We compared the microstate, power spectrum, sample entropy and DFA in four sleep stages between OSA patients and healthy controls. We found that OSA patients had an additional fifth microstate E in N1-OH/OA and N3-OH/OA stages, and microstate E was related to delta, theta and alpha power spectrum and DFA in N1-OH/OA stage.
As shown in Table 1, OSA may occur in any stage of sleep cycle, but the number of OH/OA occurrences in N1 and N2 stages is more than that in N3 stage. The pathogenesis of OSA is upper airway stenosis and obstruction3. Breathing can stop temporarily for ten seconds or even several minutes, and it can happen hundreds of times a night, leading to hypooxia, which makes the patient wake up suddenly and is difficult to enter deep sleep. Therefore, because the frequent occurrence of OH/OA causes the patient to repeatedly switch between light sleep stage (N1, N2) and wakefulness, it is difficult to enter the deep sleep stage (N3).
Microstate reflects the instantaneous state of the brain, and can identify discontinuous and nonlinear changes of global functional brain state under very high temporal resolution12, 15. It has been found that four canonical microstates A, B, C and D are related to the activities of the posterior cingulate cortex21. Combined fMRI-EEG imaging studies have shown that the neural components that generate microstates overlap with the resting-state network independently recognized by fMRI22, 23, 24. Brodbeck et al. investigated wakefulness and NREM sleep of healthy subjects, and found that microstate C was dominant in W, N1 and N3 stages, while microstate B was dominant in N2 stage. With the increase of sleep depth, the parameter GEV of microstate D gradually decreased16. Kuhn et al. investigated early NREM sleep of narcoleptic patients, and found that microstate C and D were dominant in N1 stage, microstate D was still dominant in N2 and N3 stages, and an extra microstate E appeared in N3 stage17. The authors thought that the occurrence of extra microstate E in N3 stage indicated that the corresponding neural network activity was unstable and led to fragmented sleep structure17.
Through the microstate analysis in four sleep stages in OSA patients and healthy controls, we found that OSA patients had an additional fifth microstate E in N1-OH/OA and N3-OH/OA stages, and the duration of microstate C and D was longer. The possible reason is that due to the frequent occurrence of OH/OA, the activity of the neural network that generates microstates C and D is unstable or has changed in N1-OH/OA and N3-OH/OA stages, which in turn splits the microstate E. The GEV of microstate E is low, only 2.3 ~ 2.5%, and MeanDur is also short, 22.2 ~ 22.3 ms. After fitting the microstates A, B, C, D and E back to the original EEG signals, we found that the microstate E only existed on a few patients' EEG signals. This shows that the probability of occurrence of microstate E is lower and the variability is relatively large among different subjects. In addition, we found that OSA patients had an additional microstate E in N1-OA/OH and N3-OA/OH stages, but microstate E in N1-OA/OH stage was related to delta, theta and alpha power spectrum and DFA, not related in N3-OA/OH stage. The possible reason is that the OSA patients we screened have large differences among individuals or because our sample size is too small, resulting in insufficient statistical power. In future studies, increasing the sample size of OSA patients may be able to further verify this result.
There were previous studies on the difference of wakefulness EEG power between OSA patients and healthy controls8, 9, but there were few studies on the changes of sleep EEG power in OSA patients. Delta rhythm (0.5 ~ 4.5 Hz) mainly occurs in deep sleep or coma, during which human is difficult to wake up, and human’s cortical activities lose sensory input, that is, the separation of cortical activity and thalamic activity. We found that the delta band power of OSA patients in N3-OA/OH was higher than healthy controls (P < 0.021). Grenèche and D’Rozario also found delta band power increased in OSA patients in wakefulness state8, 9. These studies show that delta power increases in OSA patients both in wakefulness and sleeping states. Monegro et al. investigated delta power in OSA patients before and/or after therapy with Continuous Positive Airway Pressure (CPAP), and found that there was an overall decrease in delta power in patients with a higher Respiratory Disturbance Index (RDI) after CPAP25. These studies show that hypoxia is related to the changes of brain delta power. It is known that there is a coupling relationship between neuroelectrophysiology and hemodynamics, and the cerebral blood flow and intracranial pressure increase in OSA patients due to a certain degree of hypoxia, which may cause delta power increase. The cerebral hypoxia is relieved after CPAP, and the relief effect is more obvious for patients with higher RDI. In addition to hypoxia, EEG changes are also related to age, including increased slow wave and lower alpha power26, so the slower EEG in OSA patients may not only be caused by sleep disorders, but may also be related to age9. In our study, the age information of 10 healthy subjects is missing in the ISRUC-Sleep data set, so further analysis of age factors is not possible.
