Assessment of pulse rate variability by the method of pulse frequency demodulation
- Junichiro Hayano1Email author,
- Allan Kardec Barros†2,
- Atsunori Kamiya†3,
- Nobuyuki Ohte†4 and
- Fumihiko Yasuma†5
© Hayano et al; licensee BioMed Central Ltd. 2005
Received: 09 May 2005
Accepted: 01 November 2005
Published: 01 November 2005
Due to its easy applicability, pulse wave has been proposed as a surrogate of electrocardiogram (ECG) for the analysis of heart rate variability (HRV). However, its smoother waveform precludes accurate measurement of pulse-to-pulse interval by fiducial-point algorithms. Here we report a pulse frequency demodulation (PFDM) technique as a method for extracting instantaneous pulse rate function directly from pulse wave signal and its usefulness for assessing pulse rate variability (PRV).
Simulated pulse wave signals with known pulse interval functions and actual pulse wave signals obtained from 30 subjects with a trans-dermal pulse wave device were analyzed by PFDM. The results were compared with heart rate and HRV assessed from simultaneously recorded ECG.
Analysis of simulated data revealed that the PFDM faithfully demodulates source interval function with preserving the frequency characteristics of the function, even when the intervals fluctuate rapidly over a wide range and when the signals include fluctuations in pulse height and baseline. Analysis of actual data revealed that individual means of low and high frequency components of PRV showed good agreement with those of HRV (intraclass correlation coefficient, 0.997 and 0.981, respectively).
The PFDM of pulse wave signal provides a reliable assessment of PRV. Given the popularity of pulse wave equipments, PFDM may open new ways to the studies of long-term assessment of cardiovascular variability and dynamics.
Analysis of heart rate variability (HRV) has been standardized with using R-R intervals of electrocardiogram (ECG) as the source signal . The measurement of ECG, however, requires multiple electrode attachments and cable connections, which precludes frequent assessments of HRV in general populations. Beyond the uses as an autonomic functional index [2, 3] or as a prognostic marker of long-term survival [4–6], the applications of HRV have extended into many areas of health sciences and technologies, such as those as markers for assessing the levels of physical and mental demands [7, 8], the degree of fatigue , the depth of relaxation and resting [10, 11] and the comfortableness of living and occupational environments [12, 13]. For these applications, self-applicable devices that allow frequent, preferably everyday measurement in general population may be more useful.
From this aspect, analysis of pulse rate variability (PRV) from pulse wave signal has been studied as a potential surrogate of HRV analysis . In contrast to ECG, pulse wave can be recorded with a single sensor without electrode and indeed, pulse wave equipments are popular and widely used not only in hospital cares but also in health sciences and clinical homecare practices. However, there is also a problem, i.e., smoother waveform of pulse wave precludes accurate measurement of pulse-to-pulse intervals by fiducial-point algorithms such as those used for measuring R-R intervals of ECG.
Here, we report a technique of pulse frequency demodulation (PFDM) as a method for estimating the function of instantaneous pulse rate. Taking advantage of smoother waveform of pulse wave, the PFDM extracts instantaneous pulse rate directly from pulse wave signal as a function of time. The continuous function of pulse rate can be converted into instantaneous pulse interval function, which can be directly used for spectral analysis. In this study, we tested the performance of the PFDM using both simulated data and actual pulse wave signals that were recorded with a wireless trans-dermal photoelectric device. We also compared PRV estimated by the PFDM with HRV measured by R-R intervals of simultaneously recorded ECG.
Principle of PFDM
The core process of PFDM is frequency demodulation based on the method of complex demodulation (CDM) [15–17]. CDM is a non-linear time-domain method for time series analysis, which provides amplitude and frequency of non-stationary/unstable oscillatory signal as a continuous function of time. Principle and computer algorithm of CDM have been published previously . Briefly, CDM extracts the time dependent functions of instantaneous frequency through the following four steps: (1) the spectral region of interest (the frequency range of target oscillation) is shifted to zero frequency by forming a product, throughout the record, of the original signal and a complex sinusoid at a reference frequency (Fr, the center frequency of the spectral region of interest), (2) the resultant complex signal is low-pass filtered so that only frequency components around zero remain, (3) the real and imaginary parts of the low-pass filtered signal are converted to a polar form, yielding the instantaneous phase, as a function of time, of the component identified at or near the Fr, and (4) the time series of frequency of the target oscillation is calculated through adding the first-order derivative of the phase to the Fr, since the slope of the phase versus time curve indicates deviation of instantaneous operative frequency from the Fr.
