Seven healthy volunteers (six men), with average age ±S.E.M. 29.4 ± 2.3 years; height 173.4 ± 2.4 cm; weight 79.4 ± 3.5 kg were studied after they had abstained from caffeine and exercise for 24 h. All subjects were non-smokers, had no evidence of heart disease, and took no medications. This study was approved by the human research committees of the Hunter Holmes McGuire Department of Veterans Affairs Medical Center and the Medical College of Virginia at Virginia Commonwealth University. All subjects gave their written informed consent prior to participating.

Data were recorded simultaneously at 500 Hz with commercial hardware and software (WINDAQ, Dataq Instruments, Akron, OH, United States). An electrocardiogram (ECG), respiration (uncalibrated pnuemobelt), and finger photoplethysmographic arterial pressure (Finapres, Model 2300, Ohmeda, Englewood, CO, United States) were continuously measured. End tidal carbon dioxide (CO₂) concentrations were measured on a breath-by-breath basis at the mouth (Infrared Analyzer, Gambo Engström, Sweden).

Muscle sympathetic nerve activity was recorded directly (Nerve Traffic Analyzer, Model 662C-1, University of Iowa Bioengineering, Iowa City, IA, United States), as described previously (Wallin and Eckberg, 1982). Briefly, multifiber sympathetic efferent traffic from peroneal nerve muscle fascicles was lead off with tungsten microelectrodes with uninsulated tip diameters of about 2 μm. A reference electrode was inserted subcutaneously 1–2 cm from the recording electrode. Both electrodes were connected to a differential preamplifier, and then to an amplifier (total gain of 70,000) where the nerve signal was band-pass filtered (700–2000 Hz), and integrated (time constant 0.1 s) to obtain mean voltage neurograms. Satisfactory recordings of sympathetic nerve activity were defined by pulse synchronous bursts that increased during end-expiratory apnea or Valsalva straining, and did not change during tactile or auditory stimulation.

The subjects were asked to follow three protocols: 1) subjects rested quietly, and breathed spontaneously (uncontrolled frequency) for five minutes; 2) subjects controlled their breathing rate at fifteen breaths per minute (metronome) for five minutes; and 3) subjects voluntarily hyperventilated while breathing 100% oxygen for 2 min, inspired, and held their breath for as long as possible. The protocol was repeated following a saline control, then a complete parasympathetic (atropine sulfate, 0.04 mg/kg, intravenous) blockade, followed by a sympathetic (propranolol, 0.2 mg/kg, intravenous) blockade (hereafter referred to as the double blockade). On a separate day, each subject returned for the saline control and sympathetic blockade only. Sympathetic nerve activity was analyzed with custom programs developed for use with commercial software (WINDAQ, Dataq Instruments, Akron, OH, United States). Bursts with a signal-to-noise ratio > 3:1, and latencies from preceding R waves of about 1.3 s were automatically detected (Fagius et al., 1987). One observer manually over-read results of automated analyses. The measurements were continuous between the stages of the experimental protocol.

A repeated measures scheme was adopted to look for significant changes in the time averaged values and wavelet power of the signals for each treatment of the protocol. Given the small sample size, we could not confidently assume a normal distribution of the means for each treatment. The Friedman analysis of variance by ranks was therefore used as a non-parametric test of significance, given by p < 0.05. Tukey’s honest significant difference criterion was used to account for the multiple comparisons made for the data generated by each analytical method. In the case where the results were compared from both the spontaneous breathing and fixed-frequency breathing data, the two sets of data were included as separate rows in the same multiple comparisons test.

In the case of the phase coherence and conditional mutual information, the Wilcoxon signed-rank test was applied to the paired data of the calculated values and the 95% level from the surrogates. As in the other tests, significant interactions were defined as those with p < 0.05. The sum of the number of significant interactions for the specific treatments was calculated for each subject. The Friedman analysis of variance was then applied to these values to test the null hypothesis that the number of significant interactions is the same against the alternative hypothesis that the number of significant interactions changed between the treatments.

As expected, atropine significantly reduced the mean R-R interval and increased mean systolic and diastolic blood pressure values (Table 1). Total power in the R-R interval, systolic pressure and diastolic pressure signals also appeared to be reduced. In contrast, the increase in R-R interval after propranolol was not found to be significant but the mean breathing rate did lower significantly when compared to saline. It can also be seen that the values for the double atropine-propranolol blockade follow those for atropine more closely than the values for propranolol.

