sleep studies, was calculated by subtraction of the lowest values in the Figure 2 bar graphs from the highest. In Figure 4, preoperative RCREC averaged across 10-minute intervals are shown through the night in relation to their statistical significance, sleep stages, and occurrence of apneas or hypopneas. After the operation, the delta, theta, and sigma RCREC diminished substantially, whereas alpha and beta RCREC increased (Figure 3).

Figure 1—As shown in this schematic diagram of a thermocouple-derived oral/nasal airflow signal, respiratory cycle segments were defined by maxima, minima, and midpoints.

Table 1—Polysomnographic results before and 1 year after adenotonsillectomy. Total recording time, min Total sleep time, min Sleep efficiency (total sleep time * 100 / total recording time), % Stage 1 sleep, % Stage 2 sleep, % Stage 3 and 4 sleep, % Stage REM sleep, % Obstructive apnea index, events per hour of sleep Apnea-hypopnea index, events per hour of sleep Minimum oxygen saturation, % Percentage of sleep time spent with end-tidal CO2 > 50 mm Hg

Preoperative 567.5 508.5

Postoperative 617.5 541.0

89.6 17.5 43.5 26.8 12.2

87.6 8.0 48.0 25.6 18.4

8.8

0.1

13.9 90

0.2 93

0.6

0.0

REM refers to rapid eye movement.

SLEEP, Vol. 27, No. 1, 2004

111


PREOPERATIVE

To confirm that detection of RCREC was not dependent on long periods of data that include more than 2000 respiratory cycles, and also not a byproduct of hypopneas, a shorter segment of 101 consecutive respiratory cycles from the preoperative study was analyzed. Nine cycles occurred during stage 1 sleep, and 92 during stage 2 sleep. Visual inspection of this portion of the recording revealed no apneas or hypopneas, but delta RCREC was highly significant (RCREC = 0.40, F = 8.16, P = .00003).

POSTOPERATIVE

Control Analyses and Simulations

A

Two types of control analyses were performed to check whether RCREC might be an artifact. First, the EEG signal from the subject was compared to the airflow signal of another child who had been studied in the same protocol; adjustment was made for any potential correlation between the airflow of the 2 subjects. This analysis showed no significant RCREC. Second, a random-noise signal (zero-mean, unit variance Gaussian) was used, instead of the C3A2 EEG signal, for comparison to the subject’s preoperative respiratory cycle segments. A randomly varied start point within the noise signal was used to produce 1000 iterations of this analysis. For delta RCREC, 107 iterations showed ANOVA results with P < .05; for theta RCREC, 104 iterations; for alpha RCREC, 80 iterations. The number of positive samples that would have been expected in each case is 50. This control analysis suggests that the distribution of results generated by the algorithm does not follow a perfect F-distribution. To account for the resulting level of uncertainty, the level of significance used in ANOVA analyses was set at P < .01 rather than P < .05. This minor adjustment probably overcompensated for the problem: by this criterion, the number of positive associations identified in the 1000 randomnoise iterations was 27, 40, and 23 (ie, < 50 for delta, theta, and alpha RCREC, respectively).

B

C

Variations in Analytic Approach Several variations in the analytic approach were tested (Table 2). Selection of the “baseline” approach described above was based on the magnitude of RCREC for the different frequency bands, ANOVA tests for significance of RCREC, and performance of the RCREC values as correlates of outcomes.4 Finally, early analyses of EEG frequency bands as narrow as 1 Hz showed that adjacent frequencies tended to behave similarly. This suggested that analysis of wider, more physiologic frequency bands (delta, theta, alpha, sigma, and beta) did not obscure valuable information.