In PFDM, the CDM is customized for analyzing pulse wave signal. Pulse frequency (instantaneous pulse rate) could change widely from less than 0.5 Hz (30 bpm) to more than 3 Hz (180 bpm). Because CDM extracts frequency of oscillatory components within an assigned spectral region (range of CDM filter), an appropriate frequency range needs to be selected so that it covers the possible range of pulse frequency. This is, however, a trade-off with the requirement that the frequency range of CDM filter needs to be narrow enough to avoid influence of harmonics and subharmonics of the fundamental oscillation (pulse wave).
When frequency of pulse wave decreases to a frequency as low as f Hz, the upper limit of the frequency range should be less than 2f Hz (frequency of the 2nd harmonic of the pulse wave). Likewise, when frequency of pulse wave increases to a frequency as high as f Hz, the lower limit of the frequency range should be greater than f/2 Hz (frequency of the 2nd subharmonic of the pulse wave). If the range for CDM filter is expressed as Fr ± Fc, where Fr is the reference frequency (the center frequency of the spectral region of interest) and Fc is the corner frequency of the low-pass filter, then the necessary condition for avoiding the influence of harmonics and subharmonics is
Fc < Fr/3.
For example, when mean pulse rate is 60 bpm, the widest possible range for CDM filter is 60 ± 20 (40 to 80) bpm; if pulse frequency deviates from this range during analyzing period, CDM would not provide an adequate estimation of pulse rate.
To overcome this problem, we devised the algorithm of PFDM incorporating the following three features: (1) overlapping short-segmentation of data, (2) adaptive determination of the Fr, (3) iterative algorithm for optimizing Fr, and (4) stepwise convergence of Fc toward Fr/3 during the iterative process. Briefly, data are first divided into short segments so that pulse frequency within each segment can be expected to remain within the range that can be covered by a feasible CDM filter. In each segment, the mean pulse frequency of the previous segment is used as the initial Fr value (adaptation). The Fr is further optimized through iteration processes, i.e., mean pulse frequency calculated is iteratively used as the Fr in the next iteration until the mean pulse frequency calculated agrees with the Fr used. Finally, to further guard against the possibility that pulse frequency deviates from the range of CDM filter, an Fc of Fr/2 is used for the initial process and it is converged to Fr/3 in the following iteration processes. The effectiveness of each of these processes is demonstrated in the simulation studies.
Measurement of actual data
For both simulated and actual data analysis, pulse wave data were sampled at 20 Hz. The PFDM was performed with custom-made software written with FORTRAN 95 (Salford Software Ltd, Old Trafford, Manchester, UK). For the PFDM, the data segments had a length of 30 s with overlapping for 10 s at both ends. For the iteration for optimizing Fr, the tolerance for the difference between the mean pulse frequency and Fr was set at 0.001 bpm. Although the instantaneous pulse frequencies were calculated for every sampling point (at 20 Hz), they were averaged over every 500 ms and converted into pulse intervals in order to compare with the source interval functions (for the simulation data) or R-R intervals of ECG (for the actual data).
HRV, by definition, is the beat-to-beat variability of sinus rhythm [1, 18]. Thus, all R-R interval data involving ectopic beat(s) or heart block(s) were excluded from the analysis. The pulse wave data, however, have less information about arrhythmias. An ectopic beat, either supra-ventricular or ventricular, could result in consecutive short pulse intervals, a long pulse interval or even a normal interval depending on the timing and whether it generates a detectable pulse wave or not. Heart blocks may also result in long intervals. To avoid the effects of rhythm disturbances and noises on the analysis of PRV, we excluded all abnormal pulse intervals, which were defined as those deviating 12% or more from the local moving average over the preceding 20 s. The definition of abnormal pulse interval was determined tentatively considering expected range of physiological PRV for the subject population.
