Electroencephalographic spectral analysis: detection of cortical activity changes in sleep apnoea patients

Subjects

The polysomnograms of 15 subjects suspected of having OSAHS (14 males, aged 51±9 yrs (mean±sd), body mass index (BMI) 29±2 kg·m⁻²) and seven healthy, asymptomatic subjects (four males, aged 50±10 yrs, BMI 28±2 kg·m⁻²) in whom OSAHS had previously been excluded, were analysed. The patients' main symptom was daytime sleepiness. Autonomic, endocrine, cardiopulmonary dysfunction, periodic limb movement disorder, other neurological disorders or intake of medication which affected sleep were excluded. The study was approved by the institutional ethics committee.

Polysomnography

Polysomnography was performed using a computerised recording system consisting of: 1) two central (C3A2, C4A1) and two occipital (O1A2, O2A1) EEG tracings, two outer canthi electrodes (LEOG, REOG) and a submental electromyography (EMG) electrode; 2) tibial EMG; 3) three-lead electrocardiogram; 4) oronasal airflow detection through a thermistor sensor, thoracoabdominal movement detection through two piezoelectric belts, a digital microphone and a pulse oximeter.

Recordings were carried out between 22:00–7:00 h. They did not include a continuous report on the supine body position.

The used polysomnography system had a digital hardware Butterworth filter of third order with cut-off frequencies set at 0.6 and 50 Hz with 12 dB per octave.

Scoring/definitions

Scoring was performed by a single observer, blinded to subject. Approximately 7% of the data were excluded due to movement or EEG artefacts. Sleep was scored on the C4A1 and C3A2 tracings according to the Rechtschaffen and Kales 10 criteria.

Arousal scoring was based on the definition of the American Sleep Disorders Association (ASDA), modified as to the duration threshold which was set at 1 s rather than 3 s 11. Arousal scoring ranged from 1–15 s, irrespective of sleep re-occurrence or prolonged awakening subsequent to the arousal; hence, Rechtschaffen and Kales 10 awakenings were also included in the scoring. The number of arousals adjacent to apnoeas/hypopnoeas per hour of sleep, was termed the respiratory arousal index (RAI).

Apnoeas were defined as a cessation of the oronasal airflow. Hypopnoeas were defined as a clear amplitude reduction of the airflow of >50% compared to a 10 s peak amplitude during the preceding 2 min, lasting ≥10 s and associated with either an oxygen desaturation of ≥3% or an arousal 12.

Arousals and respiratory events were scored during two separate sessions. To assess reproducibility of scoring outcomes, the 15 polysomnograms were scored twice by the same researcher. In these intrarater analyses the researcher was blind to the patient details and the results of the comparison studies. Intrarater comparisons were made 12 months after the initial study was scored. Variability was calculated as the difference between the two scores in per cent.

Post acquisition analysis

The power spectrum of the C4A1 tracing 13, 14 was analysed and the power spectral density (PSD) curve was generated, i.e. power distribution as a function of the EEG's constituent frequencies. For the PSD calculation, the nonparametric Fast Fourier Transform (FFT) algorithm was applied, using the technique as described by Welch et al. 15. The normalised power values, i.e. the absolute band power value as a fraction of the total power were calculated for the following frequency bands: 1) Delta frequency band: 1–4 Hz; 2) theta frequency band: 4–8 Hz; 3) alpha frequency band: 8–12 Hz; 4) sigma frequency band: 12–16 Hz.

Off-line analysis was performed using a software package (Dynesys Ltd, Edinburgh, UK). The DC-free EEG signal (sampled with 100 Hz) was filtered with a Finite Impulse Response (FIR) low-pass filter and a cut-off frequency at 25 Hz 16. For the calculation of the power density a 128 bin Hamming window with 64 points overlap and a resolution of 0.015 Hz was applied. For the delta, theta, alpha and sigma frequency bands the area under the spectral curve was calculated using a fifth-order Gaussian integration method, yielding power expressed in microVolt² 17 (fig. 1⇓).

The periods analysed centred around the end of apnoeas/hypopnoeas, identified as increase/recovery of the airflow amplitude. To assess power changes at apnoea/hypopnoea termination, comparisons were conducted between the synonymous frequency bands 10 s before versus 10 s after the end of each apnoea/hypopnoea.

