INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

BEAT-TO-BEAT FLUCTUATIONS of R-R intervals (RRI), also known as heart rate variability (HRV), in humans contain oscillations for periods ranging from seconds to hours (20). In the frequency domain, the power spectrum of HRV has been categorized into 1) high-frequency (HF; >0.15 Hz), 2) low-frequency (LF; 0.04-0.15 Hz), 3) very low-frequency (VLF; 0.0033-0.04 Hz), and 4) ultra low-frequency (ULF; <0.0033 Hz) components (20). Although the use of HF and LF has been proposed to estimate cardiac autonomic responsiveness (10, 16), the physiological basis of slower fluctuations in HRV (i.e., the VLF and ULF components) is still unclear (20).

In the VLF and ULF bands, the power spectrum of HRV is known to have 1/f -type scaling (3, 9, 17), and both the power in the VLF and ULF bands (2, 3) as well as the slope () of the scaling (3) have been reported to be good predictors of survival for patients after myocardial infarction. Because of this potential clinical significance, the origin(s) of long-period oscillation in HRV has recently been studied (1, 15). For example, Bernardi et al. (1) analyzed 1-h HRV in healthy subjects and showed that powers in the VLF and ULF bands (<0.03 Hz in their case) were higher in exercising than in resting individuals. Roach et al. (15) analyzed HRV obtained from 24-h Holter electrocardiograms (ECG) of healthy subjects and reported that there was no evidence of any persistent oscillations within the ULF band, suggesting that the power derived from transient changes in activity levels was associated with a sleep-wake cycle. However, these studies were limited in that 1) the activity levels of subjects were not directly measured, and 2) a <24-h period of observation might not be sufficiently long to quantitatively evaluate periodicities in the circadian range.

Accordingly, in the present study, we developed a very long-term (~7 days) ambulatory monitoring system for HRV and body movements (BM), and with this device we investigated the mechanism(s) responsible for the long-period oscillation in human HRV.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Measuring device. A very long-term (7-90 days) ambulatory monitoring device (LAMD) was constructed for beat-to-beat HRV and continuous BM measurements. The LAMD consisted of an amplifier for ECG signals, two shock sensors with amplifiers measuring trunk acceleration (in the vertical and horizontal axes), an 8-bit central processing unit (CPU) with 4-MHz frequency, an 8 MB electrically erasable programmable read-only memory (EEPROM), an analog-to-digital converter sampling at 1,000 Hz, an 8-bit parallel interface for data transfer, and a direct current power supply from two commercial dry cells. In the CPU, the analog output of the ECG amplifier was band-pass filtered to yield trigger sources corresponding to QRS spikes. The resultant RRI was stored sequentially in the memory. The signals from shock sensors were recorded after being full-wave rectified and integrated over 1 s. This lightweight (200 g), small (120 × 65 × 22 mm) device operated by using several dry cells for ~3 mo of the EEPROM capacity. The LAMD also had an event button whereby time stamps could be recorded by the subjects.

Experimental protocols. Continuous, ~7-day measurements of RRI and BM for five healthy subjects who were nonsmokers (3 males and 2 females, 19-29 yr of age) were made while subjects spent their normal daily lives without vigorous exercise or alcohol consumption. Each subject gave informed consent to participate in this institutionally approved study after the test protocol had been fully described.

Data analyses. The LAMD continued to collect RRI even when the ECG signal was absent (e.g., while bathing) or when the electrodes occasionally failed. During such a "data gap," the time period was first divided into the values calculated as an average for RRI 1 min before and after the gap. The number of beats inserted in this manner was determined so that the total recording duration was not altered. Any other abnormal RRI, caused either by body movements or occasional extrasystoles, were corrected by either omitting beats (for those <300 ms) or inserting beats (for those double or triple the length of the preceding intervals).