Beta rhythm (the frequency range of beta rhythm in some literatures is 13 ~ 30 Hz, which is divided into sigma (12 ~ 15 Hz) and beta (15 ~ 32 Hz) in this study) mainly appears in the active state of brain such as active thinking. In the late stage of light sleep (N1), low-amplitude beta waves may appear. In our study, the beta power of OSA patients was lower than healthy controls in N1-OA/OH stage (P < 0.026), and sigma power of OSA patients was lower than healthy controls in N2-OA/OH stage. Grenèche et al. found that the beta power of wakefulness EEG in OSA patients was higher than healthy subjects, but in their study, OSA patients were during resting state and did not move or imagine movement8. D’Rozario et al. did not report beta band power 9. Previous studies show that beta rhythms decrease (i.e., event-related desynchronization (ERD)) during movement imagery, movement preparation and movement execution27. We infer that OSA patients during the N1 and N2 stages with frequent apneas, by changing their body posture to relieve the discomfort caused by apneas, and thus show a decrease in beta power.
Through Pearson correlation analysis, microstate E was related to delta, theta and alpha power in N1-OH/OA stage, and the sigma and beta power were independent of all microstates in all sleep stages. Previous studies revealed no conclusive results concerning the association of the four EEG microstate classes with specific power spectral distributions22, 23. However, Milz et al. investigated head-surface localization- or source-dependent power effects on the occurrence of the EEG microstate classes, and found that the EEG microstate topography was predominantly determined by intracortical sources in the alpha band28. Croce et al. investigated EEG microstates associated with intra- and inter-subject alpha variability, and observed an increase in the metrics of microstate B, with the level of intra-subject amplitude alpha oscillations, together with lower coverage of microstate D and a higher frequency of microstate C29. Although their study found the relationship between alpha power and microstate metrics, the authors also pointed out that there was no specificity for alpha power. The modulation effect on microstate metrics is not unique to the alpha band. It may be caused by fluctuations in other frequency bands29. Therefore, we can infer that when apnea occurred in N1 and N3 stages, the intensity and spatial distribution of alpha activity in the cortex changed, which induced microstate E. Or because of the mediation by alpha band and the dynamic interaction with other bands (such as delta, theta bands), the original scalp potential distribution is changed, and an additional microstate E is generated.
Sample entropy is an improved method for measuring the complexity of time series, and it has applications in evaluating the complexity of physiological time series and diagnosing pathological state18, 30, 31. Zhou et al. found that the sample entropy of sleep apnea syndrome patients was lower than that of healthy controls in each sleep stage7. We found that the sample entropy of OSA patients was lower than healthy controls in four sleep stages (P < 0.05). According to the dynamic theory of entropy, the greater the entropy is, the higher the complexity of the sequence. Low entropy indicates that the data is highly predictable and regular, while high entropy indicates that the data is chaotic and unpredictable. Meaningful and complex data have intermediate entropy. When human is awake, the brain receives information from the outside world and is still carrying out complex thinking activities, and the neurons are active, therefore EEG signals present complex randomness. With the deepening of sleep, the activity of neurons in the brain gradually decreases and the synchronization orderly increases. The EEG signals show more regular characteristics of self-regulation, of which the complexity decreases. In REM stage, along with physiological activities such as dream and eye rotation, the brain nerve activity increases and the EEG complexity increases7. Therefore, the change of sample entropy reflects the physiological mechanism of sleep, as shown in Fig. 5 (a). The sample entropy of OSA patients is lower than healthy controls. The possible reason is that brain hypoxia or other pathological conditions caused by OSA have a significant impact on brain nerve activity, resulting in a significant decrease in the activity and complexity of brain cells7.
Through Pearson correlation analysis, our results showed that the microstate parameters GEV, MeanDur, TimeCov and SegDensity in OSA patients were not correlated with the sample entropy in four sleep stages, and their p values are shown in Table 5. Murphy et al. employed sample entropy to calculate the complexity of the microstate sequence over the entire template length in subjects with psychotic disorders18. They thought that the transition of microstate was independent of the duration of microstate, so they deleted the information of microstate duration. In this way, the microstate sequence was compressed. Their results showed that there was no correlation between sequence length and entropy in psychiatric patients and healthy controls, and there was no statistically significant correlation between entropy and microstate duration at any pattern length of sample entropy18. While our study showed that sample entropy was not only independent of the duration of microstates, but also irrelevant to GEV, TimeCov and SegDensity of microstates. The difference was that Murphy et al. calculated the sample entropy on the compressed microstate sequence with template length m = 3 ~ 10, while we calculated the sample entropy on the preprocessed EEG signal with template length m = 2. Although these studies differ in details, they all show that the microstate sequence has nothing to do with the complexity of EEG signals.