The time windows were chosen to maintain signal stationarity within the applied window, as required for the application of the FFT algorithm. The 10 s window would in most cases avoid overlap between postapnoeic/-hypopnoeic intervals and the following early apnoeic/hypopnoeic periods. The further aim of the chosen windows was to maximise the signal-to-noise ratio by a compromise between occupying as much of the window as possible with potential arousals while at the same time allowing a reasonably wide window to cut down on random variance in the EEG signal (fig. 2⇓).

To determine whether the found power changes were caused by apnoeas/hypopnoeas, the authors compared the outcomes to periods of undisturbed sleep of the same duration in rapid eye movement (REM) and nonrapid eye movement (NREM) sleep, during which no apnoeas/hypopnoeas occurred. These periods were not necessarily free of snoring.

To assess the significance of the EEG power changes found across OSAHS patients these were compared to the EEG power spectrum of healthy subjects.

Statistical analysis

For the before versus after apnoea/hypopnoea termination-related comparisons as well as for the comparison between consecutive periods of undisturbed sleep across the patients, paired t-tests were performed. The Bonferroni correction of p-values for multiple comparisons was used and p≤0.01 was accepted as statistically significant during the paired comparisons.

The frequency band with the most significant power ratio (before versus after end) at apnoea/hypopnoea termination, as evaluated through paired comparisons, was used for the construction of receiver-operator-characteristic (ROC) curves 18. One ROC curve was constructed for each patient's polysomnogram. The ratio at the end of all apnoeas/hypopnoeas and the ratio around all points which separated two consecutive periods of undisturbed sleep were used for each curve. The area under the ROC curve determined how well the chosen ratio differentiated between apnoea/hypopnoea termination-related changes and undisturbed sleep (the bigger the area, the better the sensitivity).

For the comparison between OSAHS patients and healthy subjects, the mean power of the same frequency band as above was calculated for each polysomnogram using 10-s sliding windows; wake epochs and artefacts were excluded. Three ROC curves were constructed using the mean sleep-related power values of the patients and healthy controls: one for the total sleep time (TST), one for NREM sleep and one for REM sleep; each subject contributed one point to each curve. The area under the ROC curves determined how well the chosen frequency band power differentiated between OSAHS patients and healthy controls.

Independent t-test evaluated the differences between patients and healthy subjects in the band power found to be significant at the end of apnoeas/hypopnoeas.

One-tailed Pearson's correlation analysis evaluated the relationship between the apnoea/hypopnoea index (AHI) and the mean significant nocturnal band power as well as between AHI and time spent in slow-wave sleep (SWS) and REM sleep across the patients.

All comparisons and ROC curves were based on measurements within 10-s periods. A p-value <0.05 was accepted as statistically significant. Mean values and the sem or median and interquartile ranges (IQR) are quoted.

Across patients before/after comparison

All apnoeas/hypopnoeas, not affected by movement and/or EEG artefacts, have been included in the comparisons. In total, 2,596 apnoeas/hypopnoeas and 1,690 20 s (2×10 s) nonevent periods of undisturbed (baseline) sleep were evaluated across the patients. During NREM sleep 2,395 apnoeas/hypopnoeas versus 1,363 20 s baseline sleep periods were evaluated and during REM sleep 201 apnoeas/hypopnoeas versus 327 baseline periods.

The paired comparison, with each patient contributing one data point to each mean, showed that the postapnoeic/-hypopnoeic decrease in normalised theta power was the most significant (p<0.003) and consistent change during NREM and REM sleep, in arousal terminating and nonarousal apnoeas/hypopnoeas (tables 1 and 2⇓⇓). In addition, during NREM sleep alpha and sigma power increased significantly (p≤0.01) in arousal-terminating apnoeas/hypopnoeas, whereas in nonarousal apnoeas/hypopnoeas there was a significant increase in delta power (p≤0.005). During REM sleep, no significant power changes were found in nonarousal apnoeas/hypopnoeas.

During periods of undisturbed sleep no significant power changes (p≥0.2) were found across the patients (table 2⇑).

The mean normalised theta power correlated well with the patients' AHI (rho=0.6, p=0.008). AHI and time spent in SWS and in REM sleep were well reciprocally correlated (rho=−0.6, p=0.01; fig. 3⇓).