D

E Figure 2—A: Delta-frequency electroencephalogram (EEG) power (1-4 Hz) often was lower during late expiration and early inspiration than during late inspiration and early expiration (analysis of variance F = 13.7, P < .0001; panel A, left), but differences were less prominent after adenotonsillectomy (F = 9.2, P < .0001; panel A, right). B: Theta-frequency EEG power (5-7 Hz) varied with respiratory cycle phase before (P < .0001) and after (P < .001) adenotonsillectomy. C: Alpha-frequency EEG power (8-12 Hz) did not vary significantly with respiratory-cycle phase before adenotonsillectomy (P > .01) but did after the operation (P < .01). D: Sigma-frequency EEG power (13-15 Hz) varied with respiratory cycle phase before (P < .0001) and after (P < .01) adenotonsillectomy. E: Beta-frequency EEG power (16-30 Hz) did not vary significantly with respiratory cycle phase before adenotonsillectomy (P > .01) but did after the operation (P < .01).

SLEEP, Vol. 27, No. 1, 2004

112

DISCUSSION This study in a child with SDB shows for the first time that EEG power in specific frequency bands can vary in conjunction with the respiratory cycle. This phenomenon, called RCREC,


is quantifiable. The magnitudes of RCREC were highest in the delta range and showed substantial decrease within delta, theta, and sigma frequency ranges after SDB was treated. The RCREC could be identified in a short (101-breath) segment of non-apneic sleep. Two types of control analyses suggested that RCREC were not an artifact. The significance of RCREC remains unknown, but the finding has potentially important implications for SDB physiology and clinical practice. One implication is that visualized or computer-detected arousals after apneic events may not be the only mechanism by which SDB can disrupt cortical function. No prior research has defined a minimum duration that allows arousals, of sufficient frequency, to attain clinical significance.

Previous studies that used autonomic or spectral EEG markers of arousal in SDB focused on a time frame equivalent to several breaths or more.1014 One study of patients referred for suspected SDB analyzed nonspecific EEG power variation, between successive 60-second windows.15 Another study of severe sleep apneics and controls focused on slowwave activity in the first nocturnal sleep cycle.8 However, no previous research has sought to determine whether breath-to-breath changes in EEG power—too quick or subtle to be detected by the human eye— could occur in patients with SDB. In addition to the focus on respiratory cycles rather than apneas, the algorithm described in this report relies on at least 2 other advances without which RCREC might have remained undetected. First, results were averaged over many successive breaths. This approach bears some similarity to that taken in neurophysiology laboratories, where technicians apply a sensory, auditory, or visual stimulus hundreds of times to allow averaging of the subsequent cortical activity, cancellation of “noise,” and emergence of the evoked potentials of interest. Second, results for any single respiratory cycle segment were expressed relative to the frequency-specific EEG power during the entire breath cycle. This feature most likely eliminated considerable noise, as changes in sleep stages on a longer time frame could have overwhelmed and obscured small changes associated with short portions of the respiratory cycle. Although the current data demonstrate RCREC, they do not define what RCREC represent. One possibility is that they represent “microarousal” that occurs, on average, with each partially obstructed breath cycle. Polysomnography with esophageal pressure monitoring in patients with SDB often shows that non-apneic and non-hypopneic breaths are nonetheless associated with increased upper airway resistance, and probably with increased work of breathing.16 Long stretches of respiratory cycles with excessive negative intrathoracic pressures can be recorded from some women,17 children with obstructive hypoventilaFigure 3—The magnitudes of respiratory cycle-related electroencephalogram (EEG) tion,18 and also some men in our experience. Increased upper airway changes (RCREC) for delta, theta, and sigma frequencies were higher before adenotonsillectomy (black bars) than after the operation (white bars), but alpha and beta RCREC resistance, as opposed to hypoxia or hypercarbia, is most likely the cause showed some change in the opposite direction. * P < .01, ** P < .001, *** P < .0001 for of visible arousals in sleep apnea.2 Therefore, it was not unreasonable to the within-subject comparisons of respiratory cycle segment-specific EEG powers by analhypothesize, in the current investigation, that increased respiratory effort ysis of variance (under the assumption that these powers follow an F distribution; see caveat described in text). associated with each breath might cause subtle but repetitive and numerous microarousals. However, other possible explanations for RCREC also exist. The phenomenon could represent a cortical evoked potential triggered by airway obstruction. Experimental airway obstruction during human sleep generates an N550 response, after a delay of 500 to 800 microseconds, that resembles a K-complex and could increase delta power after each partly obstructed breath in SDB.19 Of note, noise could also generate arousals, K-complexes, or other evoked potentials, and therefore snoring could be another cause of RCREC. Some of the current observations are not easily explained. If effort to breathe during inspiration against a partially closed airway was responsible for the EEG changes, why did the decrement in delta power occur mostly but not entirely during inspiration? Thermocouples take time to heat or cool, but the resulting signal delay is only about 100 microseconds (ProTech, personal communication), insufficient to explain the observed alignment of EEG power changes with respiratory cycle phases. Another question is why, within respiratory cycles, the alpha RCREC tended to vary with rather than inversely to the delta RCREC. This Figure 4—The delta (1-4 Hz) respiratory cycle-related EEG changes (RCREC) for 10-minute intervals are shown across the result does not make sense if delta activity is a preoperative recording in the first panel. Each bin contained an average of 170 respiratory cycles. Despite their short lengths, some bins showed statistically significant analysis of variance (ANOVA) F values, as shown in the second panel, where the marker for deeper non-rapid eye movement dotted line indicates P < .01. The third panel shows sleep stages, and the fourth indicates the times at which apneas and hypopsleep and alpha activity is a marker for arousal. neas were scored. REM refers to rapid eye movement.