EEG signal has a long-term correlation of dynamic oscillation characteristics32, 33. Detrended fluctuation analysis (DFA) quantifies the time-domain fluctuation of time series by power-law method, and describes the scaling behavior or long-range correlation of time series with scale index, which is suitable for studying the correlation of long-range power-law functions of various unstable time series. Our study showed that the scale index of OSA patients and healthy controls was 0.5 < α < 1.0, which indicated that there was a long-range power-law continuous correlation of EEG signal (with self -similarity of fractal dimension). The scale index α of OSA patients was higher than healthy controls in four sleep stages, and the scale index α of OSA patients in N1-OH/OA and N3-OH/OA was higher than that of N2-OH/OA and R-OH/OA, as shown in Fig. 5 (b). D’Rozario et al. found that the DFA of the OSA patients was higher than healthy controls with eyes opening and closing, and the DFA of the two groups with eyes opening was higher than that with eyes closing9. Our study results were basically consistent with theirs. From the fractal dynamics, when the scale index α is between 0.5 and 1, the larger the value is, the stronger the self-similarity regularity of the signal is, which means that the information entropy is lower (when the value is 1, it is 1/f noise); conversely, the smaller the value is, the stronger the randomness of the signal is, which means that the information entropy is higher (when the value is 0.5, it appears as white noise) 34. This indicates that the EEG signals of OSA patients have stronger self-similar regularity than healthy controls, especially in N1-OH/OA and N3-OH/OA stages. The oscillation mode is smoother, which may be caused by abnormal changes of brain activity conversion in OSA patients in these two stages. The reasons for the differences in the long-range power function correlations of EEG signals in different populations and in different physiological and pathological states have not been fully clarified. Some studies have speculated that this was related to the special mechanism of neural oscillation. The differences in scale behavior or long-range correlation reflect that neuronal oscillations may be affected by special mechanisms related to their origin33.
Through Pearson correlation analysis, our results showed that microstate C, D and E were correlated with DFA in OSA patients in N1-OH/OA stage, and GEV, MeanDur, TimeCov and SegDensity of microstate E were positively correlated with DFA. Previous studies have shown that the microstate of healthy subjects exhibit scale-free and self-similar dynamic characteristics 35. Murphy et al. carried out fractal analysis on the microstate of psychiatric patients, and found that the microstate sequence has a long-term time-dependent18. This shows that the microstate is the same as the scale index α, which reflects the self-similarity of EEG signals.
At present, there is no simple and effective EEG biomarker that can reflect the negative impact of OSA on the brain, although D’Rozario et al. have shown that the DFA scale index has the potential as an EEG biomarker of neurobehavioral damage9. However, their study only compared DFA and power spectrum, and lacked comparative analysis with other EEG biomarkers (such as sample entropy, microstate, etc.). In addition, they only considered the single scene of simulated driving, and lacked the research on sleep EEG and its prognostic value. In our study, the sleep EEG of OSA patients was analyzed by the microstate method and the correlation analysis with power spectrum, sample entropy and DFA was carried out, and the additional microstate E was found, which was related to the power spectrum of delta, theta and alpha bands and DFA. This shows that the microstate also has the potential as a biomarker of OSA EEG. Zappasodi el al. investigated prognostic value of EEG microstates in acute stroke, and found that a preserved microstate B in acute phase correlated with a better effective recovery36. Therefore, whether there is a correlation between microstate and OSA score and whether it has prognostic value for OSA patients is our next research work.
Finally, fewer EEG channels were emplyed for the microstate analysis in our study, only 6 channels37, but previous studies have demonstrated that canonical microstate topographic maps are not limited by low spatial sampling18, 38. In the past 20 years, researchers have conducted many studies using microstate analysis, exploring the topography, duration, frequency and transition probability of microstates. But the interpretation of these studies is complicated by the small sample size and the differences in analytical techniques. Even in the study of canonical four microstates, there are major differences regarding the microstate topographic maps, especially for microstates C and D39. Michel et al. considered that the number of topographic map should not be a fixed value (for example, 4), but the optimal number of topographic map should be estimated according to the specific situation of data sets39. Therefore, the results of microstate analysis, especially the number of microstates generated, should be treated with caution. The limitation of this study is that the sample size is limited. OSA patients with other complications and taking drugs were excluded in our study. Whether these two types of patients affect the conclusion of this paper needs further study.