Spontaneous Hemodynamic Oscillations during Human Sleep and Sleep Stage Transitions Characterized with Near-Infrared Spectroscopy

Understanding the interaction between the nervous system and cerebral vasculature is fundamental to forming a complete picture of the neurophysiology of sleep and its role in maintaining physiological homeostasis. However, the intrinsic hemodynamics of slow-wave sleep (SWS) are still poorly known. We carried out 30 all-night sleep measurements with combined near-infrared spectroscopy (NIRS) and polysomnography to investigate spontaneous hemodynamic behavior in SWS compared to light (LS) and rapid-eye-movement sleep (REM). In particular, we concentrated on slow oscillations (3–150 mHz) in oxy- and deoxyhemoglobin concentrations, heart rate, arterial oxygen saturation, and the pulsation amplitude of the photoplethysmographic signal. We also analyzed the behavior of these variables during sleep stage transitions. The results indicate that slow spontaneous cortical and systemic hemodynamic activity is reduced in SWS compared to LS, REM, and wakefulness. This behavior may be explained by neuronal synchronization observed in electrophysiological studies of SWS and a reduction in autonomic nervous system activity. Also, sleep stage transitions are asymmetric, so that the SWS-to-LS and LS-to-REM transitions, which are associated with an increase in the complexity of cortical electrophysiological activity, are characterized by more dramatic hemodynamic changes than the opposite transitions. Thus, it appears that while the onset of SWS and termination of REM occur only as gradual processes over time, the termination of SWS and onset of REM may be triggered more abruptly by a particular physiological event or condition. The results suggest that scalp hemodynamic changes should be considered alongside cortical hemodynamic changes in NIRS sleep studies to assess the interaction between the autonomic and central nervous systems.


Introduction
Sleep is a part of constantly ongoing homeostatic regulation between neurons, glial cells, and vasculature of the brain to maintain health, adaptability, and cognitive performance [1]. Slow-wave sleep (SWS), also known as deep sleep, differs from the physiologically distinct light sleep (LS) and rapid-eye-movement sleep (REM) stages by a dramatic increase in synchronized neuronal activity and loss of information integration across cortical areas [2,3]. This activity plays a crucial role in cortical plasticity and is seen in electroencephalographic (EEG) recordings as largeamplitude delta waves (0.5-2 Hz) [3,4]. LS and particularly REM generally display much more complex and chaotic EEG patterns [2]. However, the brain also exhibits slow physiological oscillations which cannot be studied with EEG alone [5]. To form a comprehensive understanding of the neurophysiologic interactions related to sleep, bioelectric signals must be supplemented with hemodynamic and metabolic data [6][7][8][9][10].
Of the methods available for studying cerebral hemodynamics and metabolism, positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) are poorly suited for all-night sleep measurements. Both methods require the subject to lie still, and in the case of PET also ionizing radiation and poor temporal resolution limit applicability. Transcranial Doppler sonography (TCD) is particularly well suited for monitoring global cerebral blood flow (CBF), but it is insensitive to blood oxygenation changes and does not reveal regional CBF (rCBF). In contrast to the aforementioned methods, near-infrared spectroscopy (NIRS) tolerates small movements of the head, can measure cortical hemodynamic and oxygenation changes locally, and allows tracking the concentration changes of both oxygenated (D[HbO 2 ]) and deoxygenated (D[HbR]) hemoglobin (Hb). These features, combined with comfortability, portability, and relatively low cost of equipment, make NIRS an excellent method for monitoring hemodynamics during sleep [11][12][13][14].
NIRS sleep studies have primarily concentrated on investigating sleep stage transitions, often during daytime napping [10,26], baseline changes in hemodynamic parameters between sleep stages [11,[27][28][29], or the effects of sleep-disordered breathing on cerebral hemodynamics [14,30]. Almost no attention has been paid to transitions related to SWS [27], and the behavior of LFOs and VLFOs in different sleep stages have not been studied previously with NIRS.
In the present study, we examine spontaneous hemodynamic oscillations measured with NIRS in different sleep stages. We also investigate hemodynamic changes during transitions between sleep stages, particularly transitions to and from SWS, and contrast cortical, scalp, and systemic hemodynamic parameters with each other to determine how they are interrelated.