SLEEP, Vol. 27, No. 1, 2004

113


However, recent data suggest that alpha activity recorded over the frontal cortex, which could contribute to signals at central leads, does correlate with frontal delta activity.20 Furthermore, our observation with RCREC would not be the first in which alpha and delta activity are noted to occur together in pathologic conditions: “alpha-delta” sleep is a welldescribed feature of fibromyalgia and other conditions that involve chronic pain.21 Another hypothetical explanation for parallel changes in delta and alpha EEG power, within respiratory cycles, is that RCREC arises not because of arousal but because the brain (or surrounding cerebrospinal fluid) moves slightly in relation to the skull with each inspiration and expiration. Such movement might be enhanced by partial obstruction of the airway, which would increase fluctuation of intrathoracic pressures that could, in turn, affect cranial contents. However, RCREC would then be expected to be similar during the entire inspiration, or the entire expiration, which our data did not suggest. In addition, RCREC would be expected to change after adenotonsillectomy in similar directions for all frequency bands, but they did not.

Another observation that is difficult to explain is that alpha and beta RCREC increased after adenotonsillectomy. This could reflect a normal developmental phenomenon during the 1 year between the initial and follow-up studies. Alternatively, normal persons in comparison to those with SDB could have more alpha and beta RCREC, which could reflect lower underlying sleep debt. However, night-to-night reliability of RCREC measures remains untested. The changes in alpha and beta RCREC after 1 year, and perhaps also the changes in delta and theta RCREC, could represent underlying biologic variability. The current findings also have other limitations. The concept of RCREC and the algorithm used to identify it were developed and tested with data from the current subject and several others.4 Confirmation of our conclusions will require demonstration of similar results in a new sample. Investigation of RCREC as a predictor of SDB outcomes will be important to establish clinical utility. However, whether RCREC represent a consequence, cause, or concomitant feature of obstructed breathing during sleep, research into the phenomenon could provide insight into SDB pathophysiology, etiology, or comorTable 2—Variations in approach used to demonstrate respiratory cycle-related EEG changes bidity. The algorithm used to detect RCREC is likely to require further refinement. Data from (RCREC)* more patients, normal individuals, other EEG Parameter Variations Tested Delta Theta Alpha Sigma Beta derivations, and other analysis techniques may (baseline determination) provide more-efficient algorithms to assess breathing and cortical function during most of Baseline None .10 .08 .04 .08 .01 the night rather than the minority spent in a 13.7 7.6 2.1 9.9 .3