There is growing interest for assessments of the activity of the autonomic nervous system (ANS) both in research and in applied settings. Few methods exist to measure activity of certain branches of the ANS, such as microneurography to assess muscle sympathetic nerve activity or determination of noradrenalin in the blood to indicate spill-over of the sympathetic nervous system (Grassi and Esler, 1999). HRV is a widely used indirect marker of the cardiac autonomic nervous system activity (CANA) and finds applications in different fields such as psychology (Laborde et al., 2017), sport science (Aubert et al., 2003; Buchheit et al., 2004; Bellenger et al., 2016; Plews et al., 2016) and medicine, namely prenatal diagnostics (Fairchild, 2013), sleep studies (Tobaldini et al., 2013; Dodds et al., 2017), and mortality risk assessment in diseased patients (Huikuri, 1995; Malik et al., 2016; Zhou et al., 2016). In contrast to the other methods, HRV offers a non-invasive and simple method to approximate the CANA.

A fundamental constraint of HRV measurements is the rather poor reproducibility that can be improved when measurements are highly standardized. Standardized measurements are usually performed during 2–5 min in supine position after resting for at least 10 min at the same time of the day (Camm, 1996). However, this time requirement may be problematic in situations where regular (daily) measurements are needed, such as monitoring fatigue and training adaptations in athletes (Plews et al., 2014). Further, resting HRV measurements may also be problematic in young children who cannot stay motionless and relaxed on demand.

Recently, studies have analyzed HRV during sleep to investigate effects of acute stress (Hall et al., 2004; Hynynen et al., 2011), chronic stress (Hynynen et al., 2006), physical activity (Herzig et al., 2017), or diseases (Roumelioti et al., 2010; Amra et al., 2017) on the CANA. Indeed, sleep could provide a highly standardized condition to time-efficiently measure HRV. However, sleep architecture has been shown to affect the CANA in various ways (Somers et al., 1993; Silvani and Dampney, 2013), and also HRV parameters have been shown to vary between different sleep stages and as a function of circadian phase (Trinder et al., 2001; Busek et al., 2005; Vandewalle et al., 2007; Boudreau et al., 2013). Compared to wakefulness, sleep is characterized by a generalized cardiovascular deactivation and a resetting of baroreflex sensitivity (Silvani and Dampney, 2013). Rapid-eye movement (REM) sleep exhibits relatively high muscle sympathetic nerve activity, high lumbar sympathetic nerve activity, and low muscle tone (Silvani and Dampney, 2013). During REM sleep, levels of sympathetic nerve activity above wakefulness values have been recorded when bursts in sympathetic nerve activity induce blood pressure surges and sudden increases in heart rate (Somers et al., 1993). During non-rapid-eye movement (NREM) sleep, low sympathetic activity (Somers et al., 1993; Silvani and Dampney, 2013), reduced cardiac output and lower blood pressure (Somers et al., 1993) have been observed. In sleep stage 2, the occurrence of arousal stimuli accompanied by bursts of sympathetic nerve activity induce a transient increase in blood pressure and heart rate (Trinder et al., 2003). The amplitude and frequency of sympathetic-bursts and blood pressure surges are lower in slow-wave sleep (SWS, N3) than in stage 2 (Somers et al., 1993). SWS is a standardized state with reduced blood pressure variability (Silvani, 2008), constant autonomic activity and a regular breathing frequency, undisturbed by emotional stimuli (Murali et al., 2003). Hence, the absence of sympathetic activity bursts and stationarity of heart rate during SWS offers a highly standardized condition for HRV assessment (Brandenberger et al., 2005). While measurements during SWS are appealing, their validity and usefulness have only sparsely been investigated. In particular, the reliability of HRV measurements has, to our knowledge, never been investigated during specific sleep stages.

Determination of SWS usually requires PSG recordings. However, previous studies have shown close correspondence of HRV parameters and cortical activity (Charloux et al., 1998; Otzenberger et al., 1998; Brandenberger et al., 2001; Dumont et al., 2004). Further, specific HRV characteristics in SWS have been observed, with stationary heart rate and uncorrelated consecutive R-R intervals (Herzig et al., 2016). Recent studies have attempted to classify sleep stages using ECG parameters only (Ebrahimi et al., 2013; Long et al., 2014, 2017; Fonseca et al., 2015; Yoon H. et al., 2017; Yoon H. N. et al., 2017). Moreover, other studies have used a simple method to identify a segment within SWS and analyze HRV (Al Haddad et al., 2009; Herzig et al., 2016, 2017). However, the reliability of these latter methods, relying solely on HRV parameters to select a segment located in a SWS phase, without necessitating PSG recordings, have never been investigated. This method would allow simple HRV measurements in a standardized state that is least influenced by environmental factors.

The aims of this study were to assess the reproducibility of HRV parameters during the different sleep stages by quantifying the variances of HRV parameters between and within subjects, as well as between and within nights of individuals. Furthermore, using the information of PSG sleep staging, we verified the reliability of an algorithm based solely on HRV data to identify a single 5-min SWS segment within each night.