The durations of the measurements within the apnea data were shorter than the fixed 5-min intervals (see Supplementary Material S1), which meant the power in the 0.021–0.052 Hz range could not be accurately determined. Additionally, the shortness of the apnea data meant that the phase coherence and conditional mutual information could not be calculated.

The power across all frequency intervals in the R-R interval variability can be seen to strongly diminish after atropine (Figure 1A). Most importantly though, this is even true for the 0.145–0.6 Hz “respiratory frequency” interval during apnea. The corresponding oscillations that occur during apnea and are abolished by atropine are shown in the time domain in Figure 2.

The graphs of the normalised power (Figure 1B) also show that there is a shift in spectral distributions from the 0.145–0.6 Hz and 0.052–0.145 Hz intervals to the 0.021–0.052 Hz interval after atropine. The tiny size of the fluctuations after atropine might make this result seem negligible but it is still statistically significant. In all other cases most of the power is in the 0.145–0.6 Hz and 0.052–0.145 Hz intervals while after atropine most of the power is in the 0.021–0.052 Hz interval. This indicates that atropine had a greater effect in reducing the power across the higher-frequency intervals compared to the 0.021–0.052 Hz interval.

The same analyses were also performed using the frequency intervals specified by the Task Force on heart rate variability (Task Force of the European Society of Cardiology the North American Society of Pacing Electrophysiology, 1996). The results provided in the Supplementary Material S1 show the same effects described above, with the main difference being the significance over the frequency intervals of the normalied power. However, since a change in the normalised power of one interval necessarily indicates a shift of this power to one or more of the other intervals, the fact that the significance over specific intervals is different does not change the interpretation.

To simplify the interpretation, a summary of the results from the coupling and coherence analysis have been combined in Figure 3. While most of the significant interactions were common in the two saline measurements, there were also some differences. A significant interaction in the combined results was therefore defined as when the Wilcoxon signed-rank test gave p < 0.05 in one of the saline controls and at least p < 0.1 in the other control measurement.

One consideration to note when reviewing these results is that if there is coherence between two pairs of signals A–B and B–C then coherence is not necessarily found in A–C as the coherence can be generated by different fluctuations in the same frequency interval.

The results of the statistical hypothesis tests showed a significant decrease in the number of interactions after both atropine and the double blockade when compared with the saline measurements. No significant difference was seen in the number of interactions after propranolol or when comparing the spontaneous and fixed-frequency breathing measurements (Figure 4) for the same treatments.

For spontaneous breathing, the interactions from the saline controls in the 0.145–0.6 Hz interval include strong bidirectional couplings and coherence between all of the signals. This adds support to the findings found in previous studies (Porta et al., 2013). For the 0.052–0.145 Hz interval the interactions do not appear to be as strong, with only one having significant bidirectional coupling and coherence.

In contrast to the results for the diastolic pressure, the systolic pressure signal remains generally coherent with the R-R intervals in the 0.145–0.6 Hz range after atropine. These results are shown in more detail in the Supplementary Material S1, where it can be seen that the coherence in the 0.052–0.145 Hz interval is slightly less significant after propranolol, while atropine removes all significant coherence across the frequency interval.

The effect of propranolol appears to be more subtle. None of the interactions in the 0.145–0.6 Hz interval are removed, although in the 0.052–0.145 Hz interval it does remove some of the interactions with the systolic pressure. In Figure 3 there appears to be some effect of propranolol when combined with atropine as there are fewer interactions after the double blockade when compared with atropine. However, this was not found to be significant in the statistical hypothesis test.

The responses were also different in spontaneous and fixed-frequency breathing, where in fixed-frequency breathing some of the interactions between respiration and the other signals were lost. However, the difference in the total number of interactions was not found to be significant.