Statistical analyses. The mean value for power spectral densities of HRV at the minimal frequency (P_min; with a period of 45.5 h) was compared with that at the circadian frequency (P_circ; with a period of 22.8 h) by paired t-test. To test whether the observed ² was above noise level, 20 surrogate data sets for each subject in which only BM was randomly shuffled (sorted) were generated. The cross-spectral analysis was then performed on these data sets 20 times. Because this procedure would eliminate couplings between BM and HRV, the ² was expected to be low if it had substantial values for the actual data. Thereafter, the ² from the actual data was compared with those from the dummy data sets that produce 95% confidence intervals, i.e., null-effect ranges.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The LAMD could successfully record continuous RRI and BM for as long as 7 days (Fig. 1A). In this example, there was a clear circadian rhythmicity in both HRV and BM; RRI was higher and BM was lower during sleeping, and vice versa during the day. Four of five subjects demonstrated a clear circadian rhythmicity. When power spectra were calculated for HRV and BM (Fig. 1B), the maximal values were observed near a log frequency of 5, which corresponded to almost 24 h. In the frequency bands higher than this (i.e., at so-called ultradian frequencies), both HRV and BM had a 1/f -type scaling, as reported previously in the literature for HRV (3, 9, 17). The mean value of  for HRV was 1.183 (Table 1), which was also in accord with previously reported data (3, 17). The R² of linear regression on the log-log plane was very high (Table 1). With the linear trend elimination, the power densities at the lowest frequency (P_min; 6.10 × 10⁶ Hz) that could be calculated from 2¹⁴ data points were reduced on average by 39% of those at the circadian frequency (P_circ; 1.22 × 10⁵ Hz) for HRV (Table 1) and by 35% for BM, respectively. These findings indicated that there was a tendency for these scaling relationships not to be observed in the frequency bands lower than the circadian rhythm for both HRV and BM. It is notable that the power densities at the lowest frequency were calculated from data containing at least five complete circadian cycles. Without the linear trend elimination, this tendency was even stronger and P_min for HRV was significantly (P < 0.05) smaller than P_circ (Table 1). The  and R² values of the 1/f -type scaling were not affected by the linear trend elimination.

View larger version (38K): [in this window] [in a new window]   Fig. 1.   A: example of very long-term data for R-R interval (RRI; solid line) and body movement (BM; dotted line). B: power spectra of 10-s averaged RRI [i.e., heart rate variability (HRV)] and BM for 4 subjects exhibiting clear circadian rhythmicities. a.u., Arbitrary units.

Despite these broadband characteristics for HRV and BM in the ULF and the VLF bands, ² between BM and HRV was not scaled but had some clear peaks (Fig. 2, A-D, top). The most prominent was found at the circadian range (²  0.9), which was significantly (P < 0.05) higher than the null-effect range for all of the four subjects. For two subjects (Fig. 2, A and B), there were also some moderate but significant (P < 0.05) levels of coherent relationships (²  0.5) between these two signals in the ULF band. The cross-spectral analysis between BM and HRV revealed no substantial coherence in the frequency range at or above VLF (log frequency greater than 2.5 Hz).

View larger version (43K): [in this window] [in a new window]   Fig. 2.   Squared coherence (2) and contour plots of cross-Wigner distribution (cross-WD) between 10-s averaged BM and HRV for 4 subjects (A-D) exhibiting clear circadian rhythmicities. Top: observed 2 (solid line) and means ± SE of 2 values from surrogate data (BM was shuffled randomly; dotted lines). Bottom: contour plots of cross-WD, normalized by maximal values for each individual, are displayed every 7% of maximal value.

Contour plots of the cross-WD, normalized by the maximal W_xy for each individual, of these four subjects revealed clear circadian variations. The time-frequency distribution of covariance between BM and HRV was higher during the day but without any distribution during sleep (Fig. 2, A-D, bottom). In other words, the daytime recordings were characterized by simultaneous changes in BM and HRV. During the day, there was also a temporally alternating pattern, suggesting (ultradian) rhythmicity in the coupling between BM and HRV. For example, when the cross-WD during the first day in a subject who had significant ² values in the ULF band (Fig. 2A) was enlarged, there was a strong rhythmicity of a period approximating 70 min (Fig. 3A). This corresponded to the peak in ² at a log frequency of about 3.6 (Fig. 2A). In contrast, the plots for a subject who had no significant ² in the ULF band (Fig. 2D) showed less clear rhythmicity, although there still were time localizations in the distribution (Fig. 3B).

View larger version (31K): [in this window] [in a new window]   Fig. 3.   A: enlarged contour plots (first 1,000 min) of cross-WD between 10-s averaged BM and HRV for subject shown in Fig. 2A. B: same type of plots for subject shown in Fig. 2D.

The results for one subject were remarkably different from the other four subjects. According to his self-recorded behavioral map, this subject spent most of the second day sleeping at home. From the third day onward, he tried to live his usual daily life but had frequent episodes of napping during the day and insomnia during the night. Consequently, a circadian rhythm in the LAMD recording was less clear (Fig. 4A), and the power spectrum of both HRV and BM did not exhibit clear peaks at the circadian frequency (Fig. 4B). In fact, unlike the subjects who exhibited clear circadian rhythmicities, this individual's P_min for HRV was almost the same as his P_circ (Table 1). Furthermore, there was no substantial coherence between BM and HRV (Fig. 4C) and no appreciable covariance between these two signals in the time-frequency domain (Fig. 4D). However, the power spectrum of HRV was still scaled in the ultradian frequency range (Fig. 4B), with a value of  similar to that of the "regular" subjects (Table 1).