Participants
All-night NIRS-EEG sleep data were gathered from 13 healthy volunteers (9 males, 4 females, mean age 26 years, range 21-32). Each subject was measured during at least two nights to reduce the influence of the unfamiliar surroundings on sleep quality; however, no such influence was observed when comparing sleep time to total measurement time on the first and second measurement night. A total of 30 all-night measurements were conducted, with five measurements discarded due to poor quality of either NIRS or EEG data.

Ethics
All subjects gave their written informed consent prior to the measurements. The study was approved by the Ethics Committee of Helsinki University Central Hospital and was in compliance with the Finnish legislation and the declaration of Helsinki.

Data Acquisition
The NIRS data were recorded with a frequency domain instrument using 685-nm and 830-nm light [31]. An optical probe with a single light emitter fiber and three detector fiber bundles was placed on the right side of the subject's forehead just below the hairline, above the right prefrontal cortex (Fig. 1). The emitterdetector separations were 1, 4, and 5 cm, with the 1-cm channel measuring mainly scalp hemodynamics and the 4-and 5-cm channels probing also cortical hemodynamics [32]. Light attenuation changes in tissue were converted into D[HbO 2 ] and D[HbR] with the modified Beer-Lambert law (MBLL) [33]. The sum D[HbO 2 ]+D[HbR] is an indicator of blood volume changes, whereas blood flow changes cannot be directly inferred from NIRS signals. However, due to the passive nature of blood transport from the brain, cerebral blood volume (CBV) changes are often indicative of parallel CBF changes [34].
MBLL requires setting the photon pathlength in tissue and the specific extinction coefficients of HbR and HbO 2 . Although the frequency domain technique allows estimating the pathlength separately for each individual [35], technical difficulties prevented this in ten measurements. Thus, an average differential path length factor was estimated from the available phase data (20 measurements) and used for calculating the photon pathlength for all subjects. The estimated differential path length factor was 6.16 for 685 nm and 5.84 for 830 nm. Values for the specific extinction coefficients of HbR and HbO 2 were taken from literature [36].
An accelerometer attached to the NIRS probe recorded head movements that might disturb the tissue-optode contact and thus introduce motion artifacts to the data (Fig. 1). An automated algorithm was used to identify and correct motion artifacts that baselines [37]. These discontinuities were removed on the assumption that actual physiological concentration baseline changes during brief movements are negligible. This correction allowed tracking concentration changes over sleep stage transitions involving motion.

D[HbO 2 ] and D[HbR]
were estimated at a sampling rate of approximately 10 Hz and resampled to a uniform rate of exactly 10 Hz with spline interpolation. Oscillations related to respiration and heart beat were removed by low-pass filtering the concentration signals (type II Chebyshev filter, passband edge at 150 mHz). The 1-cm channel was selected to represent extracerebral hemodynamics and the 4-cm channel to represent cortical hemodynamics, since it had a better signal-to-noise ratio than the 5-cm channel. Otherwise, the 4-cm and 5-cm signals behaved similarly.
An all-night polysomnogram consisting of EEG (channels C4-A1, O2-A1, C3-A2, and O1-A2), left and right electrooculograms (EOG), and a chin electromyogram (EMG) was recorded from all subjects (Fig. 1). Based on the polysomnogram, a neurophysiologist scored the sleep into S1, S2, S3, S4, REM, wake (W), and movement time stages in 30-s epochs according to the Rechtschaffen-Kales rules [38]. For data analysis, the S3 and S4 non-REM (NREM) stages were combined into SWS, and the S1 and S2 NREM stages into LS. Also, if two similarly scored 30-s epochs were separated by one 30-s epoch scored differently, that epoch was rescored to match its surroundings. This classification corresponds closely to the new sleep scoring guidelines given in the American Academy of Sleep Medicine scoring manual [39], and clearly distinguishes fundamentally different EEG states from each other [2]. It also produces longer homogeneously scored periods, facilitating the analysis of VLFOs.
A finger pulse oximeter measured the photoplethysmographic waveform and the 10-s average of the peripheral arterial oxygen saturation (SpO 2 ). The peak-to-peak pulsation amplitude of the photoplethysmographic waveform (PPGamp) and HR were derived from the pulse oximeter signal and interpolated to a uniform sampling rate of 1 Hz and filtered similarly to the NIRS signals (type II Chebyshev filter, passband edge at 150 mHz).
PPGamp reflects the amount of blood pulsating in the blood vessels of the finger and normally decreases with vasoconstriction and increases with vasodilation [40]. PPGamp is also negatively correlated with systolic BP in normal sleep, although this correlation decreases in REM sleep [41].