A significant difference in the phase shift was found in the systolic-diastolic pressure coherence in both the 0.052–0.145 Hz and 0.145–0.6 Hz intervals (Figure 5). For both of the saline controls the oscillations at these frequencies are still coherent, meaning changes in phase and frequency are the same, but not in-phase. The phase shift is larger for higher frequencies, suggesting that there is in fact a common time-lag across the frequency intervals. The positive value of the shift indicates that the oscillations in the diastolic pressure are leading the ones in the systolic. In contrast, after atropine the shift is removed in the 0.052–0.145 Hz interval, meaning the oscillations are in phase and rise and fall together. In the 0.145–0.6 Hz interval the phase shift actually becomes slightly negative, meaning the respiratory oscillation in the systolic pressure is leading the oscillation in the diatolic pressure. The effect can also be seen in the time domain time as shown in Figure 6. In these plots the fluctations in the diastolic pressure can be seen to precede those in the sytsolic pressure after saline. However, after atropine the fluctuations in the systolic pressure are in phase with or slightly precede those in the diastolic pressure.

In addition, a significant phase shift was also observed in the 0.145–0.6 Hz interval for the phase coherence between respiration and systolic pressure (see Supplementary Material S1), indicating that after atropine the respiratory oscillation in the systolic pressure was falling further behind the actual respiration. In contrast, no positive phase shift in the phase coherence between respiration and diastolic pressure was observed after atropine. However, this was likely due to the fact that the coherence of the respiratory oscillation in the diastolic pressure signals was not found to be significant.

Although not significant, it can be seen in Figure 5 that the median phase shift for propranolol appears larger than the phase shift observed in the saline controls. It is therefore possible that propranolol had the opposite effect to atropine, creating a positive shift in phase as opposed to a negative one.

Previous studies have proposed that the mechanism behind the generation of respiratory sinus arrhythmia is purely reflex-based (Piepoli et al., 1997; Karemaker, 2009). However, this theory is inconsistent with the results found for R-R interval variability during apnea. The hypothesis of the previous studies implies that there should be no respiratory-frequency oscillations in the R-R interval variability due to the lack of baro-afferent activity generated by the breathing cycle. The results of saline measurements during apnea show clear respiratory-frequency oscillations, which provides opposing evidence for this hypothesis. An alternative hypothesis is that these oscillations are post-synaptic in origin. However, this would mean that there should be no difference in the R-R interval variability at the breathing frequency between the saline control and during either of the pharmacological blockades. Instead, the results presented here show that there is still a huge reduction in R-R interval variability in this frequency range after atropine. The remaining hypothesis is that the oscillations are pre-synaptic in origin but are not generated via the baroreflex.

The persistence of variability in the same frequency range as respiratory sinus arrhythmia during apnea has also been observed previously in dogs (Horner et al., 1995). This suggests that a significant amount of this variability originates from a central mechanism in the brainstem (Berntson et al., 1993; Elstad et al., 2018). This adds to other evidence which shows that the computation of baroreflex sensitivity based on the measurement of respiratory sinus arrhythmia is unreliable (Badra et al., 2001; Carrasco-Sosa et al., 2005; Tiinanen et al., 2008).

Several results point to a significant non-autonomic influence on the variability in the cardiovascular system. Relative to the other frequency intervals, the dynamics of the R-R interval variability in the 0.021–0.052 Hz range is more resistant to the decrease in power caused by a parasympathetic blocker and is also unaffected by a sympathetic blocker. If these oscillations were generated by mainly pre-synaptic mechanisms like those in the other frequency intervals then we would expect a similar decrease in power. This is despite the fact that baroreflex sensitivity fluctuations have been observed to be concentrated in this range rather than at the higher frequency intervals (Eckberg and Kuusela, 2005). However, this gives support to other studies which have shown that the oscillations at this frequency may be influenced by post-synaptic heart rate modulation related to thermoregulation (Fleisher et al., 1996; Sollers et al., 2002). Another study has shown that these temperature-related changes can be induced through visual information (Takakura et al., 2013). Certainly, these very low frequency oscillations are mainly influenced by autonomic-dependent mechanisms that include the parasympathetic system (Taylor et al., 1998), the renin-angiotensin system (Duprez et al., 1995), anti-baroreflex functions and chemoreceptor activity (Ponikowski et al., 1997; Francis et al., 2000). Understanding the exact cause of the oscillations that remain after autonomic blockade is therefore still up for debate. It should also be noted that the stronger effect of atropine on the other frequency intervals should not mean that they reflect proportional changes in parasympathetic activity, which has been shown in previous studies (Kollai and Mizsei, 1990).