View larger version (29K): [in this window] [in a new window]   Fig. 4.   A: very long-term data for RRI (solid line) and BM (dotted line) for a subject whose behavioral rhythms were irregular. B: power spectrum of 10-s averaged RRI and BM. C: 2 (solid line) and means ± SE of 2 values from surrogate data (dotted lines). D: contour plots of cross-WD between 10-s averaged BM and HRV.

The transient changes in the HF components of HRV for the "irregular" subject also exhibited some unique characteristics. Although there was a stable circadian pattern of changes in the HF power as well as mean BM and HR for the single subject shown in Fig. 1B (Fig. 5A) similar to that of the other three subjects, the transient changes in this particular individual (Fig. 5B) were not associated with marked circadian modulations in these variables. In addition, the increases in BM did not necessarily result in quantitatively comparable increases in mean HR and decreases in HF power when day-to-day responses were examined. For example, when data obtained during the second day when he was sleeping at home (~1,500-2,000 min) were compared with those obtained during the first night, a marked decrease in the HF component of HRV was observed without an appreciable increase in BM (Fig. 5B).

View larger version (43K): [in this window] [in a new window]   Fig. 5.   A: example of transient changes in mean BM, HR, and high-frequency component (HF power) of HRV calculated from 600 successive RRI data for subject shown in Fig. 1A. B: same type of plots for subject shown in Fig. 4A. bpm, Beats/min.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Studies on human HRV during the past three decades have demonstrated that beat-to-beat fluctuations of the cardiac intervals contain physiologically relevant information that cannot be obtained by measuring only the mean value of HR (20). However, most of these studies have focused on relatively short-term variations of RRI as estimators of cardiac autonomic responsiveness (10, 16). In the frequency domain, the HF and LF components of HRV have been used almost exclusively in these investigations.

With the development of modern Holter ECG technologies, increased attention was recently paid to the slower oscillatory components of human HRV over a period spanning from minutes to hours. It was shown that, in the VLF and ULF bands, the power spectrum of HRV has a 1/f -type scaling (3, 9, 17) and that both the power in the VLF and ULF bands (2, 3) as well as the slope () of the scaling (3) are good predictors of survival for patients after myocardial infarction. Despite the potential physiological as well as clinical significance of these descriptive findings, the origin(s) of slower fluctuations in HRV are currently unclear (20).

Unlike the short-term analysis of HRV, measurement and analyses of long HRV signals, especially those in the ULF domain, are associated with some intrinsic difficulties. Various factors including circadian, activity-resting, and sleep-wake cycles need to be considered, in addition to the short-term autonomic regulation of HR. In many of the previous studies investigating short-term HRV (e.g., Refs. 10 and 16), these behavioral factors were either controlled or ignored, particularly in laboratory-based experiments. Some authors (1, 15) have considered the possible influences of these factors on long-term HRV, but these investigations were limited because behavioral rhythms were not directly measured and the period of observation (<24 h) was too short to quantitatively evaluate periodicities as long as a circadian cycle.

The LAMD developed in the current study was designed to ameliorate these problems because it provided some behavioral information (from BM), as well as HRV, for a sufficient length of time to analyze circadian rhythmicity. One might be concerned about the validity of the use of the "inbox" accelerometer to capture a wide variety of behavioral patterns of humans, although an integrated output of the accelerometer, which has the same principle of measurement as that used in LAMD, has been shown to be highly correlated with the locomotion speed of subjects (18) and the energy expenditure (12). It is notable, however, that the values for ² between BM and HRV in subjects exhibiting clear circadian rhythmicities were 0.9 at the circadian frequencies (Fig. 2), suggesting that the variations in these variables were linearly cross-correlated. Thus it was expected that, during normal daily life, the integrated accelerogram used in the present study contained quantitative information to the extent that HRV did (the "physiological" dissociation between these two signals is discussed below). Consequently, by using this device, new findings on the relationship between long-term behavioral and HR rhythmicities were demonstrated.

First, a 1/f -type scaling in the power spectrum of human HRV (3, 9, 17) was observed only at ultradian frequencies, not in the infradian frequency range whereby periods were longer than a day (Fig. 1B and Table 1). In all of the four subjects exhibiting clear circadian rhythmicities, P_min of HRV was consistently less than P_circ, and this tendency was not altered by linear trend elimination (Table 1). This finding could not be observed by traditional (<24 h) recordings, because they are not theoretically able to detect the lowest frequency in the ULF band of HRV. These data suggest that a few to several days of recording might be sufficient for studying mechanism(s) underlying human HRV in the ULF band. However, for a definitive conclusion, data obtained over a period >7 days would be required, because in the current study the analyzed subsets of HRV contained only one cycle of oscillation at the lowest frequency (although the results were averaged for 20 subsets). This might inevitably introduce some variability into power spectral values in the infradian range.