Data Analysis
Spontaneous hemodynamic oscillations. For each continuous period of SWS, LS, REM, and W, D[HbO 2 ] and D[HbR] signal power in the VLFO and LFO bands was estimated with Welch's power spectral density (PSD) method with 6.8-min segment length (2 12 samples of NIRS data), 50-% segment overlap, and 1.2-mHz frequency resolution. Only sleep periods lasting at least 13.7 min (2 13 samples of NIRS data) were included, so that at least three segments were averaged for each PSD (Table 1). Logarithmic signal powers were tested for statistically significant (p,0.05) differences between sleep stages with one-way ANOVA followed by Tukey's honestly significant difference (HSD) post-hoc test. Separate tests were conducted for each frequency band, NIRS channel, and Hb species. To account for multiple comparisons, the ANOVA p-values were controlled with the false discovery rate (FDR) method [42]. The PSDs of HR, PPGamp, and SpO 2 were analyzed similarly, except the LFO band for SpO 2 was restricted to 40-50 mHz because of the 100-mHz sampling rate, and multiple comparisons were controlled for only over the two frequency bands.
Sleep stage transitions. Sleep stage transitions were analyzed for data segments where the sleep stages preceding and following the transition were maintained for at least 300 s. Transitions not involving LS were omitted due to their rarity. For each transition, the means of D[HbO 2 ], D[HbR], HR, PPGamp, and SpO 2 for the 300 s preceding the transition were set to zero. Then, time courses of these five parameters were averaged for each of the six transition types (SWSRLS, LSRSWS, LSRREM, REMRLS, WRLS, and LSRW).
In addition to the average time courses, transition-related baseline changes in all five parameters were investigated. The mean value of each parameter between 300 and 200 s before the transition was subtracted from the mean between 200 and 300 s The contribution of extracerebral hemodynamics to the 4-cm signals was evaluated with principal component analysis (PCA) of the 1-and 4-cm average time courses. Briefly, the method separates statistically uncorrelated components from NIRS data, so that it allows identifying hemodynamic changes that are present in the brain but not in the scalp [43].  transitions were then averaged to arrive at an average time course representing combined VLFO and LFO amplitude changes before and after the transition. The first and last segments were compared for statistically significant differences with Student's t-test with the FDR adjustment.
All signal processing was carried out in MATLAB (The MathWorks, Inc., Natick, Massachusetts, USA). Data segments clearly containing large motion or signal processing artifacts were not analyzed.

Spontaneous hemodynamic oscillations
Spectral analysis of the NIRS and systemic signals showed that the VLFO and LFO powers in SWS mostly differed statistically significantly from the powers in the other stages, particularly in the VLFO band (Fig. 2). In both bands, the power level increased typically in the order SWS-LS-REM-W, although the differences between LS and REM were mostly non-significant. Oscillation power was also higher in D[HbO 2 ] than in D[HbR]. Figure 3 shows the evolution of hemodynamic signals in the VLFO and LFO bands in the SWSRLS transition for a typical subject. The signals were relatively stable in SWS, with hemodynamic activity clearly increasing after the transition to LS, particularly in the 4-cm detector. This, along with similar time and frequency domain results from other subjects, indicates suppression of hemodynamic oscillations during SWS.

Sleep stage transitions
Average D[HbR] and D[HbO 2 ] time courses for the SWSRLS and LSRREM transitions indicated an increase in blood volume in the 1-and 4-cm NIRS channels (Fig. 4). These increases were accompanied by a simultaneous increase in HR and a decrease in PPGamp. The changes related to the SWSRLS transition appeared to be mostly transient, except for the increase in D[HbO 2 ], while changes at the LSRREM transition were more persistent. The opposite transitions, LSRSWS and REMRLS, were involved with smaller and opposite changes, or no changes at all. The WRLS transition was accompanied by an increase in D[HbR] and decreases in D[HbO 2 ], HR, and SpO 2 , while opposite changes arose in the LSRW transition. In the latter case, the changes were mostly not statistically significant due to the small number of transitions recorded.
Removing extracerebral contribution with PCA affected the interpretation of the 4-cm signals mainly in the SWSRLS transition by removing the prominent transient changes (Fig.  S1). However, when components corresponding to both HR and the 1-cm signal were removed with PCA, no apparent transitionrelated responses were present in any of the remaining 4-cm signals (Fig. S1). Removing components corresponding to PPGamp and SpO 2 did not have a similar effect (not shown).
SD changes in sleep stage transitions indicated differences in spontaneous hemodynamic activity primarily between SWS and LS (Fig. 5). This difference was clearly visible in both LSRSWS and SWSRLS transitions. In the LSRSWS transition, the level of spontaneous activity appeared to change gradually, whereas in the SWSRLS transition the transient changes visible in the average time courses (Fig. 4) dominate. However, also in the SWSRLS transition the level of spontaneous activity appeared to be higher after the transient. The results are in line with the PSD analysis of different sleep stages where the power of VLFOs and LFOs decreased in SWS compared to other stages. No statistically significant SD differences between LS and W or LS and REM were observed in the NIRS signals.