In the present study aiming at characterizing the reproducibility of HRV variables in the different sleep stages, we found better reproducibility (between-subject/within subject ICC and SDs) for HF power, RMSSD, and heart rate during SWS compared to the other sleep stages. In all sleep stages, variances between subjects generally accounted for the highest percentage of total variance, followed by between-segment variance (within same nights) and between nights within each subject. Between-subject/within subject ICCs were excellent for heart rate in all sleep stages. They were also excellent for HF power and RMSSD and good for SDNN in SWS. Fair ICCs were found for LF and LF/HF ratio in all sleep stages. When a single 5-min segment was located within SWS identified by an algorithm based on HRV data only, 87% were correctly placed within SWS according to PSG.

The variances observed during SWS were comparable to the variances reported by Schroeder et al. (2004) for HRV measurements in wake supine position, using a repeated measure study design similar to ours. The lowest residual variance (within-night variance), reflecting the best reproducibility, was found for heart rate and HF power in SWS, followed by stage 2, and highest residual variance and poorest reproducibility in REM sleep. These results are in accordance with the low cardiovascular variability reported during SWS (Franzini, 2000). Thus, while heart rate, HF power and RMSSD in SWS are highly reproducible under comparable conditions, it remains open whether it is also a stage suited to detect acute or chronic stress. Acute psychological stress was found to be reflected by lower HF power in stress situations both during NREM and REM sleep, and of note, with the difference increasing over the course of the night (Hall et al., 2004). Further research is required to assess how well HRV parameters measured in the different sleep stages may reflect the effects of different stressors.

This is, to the best of our knowledge, the first study to quantify variance components of HRV parameters as well as reproducibility in different sleep stages. Reproducibility of HRV measurements in an awake state between different days has been extensively studied (Schroeder et al., 2004; Pinna et al., 2007; Sacre et al., 2012; Schafer et al., 2015; Silva et al., 2017). During sleep, reproducibility of mean heart rate of a whole night's sleep has been investigated (Waldeck and Lambert, 2003). In these studies, ICC, a measure for relative reproducibility, for HRV measurements between different days ranged from 0.64 to 0.91 when measured in a wake state in supine position. Similar values of reproducibility were found for different HRV parameters in the studies cited above. In our study, ICCs were best for HF power (0.84), RMSSD (0.84), and SDNN (0.73) in SWS, and poorer for the two other sleep stages. For heart rate ICCs were between 0.75 and 0.77 in all sleep stages. For LF power and LF/HF power ratio ICCs were below 0.55 in all sleep stages. The low ICC for LF power and LF/HF ratio during SWS may have been a consequence of the generally very low values of LF power in all subjects, reflecting low sympathetic activity (Somers et al., 1993; Silvani and Dampney, 2013) during deep sleep. Greater heterogeneity of the population regarding the parameter of interest automatically leads to greater ICCs compared to more homogeneous populations. The generally poorer ICCs in stage 2 and REM sleep for all analyzed HRV parameters may be explained by the higher frequency and higher amplitude of sympathetic surges and therefore the lower stationarity of HRV parameters in these two stages.

Based on the results of our study, it seems advisable to assess HF power and/or RMSSD during SWS rather than another sleep stage due to a smaller between segment variance. Given that the variance was greater between segments than between nights (and this was not due to a trend over the course of the night), it seems important to average over several segments within one night. Our findings of a low between-subject/within-subject ICC for LF power and LF/HF ratio in all sleep stages questions the suitability of sleep in general for the assessment of these parameters, which are often used as marker of sympathetic activity.

In previous studies assessing the effects of sleep stages on heart rate and HRV, the highest values of heart rate were observed during REM sleep and lower heart rates during stage 2 and SWS with no significant difference between the latter two stages (Trinder et al., 2001; Boudreau et al., 2013). Similar effects of sleep stages on heart rate were observed in our study. HF power was comparable in SWS, stage 2 and REM sleep which was also the case in previous studies (Busek et al., 2005; Boudreau et al., 2013) while others have observed a reduction in HF power in REM sleep compared to the other sleep stages (Trinder et al., 2001). In our study, LF power was lowest during SWS and highest during REM sleep. This is in accordance with results from other studies and indicates a clear sleep stage dependence of LF power (Trinder et al., 2001; Busek et al., 2005; Cabiddu et al., 2012; Boudreau et al., 2013). Reported differences between sleep stages with regard to absolute LF power (up to 140%) in our and previous studies were much bigger than with regard to absolute HF power (<25%). The large differences in LF power between different sleep stages resulted in a concomitantly large difference in LF/HF ratio, being lowest in SWS, followed by stage 2 and being highest during REM sleep. These findings are also in accordance with previous studies (Scholz et al., 1997; Elsenbruch et al., 1999; Ferri et al., 2000).

