Origins of Heart Rate Variability
Inducibility and Prevalence of a Discrete, Tachycardic Event
Background—We propose that heart period sequences are linearly organized, like sentences, and that there is a lexicon of recurrent, similarly shaped transient structures like words. Each word (or lexon) has a characteristic physiological basis. One potential lexon is the transient, reversible tachycardia that is induced by exercise initiation under laboratory conditions. We hypothesized that this lexon was inducible and observable on ambulatory ECGs of most or all subjects, was morphologically similar in both induced and detected bursts, and shared a plausible origin in both circumstances.
Methods and Results—Ten healthy subjects (mean age, 36 years) underwent a protocol in which subjects rolled themselves from supine to lateral decubitus positions and back. Transient tachycardias (“bursts”) were seen in 36 of 40 rollovers. Bursts were characterized by an initial monoexponential heart period decay (K=0.39±0.23 s−1), a maximum heart period decrease of 277±109 ms after 10.8±4.5 seconds, and a subsequent return to baseline 23.3±10.8 seconds after roll initiation. The roll-induced bursts were detected with 97% sensitivity and 99% specificity with a search algorithm that incorporated morphological parameters. In 24-hour ambulatory ECGs of 10 healthy subjects (mean age, 38 years; range, 17 to 69 years), 117±59 bursts were detected. Induced and detected bursts were similar in most morphological parameters. Finally, many bursts occurred at night, when rolling over also occurs.
Conclusions—Bursts are inducible, transient tachycardias that occur clinically and constitute a lexon with an understandable physiology.
Measures of heart rate variability are often used as probes into the dynamics of the cardiovascular control system. Several mathematical approaches have been used, including spectral analysis and nonlinear analysis.1 2 3 4 These both assume the presence of long-term, deterministic signals. However, analyses with nonlinear predictability, correlation dimension, and information scaling have shown that heart period sequences do not have the characteristics of signals with long-term determinism.5 6 7 What then might be the origin of heart period variability?
Heart period sequences might be linear concatenations of temporally localized events. Time-frequency decomposition studies8 and the broad spectral width of the 0.10-Hz low-frequency band related to hemodynamic control1 2 3 4 argue that the origins of this component are transient and therefore temporally localized. Similarly, the broad spectral width of the high-frequency band and its inconstant timing and amplitude1 2 3 4 suggest that it too arises from temporally localized events. Finally, the sources of the ultralow-frequency band are localized around the times of going to bed and getting up.9 Thus, the sources of heart period variability seem to be discontinuous and localized in specific segments of the sequences.
Recently, predictability analyses of nonlinear systems have contributed complementary findings. These techniques compare results obtained from analyses of original sequences with those from sets of Fourier phase-randomized surrogate sequences. The surrogate sequences lack the deterministic features of their parent sequences. Kanters et al5 6 and we7 showed that parent heart period sequences from resting subjects were more predictable than their surrogates and that this predictability lasted for only 4 to 30 beats. Predictability is due to the recurrence of similar subsequences, suggesting that the parent sequences contain recurring subsequences.
We propose that heart period sequences are linearly organized, like sentences. There is a lexicon of recurrent, similarly shaped transient structures like words, and each word has a characteristic physiological basis.10 We coin the word “lexon” to refer to the meaningful transient structures in heart period sequences. These structures should fulfill 4 criteria:
The lexon should be reliably inducible in most or all healthy persons.
The lexon should be detectable on ambulatory heart period recordings in most or all healthy persons.
The lexons detected in ambulatory recordings should resemble lexons induced under controlled circumstances in ≥3 features.
The behavioral or physiological factors that are associated with the detected lexon should be related to those factors that induce the lexon under controlled circumstances.
In the present study, we determined whether 1 lexon candidate, the brief tachycardia that occurs at the initiation of exercise,11 12 13 satisfied these criteria. We named these events “bursts” in recognition of their brief tachycardic nature. In preliminary studies, we noted that bursts that were morphologically similar to those reported herein were induced when subjects rolled themselves from supine to lateral decubitus positions and vice versa. Because rollover actions are expected at night while most other physical activities are curtailed, and because bursts are abrupt, large-magnitude structures, we considered them a favorable lexon candidate. Our specific objectives were as follows: (1) to determine whether there was a characteristic reversible tachycardia reliably induced by supine rolling actions in normal subjects; (2) to determine whether similar bursts occurred on ambulatory ECGs in normal subjects; (3) to determine whether induced and detected bursts were similar; and (4) to determine whether the induced and detected bursts plausibly shared a common origin.