Discussion
Our investigation revealed two major findings. First, slow spontaneous hemodynamic activity was greatly reduced in SWS compared to other sleep stages, while differences between LS and REM were relatively small. The reduction occurred simultaneously with the transition to SWS and was reversed with the transitioning away from SWS.

Reduction of spontaneous hemodynamic activity in SWS
The reduction in the slow hemodynamic oscillations in SWS coincides with the presence of delta waves in the EEG, suggesting a possible common origin for them. Since the delta waves are generated by neuronal activity that is relatively constant over the local cortical area [4], variations in rCBF could also be expected to be suppressed in SWS. Since suppression of hemodynamic oscillations was evident in both systemic and cortical signals during SWS, the question may also be raised whether there is a synergetic benefit from this simultaneous reduction. For example, since SWS has been associated with diminished vasoreactivity to hypoxia [45], this could necessitate the suppression of systemic hemodynamic oscillations to ensure constant CBF and cerebral oxygenation.
In contrast to our results on HR oscillations, in a previous study the relative contributions of VLFOs and LFOs in HR variability were observed to increase in REM sleep compared to NREM, while there were practically no such differences between LS and SWS [46]. However, the same study indicated that the absolute spectral power increases from SWS to LS to REM, which corresponds to our results. This latter pattern has also been observed for sympathetic nervous system activity, which plays a key role in the regulation of systemic hemodynamic variables and homeostasis [47].

Sleep stage transitions
This is the first study to present a comprehensive account of the time evolution of hemodynamic parameters in sleep stage transitions. Our observations on the WRLS and LSRW transitions are consistent with a general reduction in physiological activity and alertness in sleep compared to wakefulness, and in agreement with previous studies [10,11,[26][27][28]. In some cases, our results on sleep stage transitions cannot be directly compared with the results of previous studies since they have compared hemodynamics and metabolism between REM, SWS, and W instead of comparing these states to LS [27,48]. In some studies, hemodynamic parameters such as rCBF have been studied as a function of EEG delta activity instead of sleep stage transitions [49][50][51].
In the SWSRLS transition, the systemic parameters were the first to react, suggesting peripheral vasoconstriction and an increase in HR. The sudden and transient nature of these changes suggests that the transition may be linked to a transient arousal [47]. The shift in systemic parameters is followed by NIRS signal changes indicating an increase in the cortical and scalp blood volumes. In contrast, TCD evidence suggests that arousals during NREM sleep not resulting in awakening are more likely to lead to a decrease in CBF and CBV [52]. We speculate that the hemodynamic changes at the SWSRLS transition may also play a significant role in the termination of SWS, e.g., by actively stimulating neurons into moving away from the delta wave activation pattern.
In the LSRREM transition, the changes in both systemic parameters and NIRS signals persisted well into the REM period, which may reflect the higher sympathetic activity in REM compared to other sleep stages [11,47]. The onset of REM sleep is accompanied by rapid eye movements which have been reported to activate the prefrontal cortex in a previous NIRS study [11]. Thus, changes in cortical hemodynamics could be related to cerebral activity produced by dreaming. However, rapid-eye-movement-related fMRI studies have not shown activity in the prefrontal cortex close to our measurement site [53,54].
A statistically significant change in the amplitude of VLFOs and LFOs was visible in all hemodynamic variables within 300 s of the LSRSWS transition, as well as in all variables except D[HbR] in conjunction with the opposite SWSRLS transition. The onset of SWS was primarily characterized by a reduction in the amplitude of slow hemodynamic oscillations, not as a baseline change in hemodynamic parameters. Since the oscillations are to at least some extent controlled by feedback loops in the autonomic nervous system, the results may be interpreted to reflect differences in homeostatic regulation between SWS and the other stages.
Many sleep disorders, such as obstructive sleep apnea and the restless legs syndrome, lead to an increase in the frequency of sleep stage transitions to LS and W, and a decrease in the duration of continuous periods spent in one stage [55,56]. Since sleep stage transitions interfere with the prevalent homeostatic oscillation patterns, an increase in the transition frequency could contribute to the detrimental effects of sleep disorders by disrupting and damaging the homeostatic feedback system over the long term. Decreased cardiovascular oscillations have been associated with various diseases, including diabetes, obstructive sleep apnea, and cardiovascular mortality [57].