The results of the coupling and coherence analysis also provide evidence of non-autonomic interactions. In particular, the persistence of the R-R interval–systolic pressure interaction in the 0.145–0.6 Hz frequency interval after both blocking drugs also suggests that there may be a non-autonomic mechanism involved in its generation. This could be related the mechanical influence of the respiration on the R-R interval via the pressure applied on the sinus node (Swenne, 2013).

Indications of the influence of non-autonomic mechanisms on the cardiovascular system can also been seen in the differences between the systolic and diastolic blood pressure variability. While general blood pressure variability has been observed since the 18th century [see (Billman, 2011) and the references therein], very little work has been done to investigate these differences. However, a change in the phase shift in the systolic-diastolic blood pressure oscillations after atropine was observed by Triedman and Saul in a paper from 1994 (Triedman and Saul, 1994). In that work they measured the coherence between central venous pressure fluctuations and blood pressure variability using frequency domain analysis. They showed that a change in venous pressure was reflected in the systolic pressure after a time lag of 1.55 s and in the diastolic pressure after a time lag of 2.1 s. The phase shifts between venous pressure and the systolic and diastolic pressure variations were shown to equalise across much of the frequency domain after autonomic blockade. Furthermore, they noted that the phase relationship between the instantaneous lung volume and the diastolic pressure variations changed by 90° after autonomic blockade, causing it to match the relation between lung volume and systolic pressure variations.

The explanation that the authors give for the change in phase shift is that during autonomic blockade the blood pressure variations reflect the mechanical effects of intrathoracic pressure on arterial pressure. In contrast, under control conditions a different phase relation is seen between respiration and blood pressure oscillations because the heart-respiration interaction is the predominant modulator.

The results presented here in Figure 5 confirm the findings of Triedman and Saul. The phase shift in the respiration-systolic pressure phase coherence suggest that it is in fact the systolic rather than the diastolic pressure variations which change in phase, although a less-pronounced phase shift in systolic variability was also reported by Triedman and Saul (1994). However, the significant phase shifts below the respiratory frequency range do not align with the assumption that the blood pressure variability after autonomic blockade originates from intrathoracic pressure changes. In Triedman and Saul’s paper, this was explained by the experimental protocol where the respiratory frequency distribution was artificially widened using random-interval breathing. This explanation does not hold for the current results though, where significant phase shifts in the lower frequency interval were seen during fixed-frequency breathing. Rather, this suggests there are other sources of non-autonomic blood pressure variability.

Evidence for the modulation of blood pressure variability by a mechanism other than autonomic effects has been observed in rats, where the variations in systolic blood pressure were attributed to changes in stroke volume (Japundzic et al., 1990). A subsequent study with human subjects found these respiration-synchronous fluctuations to be inversely proportional to those in the R-R interval variability (Toska and Eriksen, 1993). The superposition of these two sources of fluctuations actually reduces the cardiac output variation under control conditions. It was observed that when the R-R interval fluctuations are blocked by atropine this results in an increase in fluctuations of cardiac output and arterial pressure. An alternative explanation for this effect was proposed by Taylor and Eckberg (1996), who observed reduced arterial pressure fluctuations during fixed-rate atrial pacing. They suggested that altered arterial compliance caused by the blocking drugs is the reason for the increased blood pressure variability seen in some of the earlier studies.

Neither of these explanations fully describe the results seen here. Indeed, comparing Figure 1 and the spectral analysis of the systolic blood pressure signals (provided in the Supplementary Material S1) reveals little significant change in the blood pressure variability after atropine despite a huge reduction in the R-R interval variability, with the most striking difference in the 0.145–0.6 Hz interval. This seems to confirm that the blood pressure variability is at least partly-generated by non-autonomic mechanisms. However, given that the two sources of variability indicated by Figure 5 are out of phase by < 90 ° , the principle of wave superposition means that the power should decrease when the heart-respiration interaction is removed. This cannot simply be the effect of intrathoracic pressure changes, the magnitude of which should be correlated with the respiration total power which did not change significantly between the treatments (Table 1). Furthermore, the increased arterial compliance cannot be the only mechanism responsible for maintaining the power of the blood pressure variability since it does decrease in the 0.052–0.145 Hz interval, which is significant during both spontaneous breathing and apnea.