There were 2 protocols. In protocol 1, we induced bursts under controlled circumstances to parameterize their morphology. In protocol 2, we used these morphological parameters to locate bursts in ambulatory recordings and to validate their identity with the protocol 1 bursts. All subjects were healthy, consenting volunteers who were not receiving medications.
Protocol 1: Bursts Induced by Rolling Actions
Ten healthy subjects (6 men, 4 women; mean age, 34 years [range, 20 to 49 years]) lay quietly supine for 10 minutes. They then rolled from supine to a right lateral decubitus position for 1 minute, then rolled back to supine for 1 minute, then rolled to a left lateral decubitus position for 1 minute, and finally rolled back to supine for 1 minute. Each subject completed the 4 rolling actions. Subjects were instrumented with ECG leads and the Finapres noninvasive blood pressure monitor (Ohmeda) finger cuff. Outputs were digitized at 200 Hz, delineated, and manually overread in CVSoft (Odessa Software). We used the time of the first R wave of each heart period as the time of that beat’s heart period value. Mean arterial pressure (MAP) values for each beat were calculated as MAP=(systolic pressure/3)+[2×(diastolic pressure/3)]. The resulting sequences were stored in Matlab (The Mathworks).
We needed morphological parameters to construct an algorithm to detect bursts in heart period sequences and to compare the gold-standard induced bursts with the detected bursts. We first reviewed the heart period sequences spanning the rollovers. A burst was considered to have occurred if the sequence of heart period values visually resembled those reported previously.11 12 13 Thirty-six of 40 rollovers induced acceptable bursts. Next, we identified and quantified morphological features shared by all 36 bursts.
The 3 general features of a burst are its duration, magnitude, and shape (Figure 1⇓; Table 1⇓). The duration parameters are the intervals between the initial beat, the minimum heart period beat, and the final beat. Respiratory sinus arrhythmia was evident in the interburst intervals of most subjects (Figure 2⇓), which made delineation of the initial and final beats subject to imprecision. To circumvent this, we also delineated the first beat in a burst to exceed 50% of the tachycardic magnitude and the first beat to exceed 50% of the heart period recovery.
The magnitude parameters are the heart periods at baseline and at the trough of the burst, as well as the difference between them (Figure 1⇑; Table 1⇑). The shape parameters depended on the observation that the initial rapid heart period decay resembled a monoexponential decay function. We fit the heart period decay to a monoexponential curve and obtained a theoretical asymptote heart period, a decay rate constant (K), and a correlation coefficient quantifying the fit of the model to the heart period data. We also analyzed the MAP sequences recorded for each of the bursts and report the magnitude and timing of the hypotension.11 12 13
Our goal was to detect bursts using the fewest parameters. We started with broad inclusion criteria for the simplest of morphological parameters, which ensured initial high sensitivity for the detection algorithm.14 We then iteratively added new parameters, each time finding the optimal value(s) for the newly added parameter. Six cycles of this process identified a set of criteria with high sensitivity and specificity (see Table 1⇑ for definitions of parameters). These included 2 magnitude parameters (ΔHPmag and HPmin−HPasymptote), 2 duration parameters (tmin and t50%rec), a shape parameter (expfit), and a parameter (stability of 15 preceding beats) designed to eliminate false detection of high-magnitude, transient bradycardic events described previously.15 16 The order of the parameters and the optimized values are presented in the Results section. The gold-standard–positive population for the receiver operator characteristic analyses consisted of the 36 bursts induced in protocol 1; the negative controls were all large-magnitude tachycardic events in the protocol 1 heart period sequences. These were defined broadly by the first 2 criteria of the algorithm: a heart period decrease ≥102 ms over a time interval of 3 to 24 seconds.
Protocol 2: Detection of Bursts on Ambulatory ECGs
Ten subjects (6 males, 4 females; mean age, 38 years [range, 17 to 69 years]) underwent ambulatory ECG for 24 hours. The recording system incorporated a synchronization pulse to reduce variability due to tape-speed fluctuations.9 The recordings were analyzed with the Marquette 8000 scanner with version 5.7 of the Marquette Arrhythmia Analysis Program to identify and label each QRS. Entire recordings were analyzed to eliminate cycles in which ventricular beats had abnormal morphological characteristics or were without normal P waves. These beats were replaced by linear interpolation between adjacent normal beats. Unclassified beats were corrected manually and verified. No recordings had ectopy more severe than isolated complexes or atrial couplets. The proportion of ectopic complexes ranged from 0 to 0.17%, and no patients had >7 ectopic beats/h. The corrected sequences of ≈105 beats were then transferred into Matlab.