Limitations
Our data are based on hemodynamic recordings from the right prefrontal cortex. While this is an inviting location to measure with NIRS during sleep due to the lack of hair and reduced risk of probe displacement from rubbing against the pillow, the results may not be applicable to other brain areas. Thus, optical topography would be necessary to compare hemodynamic signals between different cortical areas. Also, while we have presented and discussed here the general behavior of LFOs and VLFOs in different sleep stages, a more detailed analysis of specific frequency bands in the 0-150 mHz range would provide a more quantitative reference for future NIRS sleep studies.
We used a 4-cm NIRS channel to represent cortical hemodynamics. This carries the risk of attributing hemodynamic changes in the scalp to cortical circulation, especially since the 1and 4-cm signals resemble each other to some extent. For example, the time domain behavior of the 4-cm NIRS signals in all sleep stage transitions could be explained by the influence of systemic hemodynamics, as was demonstrated by PCA with the 1cm signal and HR as physiological noise models. However, this does not show that there are no cortical hemodynamic changes associated with the transitions, only that any such changes must correlate with systemic hemodynamic changes and cannot be isolated with PCA or any other method that relies on the systemic signals to represent noise. Moreover, since systemic hemodynamic changes reflect the behavior of the autonomic nervous system, they also reflect neurophysiological changes related to sleep. Since the influence of autonomic nervous system activity on cerebral hemodynamics is a relatively unknown subject [58], it is difficult to determine in what proportions the 4-cm signals represent contribution from surface tissue, cortical hemodynamic changes due to cortical activation, and cortical hemodynamic changes arising from systemic influences.
It is generally acknowledged that the baselines of hemodynamic parameters such as HR and BP change during the course of the night [46,59]. In our data, instrumental drift prevented reliable tracking of D[HbO 2 ] and D[HbR] baselines over several hours, and we did not compare VLFO and LFO power levels between different parts of the night. Investigating the overnight behavior of the parameters discussed in this study might also provide new insights into cerebral hemodynamics during sleep.

Conclusions
This is the first effort to compare spontaneous hemodynamic VLFOs and LFOs between different sleep stages, and to present a comprehensive account of hemodynamic changes in sleep stage transitions to and from SWS. The results show that SWS is associated with a significantly lower level of slow spontaneous hemodynamic activity than LS, REM, and W. This may be explained by the differences in neuronal activation patterns between sleep stages as well as a reduction in autonomic nervous system activity. Also, transitions between sleep stages are asymmetric, so that the SWSRLS and LSRREM transitions corresponding to increasing complexity of neuronal activity are involved with more dramatic hemodynamic changes than the gradually occurring opposite transitions. This may be partly explained by increased sympathetic activity related to the transition or the new sleep stage. These observations indicate that monitoring spontaneous cerebral hemodynamic oscillations with NIRS provides novel information on the neurophysiological characteristics of sleep. They also encourage examining the interaction between autonomic and cortical hemodynamics in more detail. Figure S1 Time course averages around sleep transitions after PCA filtering. The first row shows the original 4-cm transition averages, the second row the 4-cm averages after removing the component corresponding to the 1-cm averages, and the third row the 4-cm averages after removing components corresponding to HR and the 1-cm averages. LFOs have been removed from the signals for graphical purposes. Note that the PCA filtering removes all visible changes from the 4-cm averages, suggesting that any cortical hemodynamic changes either correlate with systemic hemodynamics or are too small to be detected. In some cases, the PCA-filtered averages of the second row still visually resemble the 1-cm averages. This is most likely because the identification and removal of the 1-cm component was based on a simple statistical criterion instead of visual inspection or a complex physiological or anatomical model.