Instead, a possible candidate for the effect seen is a change in the myogenic response. Coherent oscillations related to myogenic activity have been shown to be ubiquitous throughout the cardiovascular system (Stefanovska and Hožič, 2000). Myogenic activity has previously only been found to generate oscillations in the 0.052–0.145 Hz range (Kvernmo et al., 1998), although the adjustment of myogenic activity to the respiration cycle has also been illustrated (Stefanovska et al., 2001). It is therefore possible that the actual effect being observed is an increase in the frequency of these myogenic oscillations, brought about by the change in the blood flow dynamics. If the myogenic oscillations were to move sufficiently close to the 0.145 Hz boundary, as hypothesised to occur under certain conditions in (Stefanovska et al., 2001), then some of their power would contribute to the 0.145–0.6 Hz interval. This would explain the reduction of the power in the 0.052–0.145 Hz interval.

Besides the physiological implications, the phase shift in Figure 5 appears to act as a marker for functioning autonomic regulation of the cardiovascular system. It has already been shown that the phase relationship between heart rate and systolic pressure variations changes with ageing (Milan-Mattos et al., 2018)). However, Triedman and Saul (Triedman and Saul, 1994) only compared the phase shifts in the coherence of systolic and diastolic pressure variability relative to the central venous pressure and instantaneous lung volume. Here we have shown that there is in fact a change in the direct phase relation between the systolic and diastolic pressure variability. It is also interesting to note that, even without autonomic blockade, on timescales of 30 min and longer the systolic and diastolic blood pressure variations are strongly correlated (Mancia et al., 1983). This may therefore help to identify the maximum timescale of the autonomic regulation.

Since atropine is a parasympathetic blocker it does not have a direct effect on the sympathetic nervous system. However, in the sympathetic nerve activity signals a slight decrease in the power of the 0.6–5 Hz interval can be seen after atropine, which is significant for the double blockade during fixed-frequency breathing (Figure 7). An increase in power across all frequency intervals can also be seen during apnea, as reported previously (Morgan et al., 1995). Atropine appears to reduce this effect, while the change in response after propranolol was a reduction in the median powers but with a wide range of responses from the different subjects.

This reveals an important physiological mechanism relating to the role of the parasympathetic system during apnea. The reduction in the power of the sympathetic nerve activity after atropine suggests that the sympathetic system is suppressed when the parasympathetic system is blocked. Such a feedback mechanism has been evidenced previously (Montano et al., 1998). Further study of the effect of atropine on the increase in sympathetic activity during apnea is needed to determine whether this effect is significant compared to any reduction in activity during spontaneous or fixed frequency breathing. It has previously been proposed that the increase in sympathetic activity power during apnea is related to stimulation of the carotid bodies from hypoxia or hypercapnia (Daly et al., 1979; Somers and Abboud, 1993). However, the fact that the subjects in the current study hyperventilated while breathing 100% oxygen for 2 min prior to the apnea measurements would suggest that this is not the cause. Instead, it is possible that a separate parasympathetic-modulated mechanism contributes to this increase in sympathetic activity. This seems consistent with the nonlinearity of the autonomic interactions within the cardiorespiratory system and the role of vagus nerve activity in obstructive sleep apnea (Parhizgar et al., 2011).

The results highlight some important methodological implications. The reason for the apparent loss of significant interactions in the 0.145–0.6 Hz interval during fixed-frequency breathing is that the breathing cycle is an oscillation with a stationary frequency. This means that the dynamics of the Fourier transform surrogate time series in the 0.145–0.6 Hz interval are indistinguishable from the measured data. The only difference that remains is the effect of the interactions on the shape and amplitude of cycles of the tidal volume, which might not be as strong as the effect on the breathing rate in the conditional mutual information and do not influence the wavelet phase coherence. This effect is most surprising for the respiration to R-R interval coupling which characterises the respiratory sinus arrhythmia. This interaction has been explored in-depth in previous studies (Grossman and Taylor, 2007), and is seen in the R-R interval variability power in the 0.145–0.6 Hz interval (Figure 1). Since it is common to use fixed-frequency respiration as a way to limit the adverse effects that breathing has on other physiological parameters, this has important implications for the type of methods that should be used to determine the strength of the respiration - heart coupling. Other methods which are able to directly model and track the changes in the strength of interactions in time may be more suitable (Stankovski et al., 2012). It is also worth noting that experiments using ramped breathing do not face this problem even though the frequency is still externally paced (Stankovski et al., 2013).