Derivation of Surrogate Control Sequences
We needed an objective way to distinguish the putative bursts in the original heart period sequences from other transient tachycardias that may have arisen by chance. To do this, we first derived phase-randomized surrogate sequences for each original sequence. These surrogate sequences have the same linear temporal correlation as their original sequences but lack any phase-dependent (or deterministic) structures that may be present in the original sequences. To construct surrogates,5 6 7 we made fast Fourier transformations (FFTs) of the original heart period sequences, replaced the phases of the Fourier coefficients with values drawn randomly from a uniform distribution between 0 and 2π, then applied the inverse FFT.
We used Mann-Whitney nonparametric comparisons of measured variables with idealized parametric distributions to test for normalcy of distributions. Normal distributions are reported as mean±SD. Nonnormal distributions are reported as 25%, median, and 75% quartile values. ANOVAs were performed on grouped variables of normal distributions, and Kruskal-Wallis tests were performed on grouped variables of nonnormal distributions. Unpaired t tests were used to compare normal distributions, and Mann-Whitney U nonparametric tests were used for comparing nonnormal distributions. P<0.05 was deemed to be significant.
Prevalence of Induced Bursts
Thirty-six (90%) of the 40 rolls resulted in bursts that were easily identified by visual inspection. Figure 2⇑ depicts the 4 recordings that were excluded from the analysis, as well as 4 bursts (each from a different subject) that were characteristic of those that were analyzed further. All subjects produced ≥2 of the 4 possible bursts, satisfying the first criterion of generalized inducibility.
Morphological Parameters of Induced Bursts
Table 2⇓ summarizes the values obtained for morphological parameters of the induced bursts. Values for all of the morphological parameters of the induced bursts were normally distributed. Three of the 4 duration parameters, specifically t50%mag (5.9±3.0 seconds), t50%rec (13.3±4.7 seconds), and tburst (23.3±10.8 seconds), were invariant among the subjects. The fourth duration parameter, tmin, was significantly different in only 1 subject (ANOVA post hoc analysis). Thus, measures of the duration of the bursts were generally constant among subjects.
In contrast, the values of the magnitude parameters were more variable. These included HPbaseline (993±189 ms), HPmin (716±104 ms), and ΔHPmag (277±109 ms). However, this variance appeared to be due to differences in the baseline heart periods. Figure 3⇓ shows that both HPmin and ΔHPmag correlated well (P<0.001; r2=0.8) with HPbaseline, which itself was subject dependent. Thus, the variances of the values of the magnitude parameters reflect more the variance in the baseline heart periods rather than intersubject variability in burst magnitude.
The values of the 2 shape parameters showed limited intersubject variability. The initial heart period decrease was well fit by the monoexponential decay model (expfit=0.98±0.02) in all subjects, but the decay constants (0.39±0.23 ms/s), although similar, were significantly subject dependent. Thus, the values of the shape parameters showed only limited intersubject variability.
Finally, all bursts were accompanied by a relative hypotension. MAP dropped 17.8±11.7 mm Hg over 12.1±5.9 seconds.
The optimized burst-detection algorithm incorporated 6 criteria: (1) preceding heart period quiescence, defined as all of the 15 beats preceding the first beat of the burst having heart periods ≥78 ms greater than the trough heart period HPmin; (2) a high-magnitude drop in heart period, defined as a drop in heart period ΔHPmag >102 ms; (3) a bounded time to the trough heart period, defined as tmin between 3.3 and 24 seconds; (4) a bounded time to the point at which heart period recovered 50% toward its baseline value, defined as t50%rec between 8.8 and 31.4 seconds; (5) a high degree of fit of the initial heart period drop to a monoexponential curve, defined as expfit >0.95; and (6) close agreement between the theoretical trough heart period from the monoexponential equation and the measured trough heart period, defined as HPmin−HPasymptote between –63 and 128 ms. The sensitivity of the burst-detection algorithm for the 36 gold-standard inducible bursts was 97%, and the specificity for the negative truth set defined by criteria 1 and 2 was 99%. The positive predictive value for the detection algorithm was 95%.
Prevalence and Morphology of Detected Bursts
Bursts were detected with the algorithm in all 10 of the 24-hour heart period sequences (Table 2⇑). An average of 117±59 detected bursts (range, 3 to 206 bursts) were obtained for the ten 24-hour heart period sequences. All subjects had bursts, satisfying the second criterion of generalized prevalence. Control sequences were derived from the original sequences by phase randomization. There were significantly fewer bursts in the control sequences for each of the original sequences, as well as in all 10 control sequences as a whole compared with the original population.