Second, there were coherent relationships between BM and HRV at specific ultradian as well as circadian frequencies. This result was partially in agreement with the findings of Bernardi et al. (1) and Roach et al. (15) wherein long-period changes in the activity levels of subjects were accompanied by those in HRV. However, unlike these previous investigations, the present study measured BM and found that the relationships between BM and HRV in the ULF band were seen only at certain specific frequencies. The ² between BM and HRV was very high (Fig. 2, A-D), suggesting that the sleep-wake cycle was a major determinant of circadian rhythmicities in both BM and HRV. There were also moderate levels of coherence at log frequencies of about 4.2 (the period corresponding to ~4.4 h) and at 3.6 (1.1 h) (Fig. 2, A and B).

Because the HRV at the ultradian frequencies contains oscillations with multiple scales and even nonstationarities, a wavelet-based time-frequency analysis was recently used to evaluate the contributions of temporally localized changes in HR to overall measures of HRV (7, 8, 15, 22). In the present study, another version of time-frequency analysis (i.e., the cross-WD) was used to evaluate the temporal localizations of the coherent relationships between BM and HRV because, unlike wavelet analysis, it allowed us to calculate the covariance between these two signals. Consequently, we observed that the relationships between BM and HRV were temporally localized in the daytime and that the rhythmicity in the cross-WD was partially responsible for the moderate levels of ² at the ultradian frequencies (Figs. 2 and 3). Whether these coherent oscillations in BM and HRV were caused by some physiological process is currently unclear. However, considering the finding that 60-100 min periodicities were found in the electroencephalogram-based arousal level (14), they may be related to the activity-resting rhythm of subjects. Diurnal variations in HRV synchronized with food intakes (6) may also account for the rhythmicities.

Finally, it was interesting to note that the 1/f -type scaling in the power spectrum of HRV was observed despite the limited contributions of BM at only specific frequencies (Fig. 2). This was so even when a subject followed an irregular life style (Table 1) without any coherent relationship between BM and HRV (Fig. 4). Our results contrasted with the recent report by Roach et al. (15), which concluded that there was no evidence of any persistent oscillation within the ULF band (this conclusion is also challenged by the results for the cross-WD) and that power came mainly from transient changes in activity levels of the subjects. That is, in the present study, the level of activity (measured by BM) was not necessarily coherent with long-period HR or HRV (HF) modulations (Figs. 4 and 5B). In contrast to the modeling by Roach et al. (15) of very long-term HRV by an on-off response for day-night transition as well as harmonic oscillators in HF and LF bands, our data further indicated that there were temporally localized influences of BM on daytime HRV and that the 1/f -type scaling in the power spectrum of HRV was not simply caused by the on-off transition of physical activity.

As reported previously (4, 5), during normal daily life the mean HR and BM were lower and the HF power of HRV, an index of cardiac vagal activity (10, 16), was higher during sleeping than during waking. However, whether these variations in autonomic outflow (e.g., mean HR and HF power) result secondarily from behavioral changes (i.e., mean BM) and are therefore diurnal in nature has not been clarified. The results of the present study clearly showed that the cardiac autonomic rhythmicity can be dissociated from the behavioral pattern. A similar finding on the dissociation between the magnitude of the autonomic rhythmicity from the behavioral rhythm has been previously reported using a rat model of chronic heart failure (21). Any existence of the temporal dissociation between these two rhythmicities was also confirmed in the present study by the decreased ² (Fig. 4C) and the cross-WD between BM and HRV (Fig. 4D) in the subject with an irregular life style. Thus we conclude that long-term HRV cannot be explained solely by activity levels of the subjects. The 1/f -type scaling in the power spectrum of HRV might possibly reflect an intrinsic autonomic mechanism operating over many hours.

In perspective, the results of the present study confirmed that BM was an important modulator of 1/f -type scaling in human HRV at ultradian frequencies. Considering that the long-period oscillations in HRV have been reported to contain some clinically relevant information for predicting mortality of postinfarction patients (2, 3) and in discriminating cardiac disease patients from healthy individuals (7, 8, 22), the possible effects of patients' behavioral patterns on the long-period oscillations in human HRV should be carefully taken into account for future research.

Also, the present study showed that the covariance of long-term HRV with BM was temporally localized. This suggested the need to use methodologies that enable us to examine variations with multiple scales (7, 8, 15, 22), in addition to the conventional method (3, 9, 17), to analyze a uniform 1/f -type scaling to elucidate the mechanism(s) responsible for the long-period oscillation in human HRV.