To assess how well theta power fluctuations can differentiate undisturbed sleep from apnoea-/hypopnoea-related EEG changes, one ROC curve was constructed for each patient. Each curve consisted of the theta_normalised power ratios (theta_{before_end}: theta_{after_end}) at the end of all apnoeas/hypopnoeas and around the point separating all the consecutive 10 s periods of undisturbed sleep used in the intra-individual comparisons. The areas under the ROC curves, determining the sensitivity of the method, varied between 0.6–0.9 across the 15 polysomnograms (table 3⇓).

Obstructive sleep apnoea/hypopnoea syndrome patients versus healthy subjects

Mean normalised theta power across patients was 25±0.5 normalised units (n.u.) and significantly lower than in healthy subjects 28±1 n.u. (F=2.4, p=0.03). In patients, mean delta power was 46±0.1 n.u., alpha power 17±0.05 n.u. and sigma power 12±0.02 n.u. In the healthy controls, mean delta power was 44±0.1 n.u., alpha power 17±0.1 n.u. and sigma power 11±0.1 n.u.; none of the three bands differed significantly between the two groups (p>0.2).

To assess how well theta power fluctuations can differentiate OSAHS patients from healthy controls, ROC curves were constructed. During NREM sleep the area under the curve was 0.76±0.1 (p=0.05), during REM sleep 0.82±0.09 (p=0.02) and 0.76±0.1 (p=0.05) during the TST (fig. 4).⇓

Spectral power, sleep states and arousal

The most significant and consistent change across the 15 patients was the decrease in normalised theta power at the end of apnoeas/hypopnoeas. During NREM sleep when arousals were visible, the decrease in normalised theta derived from the increase in alpha and sigma power. When arousals were not visible, the theta fraction decreased as a result of delta power increase.

Across the 12 patients, REM-related arousal terminating apnoeas/hypopnoeas were associated with decrease in theta power fraction alone, which indicates a predominant decrease in the absolute theta power. The 87 REM-related apnoeas/hypopnoeas which were not terminated by visible arousals, were not associated with any significant spectral power changes. This may be due to the limited number of these events across the 12 patients (mean: 6, range: 1–21 per patient). Furthermore, it suggests that during REM sleep, apnoea/hypopnoea termination evolves subcortical regions but does not always evoke a cortical reaction.

The mean theta power fraction was significantly lower in the patient group compared to the healthy subjects as a result of the overall increase in sigma and delta power and despite the similar time the two groups spent in NREM 1 and 2, SWS and REM sleep.

As the subjects were matched for age and BMI, the findings suggest that the differences in theta power fraction between the two groups are related to the postapnoeic/-hypopnoeic power decrease and the recurrent visible and nonvisible respiratory arousals. These frequent fluctuations of spectral power may contribute to the patients' daytime symptoms. To assess the significance of this assumption, comparisons between objective daytime measurements (such as MSLT) and nocturnal EEG power spectrum would be necessary.

The present findings are consistent with the outcomes of Svanborg et al. 9 who analysed the EEG power spectrum in the course of apnoeic/hypopnoeic episodes based on 2-s epochs and found an increase in delta power towards the end of apnoeas/hypopnoeas and a decrease at their termination as faster activities became more prominent. Berry et al. 6 identified a cyclical delta power increase in sleep apnoeics, parallel to increased inspiratory effort. The EEG spectral analysis was performed using sequential 1 s windows.

Both studies identified differences between NREM and REM sleep, as no significant changes in delta band power were identified during REM-related apnoeas/hypopnoeas.

Black et al. 19 analysed the EEG power spectrum around the peaks of oesophageal pressure (P_oes) using 4-s windows, during NREM 2 in 15 upper airway resistance syndrome patients, and similarly identified a progressive increase in delta power preceding those peaks. An increase in the power of higher frequencies was identified at P_oes reversal. Although changes were more significant at arousal terminating “events”, they were also present when arousals were not visible.

The present study focused around the end of apnoeas/hypopnoeas and did not look into EEG power changes in the evolution of these events. Therefore, the study could not confirm the previously found significant delta power increase towards the end of apnoeas/hypopnoeas, identified through consecutive comparisons of short intervals prior to and during apnoeas/hypopnoeas.

The aim of the present study was to improve detection of EEG changes at the end of apnoeas/hypopnoeas. Using longer windows, the analysis looked into changes in delta, theta, alpha and sigma power and their interaction and delivered 10-s based power values. The most consistent finding across the 2,596 events studied, is the theta power decrease. In the majority of NREM apnoeas/hypopnoeas this resulted from the increase in higher frequencies, which made the EEG changes visually detectable. The rise in high activity is consistent with the findings of Svanborg and Guilleminault 9 and Black et al. 19.