The magnitudes of the induced and detected bursts were generally similar (Table 3⇓). They had statistically similar t50%mag, tmin, and t50%rec and similar but not identical total durations of tburst. Figure 4⇓ depicts the differences in tburst between induced and detected bursts. The distributions have similar medians and modes, but detected bursts are skewed by a number of very long events. These differences are addressed in the Discussion. In contrast, induced and detected bursts had different magnitudes. Detected bursts started and finished at shorter heart periods and had slightly smaller magnitudes. The initial drops in heart period in both induced and detected bursts were well fit by monoexponential equations. However, the decay constants of the detected bursts were slightly but significantly slower. In general, induced and detected bursts were similar in duration and shape but different in magnitude, satisfying the third criterion of morphological similarity.
Diurnal Distribution of Bursts
The occurrence of bursts with exercise initiation and rolling over suggests that they should be detectable both at night and during daytime hours. Figure 5⇓ plots the number of detected busts per hour over a 24-hour day for the 10 subjects of protocol 2. Bursts occurred throughout the recordings, and their incidence was independent of time of day (P=0.9). Figure 6⇓ depicts detected bursts that occurred during normal sleeping and waking hours. The bursts are easily seen on macroscopic heart period traces and accurately detected by the burst-detection algorithm. During sleeping hours, the interburst sequences were characterized by elevated heart period values and the marked respiratory sinus arrhythmia of supine subjects. Their prevalence throughout the recordings is consistent with their being due to exercise initiation (including rolling over in bed), partially satisfying the fourth criterion of teleological plausibility.
We propose a novel, lexical approach to the study of heart period variability, propose formal criteria for judging the validity of candidate lexical events, and describe the first such lexon. This approach has potential applications in both basic and clinical studies.
The Burst Lexon
The candidate lexon that we chose was a brief, reversible tachycardia previously reported to occur during standing up and during initiation of cycling.11 12 13 It was chosen because of its large size, ease of induction, and apparent stereotypical shape. It is inducible and detectable in most or all subjects; the detected bursts resemble the induced bursts; and the induced and detected bursts share plausibly related provocative factors. Thus, a burst is a discrete, imperfectly reproducible heart period event that is related to known physiological behavior and is easily detectable on ambulatory heart period recordings.
Applications of Lexical Approach
We suggest that heart period sequences be modeled as concatenated assemblages of characteristically scaled and characteristically shaped events, or lexons. The lexical approach has several strengths and applications. It does not assume the existence of any global properties; rather, it focuses on temporally localized structures within the recordings. It overcomes the problem of stationarity that confounds frequency-domain analyses and is based on the premise that there is useful information in the very structures that might cause lack of stationarity. Similarly, it does not require long-term statistical constancy in the sequence.
Second, it allows investigators to form reasonable deductions and hypotheses about the physiology that accompanies heart period changes recorded on ambulatory ECGs. For example, the first 3 to 5 seconds of burst tachycardia may be due to activation of skeletal mechanoreceptors, with the afferent impulses carried by type III and IV nerve fibers.12 17 18 19 Activation of both large and small muscles can cause this phase. The second phase may be a compensatory tachycardia due to unloading of arterial baroreceptors by hypotension.12 Thus, the observation of a burst on a heart period recording suggests the reasonable inference of vagal withdrawal, hypotension, and subsequent increased sympathetic activity mediated by baroreceptors.
Third, the lexical approach facilitates the investigation of temporal physiological transients that might precede clinical events such as arrhythmias and syncope. The identification of specific lexons preceding arrhythmias, coupled with an understanding of their physiology, may prove particular fruitful.
Sources of Variance in Bursts
Although bursts are generally similar, there were differences noted between subjects and between induced and detected bursts. Most of the larger differences were in magnitude measures. There is no published physiological explanation for burst magnitudes, but it is noteworthy that the magnitude variables correlated well with the baseline heart period (Figure 3⇑). We speculate that some of those factors that affect resting heart period also affect the magnitude of the tachycardic response induced by rollover initiation. The burst-detection algorithm was sufficiently robust to detect bursts with both high sensitivity and high specificity in all subjects despite these differences in magnitudes.
The bursts detected on ambulatory recordings were slightly longer than those induced by rolling over. Figure 5⇑ shows that the detected burst distribution has a pronounced tail of lengthy tburst times. Given that bursts can also be induced by sitting up, standing up, starting to walk, and starting to cycle11 12 13 (and D. Roach, MD, unpublished data, 1998), it may be that these long tburst values represent the initiation of exercise and that the heart period values do not fully recover until the exercise is completed. We speculate that the burst may be a universal response to the initiation of activity from relative quiescence.