The effects of fixed-frequency breathing data can also interact with the effects of the treatment and cause ambiguous results. This can be seen in the respiration-systolic pressure coherence, where it is only for atropine and the double blockade that the coherence is not lost during fixed-frequency breathing (Figure 3). This could perhaps be explained if the amplitude of the fixed-frequency respiration-related fluctations in the systolic pressure were reduced by the effect of atropine. This would increase the relative amplitude of any background fluctuations in the 0.145–0.6 Hz interval. Consequently, the surrogate data would no longer contain a single stationary oscillation and the coherence would appear significant again. However, no significant reduction in amplitude in the 0.145–0.6 Hz interval can be seen in the systolic blood pressure variability. One explanation might again be the mechanical impact that the respiration has on the cardiovascular system. The interaction is dominated by the autonomic baroreflex in healthy individuals, but when the baroreflex is diminished the mechanical effects can start to dominate instead (Swenne, 2013). Breathing also mechanically alters left ventricular stroke volumes (Ruskin et al., 1973). If this effect on the blood pressure occurs over a different time delay then the result is two oscillations around the same frequency superimposed on one another. This would give rise to phase variability in the surrogate data, which makes the underlying coherence appear significant.

It is worth mentioning some of the limitations of the study to provide some more context for the results. The most obvious is the limited number of subjects which means the data is more strongly influenced by variations within individuals. This issue was tackled by using stricter non-parametric statistical hypothesis tests, using surrogate data and comparing against the saline controls at all times. While this may have reduced the likelihood of false positive results, it also means that some non-significant results may become significant if there were more subjects. However, the statistical power of the study was also improved by the large effect sizes seen for the different pharmacological blockades and as such these significant effects were still statistically observable.

The length of the data in time is another limitation of this study. It has already been mentioned that the shortness of the apnea data reduced the range of frequencies that could be observed, in addition to preventing the analysis of coherence and couplings. For the other 5 min long parts of the protocol these types of limitations are still relevant. For example, some of the responses to the blocking drugs may have not been observed because they occurred on timescales longer than 5 min. For low frequency oscillations that were in the observable range there are still fewer cycles of the oscillations, making it more difficult for changes at these frequencies to be statistically significant, causing a bias towards finding significance in results related to high-frequency dynamics. This might explain why the vast majority of the significant results correspond to changes after atropine, which affects the high-frequency vagal activity, while the changes after propranolol may be limited to the low-frequency ranges associated with sympathetic activity.

In the case of the time-frequency analysis, the choice of the frequency intervals can result in bias towards detecting some phenomena over others. To mitigate this, the analysis of the R-R interval variability was performed using two sets of intervals and no anomalous results were found.

A limitation of the observed changes in the power spectral distributions, both absolute and normalised, is that they are found by making comparisons across the whole group. This differs from the phase coherence and conditional mutual information results where surrogate data is used to account for the intra-individual variability. One possible way to achieve this when comparing spectral data would be to use inter-subject surrogate analysis (Toledo et al., 2002). However, this was not possible in the current study as it would require a larger number of subjects in order to capture the full range of variations in the spectral distributions of the control measurements.

Finally, the conditional mutual information-based method for finding the strength of interactions between two raw time series, while very powerful, does not differentiate between the various types of couplings. For example, the significant drop in conditional mutual information for interactions involving respiration during fixed-frequency breathing hints that the interactions observed during spontaneous breathing may be based on the modulation of the breathing rate rather than amplitude-based effects. The development of dynamical Bayesian inference of generalised coupling functions might be useful in this case (Stankovski et al., 2017).

Table 2 gives an overview of the physiological hypotheses that have been tested in the current study, along with the details of comparable studies that tested the same hypotheses. This work has made several contributions to the understanding of cardiorespiratory and cardiovascular interatctions, which we now summarise in the following list of findings:

• The results have revealed that respiratory sinus arrhythmia is centrally mediated and not generated solely by a reflex mechanism of the dynamic pressure changes caused by the breathing cycle.