Furthermore, the present study found that events with nonvisible EEG changes and no alpha/sigma increase induced a significant delta increase. Hence, this rise in slow activity in response to apnoeas/hypopnoeas represents, as previously suggested 9, a form of arousal. DeCarliet al.20 support the postevent alpha and/or theta increase but not delta alone, based on the outcomes of wavelet transform, a different technique of EEG analysis.

The present study confirms the previously found power differences between REM- and NREM-related apnoeas/hypopnoeas: visible changes in REM sleep were associated with significant theta decrease alone, nonvisible changes showed no significant power changes; hence, no changes in delta power have been detected 6, 9. The present findings provide information on differences between arousal and nonarousal REM- and NREM-related apnoeas/hypopnoeas. Previous studies did not look at these differences either because only NREM-related apnoeas/hypopnoeas were studied 19 or because differences in respect to the visibility of arousals were not evaluated 9.

The increased EMG activity during visible cortical arousals could potentially alter the EEG power spectrum and induce differences between visible versus nonvisible arousals. Muscle artefacts (movement and breathing) would influence alpha, sigma and low delta (0.2–2.2 Hz) but not theta frequency band power, as Brunner et al. have demonstrated 21. However, the importance of this factor would definitely be in certainty evaluated through the elimination of EMG influences from the EEG signal, which was not performed during the present study.

Sleep quality

Sleep quality was not significantly different between patients and healthy controls. This was due to: 1) the impaired sleep quality of healthy subjects under in-lab conditions; and 2) the inconsistency among patients as to their disease severity ranging from mild, moderate to severe, yielding differences in their sleep quality. The latter is supported by the good reciprocal correlation found between AHI and time spent in SWS and REM sleep and the good correlation between AHI and theta power across the patients. The differences in sleep quality between patients and healthy controls are therefore better preserved through the range of time spent in each sleep state.

Study limitations

Limitations of this study include definitions, the use of thermistor for airflow detection, the unequal number between apnoeas/hypopnoeas and periods of undisturbed sleep used in the comparisons, as well as the number of subjects studied. A further limitation may be the use of only two bins: NREM and REM sleep.

There are many definitions of hypopnoeas and arousals used in the literature. The hypopnoea definition presently used, fulfils the ASDA recommendations 11. The inclusion of arousals or desaturation in the definition may lead to a greater association with arousals than would be found using other ASDA accepted definitions, i.e. based on thoraco-abdominal movement or nasal pressure alone. The arousal definition used, has been used in previous studies and validated against respiratory events and outcomes 4. The thermistor is a widely used good qualitative detector of airflow detection. It is unlikely that the limited accuracy of this sensor may have altered the outcomes of the EEG spectral analysis. This is supported by the fact that the findings are in keeping with the outcomes of Black et al. 19, who used oesophageal manomentry to monitor changes in respiratory effort.

All the 2,596 apnoeas/hypopnoeas and 1,690 undisturbed sleep periods, which were not affected by artefacts, across the 15 polysomnograms were analysed. This represents a large body of data with which to answer the questions posed. The present results are not only robust but represent the largest body of events which have been so analysed. The study is thus adequately powered to detect clinically important differences.

The two bins, NREM and REM sleep, were selected as a result of their different characteristics: REM sleep is a low voltage, high frequent state; its baseline, nonevent, power spectrum is different from the NREM spectrum. However, it was beyond this paper's aims to look into apnoea/hypopnoea-related differences across the four different NREM sleep stages.

In conclusion, the present study has shown that novel techniques may improve and simplify detection of electrographical changes at apnoea/hypopnoea termination. The most consistent significant change is the decrease in theta power, which during NREM sleep is either associated with an increase in high frequencies (alpha, sigma) or delta increase.

The detection of theta power changes is, as the ROC curves demonstrate, a sensitive method to pick up apnoea-/hypopnoea-related EEG changes and also to differentiate between OSAHS patients and healthy subjects. Since part of the REM-related apnoeas/hypopnoeas is not associated with significant theta power changes, the proposed method may only be used as part of a tool for improved and simplified diagnosis of OSAHS.

Further validation, estimation of a cut-off value for the differentiation between patients and nonpatients and correlation analysis between this electroencephalographic activity marker and sleepiness severity as well as cognitive function impairment, are necessary.