Many studies reported a rapid decrease in heart rate concomitant with the wake-sleep transition (Shinar et al., 2006) and a more gradual decrease thereafter throughout the night (Burgess et al., 1999). This has been clearly shown when the different sleep stages were analyzed separately (Cajochen et al., 1994; Trinder et al., 2001). When heart rate is analyzed irrespective of the sleep stages, then often an increase in heart rate toward the morning hours was observed. Reasons for this may be the higher abundance of REM sleep in the later part of sleep, the circadian increase in cortisol concentration in the blood and/or the increase in body temperature toward the morning (Burgess et al., 1997). Further, heart rate has been found to steeply increase at the end of each SWS phase, when an arousal and the concomitant blood pressure surge initiates a transition into REM sleep (Bonnet and Arand, 1997). These autonomic processes precede the central activities according to which the sleep stages are classified. Unexpectedly, in our study, a mean increase in heart rate of 3.2 ± 0.4 (SD) bpm was observed in the first SWS phase. This was followed by a mean decrease in heart rate of (2.9 ± 0.5 bpm) over the rest of the night. The initial increase in heart rate in the first SWS phase has not been reported previously, however, some data seems to be in agreement with our finding (Cajochen et al., 1994; Burgess et al., 1999; Brandenberger et al., 2005). This acceleration of heart rate in the first SWS phase was not accompanied by a concomitant decrease in HF power as would be expected when caused by a decreased vagal activity or by an increased sympathetic activity. We suspect factors other than the CANA to be responsible for this discrepancy.

The accuracy of our algorithm for the detection of a segment in SWS by relying solely on HRV data was satisfactory with most of the segments (87%) being accurately placed within a phase of SWS (as verified by PSG). Different HRV based algorithms to perform sleep staging have been proposed and compared to sleep staging by PSG. Most algorithms have used HRV frequency domain parameters (Shinar et al., 2001; Mendez et al., 2009). Other studies have used a combination of different HRV parameters (time domain, frequency domain, non-linear parameters) with the aim to improve the precision of their classification algorithm (Xiao et al., 2013; Fonseca et al., 2015; Yoon H. et al., 2017; Yoon H. N. et al., 2017). Compared to the traditional sleep staging by PSG, an accuracy of 69% in the classification of sleep stages was achieved in the study by Fonseca et al. (2015). In an attempt to classify sleep into wake-REM-NREM sleep, an accuracy of 79% has been reported using a combination of frequency domain parameters (Mendez et al., 2009). In one early study (Shinar et al., 2001), the use of HRV (LF/HF ratio) to identify SWS resulted in 80% correct identification of SWS. Recently, using HRV parameters besides respiratory signals measured by respiratory inductance plethysmography, an accuracy of 89% was reported for the detection of slow wave sleep (Long et al., 2017). Similarly, using a combination of time domain, frequency domain and non-linear parameters resulted in an Cohens kappa of 0.56 for the detection of SWS (Yoon H. N. et al., 2017). Combining HRV frequency domain parameters with actigraphy resulted in an accuracy of 74% (Muzet et al., 2016), comparable to the methods based on HRV only. In contrast to these studies, our aim was not to perform an actual sleep staging but to verify that a segment identified by HRV analysis was located within SWS This explains the rather high accuracy of our algorithm, based on rRR values, a simple HRV measure. In one of our previous studies in elite athletes (Herzig et al., 2016), the same algorithm resulted in a yield of 75% of segments placed in SWS according to non-conventional PSG sleep staging with a mobile 3-lead EEG system.

A limitation of the present study was the absence of a direct measurement of the CANA. However, the aim of this study was to quantify the reproducibility of HRV parameters without making direct inferences on the CANA. Further, we have only quantified reproducibility under standardized conditions, it will be the task of future studies to assess the sensitivity of HRV parameters in different sleep stages to detect disturbances, such as psychological or physical acute or chronic stress. While cortical activity largely follows autonomic activity (Jurysta et al., 2003) some dissociation may occur. Further, data of the present study was collected for a different study assessing a treatment effect. However, the small between nights variance support the absence of a treatment or night order effect. Last but not least, the present study was conducted in healthy subjects, and findings cannot be extrapolated to diseased populations. Strengths of the present study are the statistical analysis of the nested design separating variances due to different factors. This is the first study to assess the reproducibility of HRV parameters during sleep and to compare reproducibility of these parameters between the different sleep stages. The simple algorithm to identify SWS used in the present study will allow wide usage of this efficient method to measure HRV parameters in SWS.

In conclusion, SWS offers a stationary and reproducible phase for determination of heart rate, HF power and RMSSD, and can satisfactorily be detected based on R-R intervals derived from ECG recordings, without the need of full PSG. LF power and LF/HF ratio had fair to poor reproducibility in all sleep stages. Between segment variance was high compared to between night variance for all HRV parameters except heart rate, indicating that averaging over segments within each night is advisable.

XO, RR, and PA designed the experiment and conducted the data collection. DH, PE, MW, XO, and PA were involved in data analysis. DH and PE composed the manuscript. All authors reviewed and approved the manuscript.