It might be that the induced and detected bursts are fortuitous coincidences. We used 2 methods to address this problem. First, we constructed phase-scrambled control sequences that contained the statistical and spectral characteristics of the original sequences but lacked deterministic structures. There were significantly fewer bursts in the control sequences. Because phase scrambling preserves some of the properties of the shapes, it is not surprising that some residual subsequences were detected. We also demonstrated the morphological similarity of the induced and detected bursts. Although some of these comparisons might be biased because the same parameters were used in the detection algorithm as in the comparison, many were not. Specifically, ΔHPmag and expfit were 1-sided detection criteria, and tburst, t50%mag, and K were not used as detection criteria but as comparative criteria.
We only examined the response to rolling over. However, bursts occur at the initiation of standing up and cycling11 12 13 and with sitting up and initiation of walking (R. Sheldon, MD, et al, unpublished data, 1998). We have not proven that the bursts detected on Holter recordings are due to exercise initiation or rolling over in bed, and it remains possible that other activities and physiological perturbations cause bursts.
This study was supported by the Medical Research Council of Canada (PG11188), Ottawa, Canada, and the Calgary General Hospital, Calgary, Alberta, Canada (to Dr Sheldon).
- Received November 7, 1998.
- Revision received March 23, 1999.
- Accepted April 9, 1999.
- Copyright © 1999 by American Heart Association
Akselrod SD, Gordon FA, Ubel DC, Shannon A, Barger C, Cohen RJ. Power spectrum analysis of heart rate fluctuation: a quantitative probe of beat-to-beat cardiovascular control. Science. 1981;213:220–222.
Malliani A, Pagani M, Lombardi F, Cerutti S. Cardiovascular neural regulation explored in the frequency domain. Circulation. 1991;84:482–492.
Pagani M, Lombardi F, Suzzetti S, Rimoldi O, Furlan R, Pizzinelli P, Sandrone G, Malfatto G, Dell’Orto S, Piccaluga E, Turiel M, Baselli G, Cerutti S, Malliani A. Power spectral analysis of heart rate for arterial pressure variabilities as a marker of sympatho-vagal interaction in man and conscious dog. Circ Res. 1986;59:178–193.
Saul JP, Berger RD, Albrecht P, Stein SP, Chen MH, Cohen RJ. Transfer function analysis of the circulation: unique insights into cardiovascular regulation. Am J Physiol. 1991;261:H1231–H1245.
Roach DE, Sheldon RS. Information scaling properties of heart rate variability. Am J Physiol. 1998;274:H1970–H1971.
Roach D, Sheldon A, Wilson W, Sheldon R. Temporally localized contributions to measures of large-scale heart rate variability. Am J Physiol. 1998;274:H1465–H1471.
Ewing DJ, Hume L, Campbell IW, Murray A, Neilson JMM, Clarke BF. Autonomic mechanisms in the initial heart rate response to standing. J Appl Physiol. 1980;49:809–814.
Borst C, Wieling W, Brederode JFMV, Hond A, Rijk LGD, Dunning AJ. Mechanisms of initial heart rate response to postural change. Am J Physiol. 1982;243:H678–H681.
Roach DE, Thakore E, Sheldon RS. Large magnitude, transient, bradycardic events in rabbits. Am J Physiol. In press.
Roach DE, Thakore E, Sheldon RS. Autonomic control of rabbit faintlets. Pacing Clin Electrophysiol. 1998;21:927. Abstract.
Shepherd JT, Blomqvist CG, Lind AR, Mitchell JH, Saltin B. Static (isometric) exercise: retrospection and introspection. Circ Res. 1981;48(suppl I):179–188.
Hollander AP, Bouman LN. Cardiac acceleration in man elicited by a muscle-heart reflex. J Appl Physiol. 1975;38:272–278.
Borst C, Hollander AP, Bouman LN. Cardiac acceleration elicited by voluntary muscle contractions of minimal duration. J Appl Physiol. 1972;32:70–77.
Smith S. Reduced sinus arrhythmia in diabetic autonomic neuropathy: diagnostic value of an age-related normal value. BMJ. 1982;285:1599–1601.
O’Brien IAD, O’Hare P, Corrall RJM. Heart rate variability in healthy subjects: effect of age and the derivation of normal ranges for tests of autonomic function. Br Heart J. 1986;55:348–354.