The ECG data were analyzed with a Holter scanner (DSC-3100, Nihon Koden, Tokyo, Japan), on which the results of automatic labeling of QRS complexes were reviewed and manually edited for all errors. The ECG were analyzed with a sampling frequency of 125 Hz and, thus, the R-R intervals were measured at a time resolution of 8 ms. The time series data of R-R interval were interpolated along the time axis with a horizontal-step function (R-R interval was considered as constant during each R-R interval) and resampled at 2 Hz.
Fast Fourier transformation with a Hanning window was performed for sequential 5-minute segments of both pulse interval and R-R interval data. A segment was excluded from the analysis, if the ratio of valid data points in the segment was <80%. In the segments analyzed, defected parts of data, if any, were interpolated by the horizontal-step function. After correcting for the loss of variance resulting from the window process, power spectral density was integrated over 0.04–0.15 Hz and 0.15–0.45 Hz for assessing the power of low frequency (LF) and high frequency (HF) components, respectively. The power of these components was converted into mean amplitude ([2 × power]1/2) to reduce the skewness from the normal distribution.
The agreement between PRV and HRV measures was evaluated from two aspects; (1) agreement when these measures are used for assessing intra-individual variations and (2) agreement when they are used for inter-individual comparisons.
To examine the former agreement, minute-to-minute pulse rate and heart rate and spectral components of PRV and HRV for 5-min segments were compared within each subject. The agreement between corresponding values were evaluated with (a) the difference-against-mean plot and the 'limits of agreement' of Bland and Altman method  and (b) intraclass correlation coefficients for 2-way mixed effects analysis of variance with defining segments as the random factor and methods as the fixed factor . The statistical adequacy of the segments as random factor is partly supported by the fact that long-term heart rate fluctuation has the characteristics of random fractal noise [1, 21].
To examine the latter agreement for inter-individual comparisons, the minute-to-minute and 5-min segment values were averaged over the entire recording length for individual subjects. The agreement between the corresponding mean values were evaluated with (a) the Bland and Altman method same as above  and (b) intraclass correlation coefficients for 2-way mixed effects analysis of variance with defining subjects as the random factor and methods as the fixed factor .
Data were presented as mean (SD) or median (range). Type 1 error level was set at probability (p value) of < 0.05.
Simulation studies were performed for evaluating the performance of the PFDM, particularly the effects of (1) Fr adaptation with data segmentation, (2) Fr optimization with iterative algorithm, (3) Fc convergence, and (4) robustness against the fluctuation in pulse height and baseline trend. Also, the frequency characteristics of the PFDM were examined to test if the output from the PFDM is appropriate for frequency domain analyses.
Effects of Fr adaptation and iterative optimizations
Effects of Fc convergence
Robustness against pulse height and baseline fluctuations
Analysis of actual data
Actual data of overnight recordings of pulse wave together with simultaneous ECG were obtained from 30 subjects out of 33. Data were lost due to technical problems in a subject and were excluded due to atrial fibrillation during the entire recording periods in two subjects. The mean ± SD length of data in the 30 subjects was 6.0 ± 0.8 hr. In these data, median ratio (range) of 5-min segments that met the criteria (valid data ≥80%) for spectral analysis was 87 (63–95) %.
The presences of atrial fibrillation in the two excluded subjects were detected by the PFDM of pulse wave. In both subjects, more than 90% of the pulse intervals deviated 12% or more from the local moving average and none of the 5-min segments met the inclusion criteria (valid data ≥80%).
Agreement between pulse rate and heart rate
The analysis of agreement for inter-individual comparisons showed almost perfect agreement between the mean pulse rate and heart rate of individual subjects. The intraclass correlation coefficient was 0.999, the mean difference was 0.00 bpm, and the upper and lower limits of agreement were 0.03 and -0.03 bpm, respectively.
Agreement of spectral components of PRV and HRV
Agreement between PRV and HRV for inter-individual comparisons
PRV Mean (SD) ms
HRV Mean (SD) ms
Mean difference ms
Upper limit of agreement ms
Lower limit of agreement ms
To evaluate the agreement for inter-individual comparisons, mean values of LF and HF amplitude of individual subjects were compared between PRV and HRV. The results showed good agreement for the mean LF amplitude and acceptable agreement for the mean HF amplitude (Table 1).
Influence of ventricular ectopies (VE)
In this study, abnormal pulse intervals caused by ectopies and heart blocks were excluded whenever they deviated 12% or more from the local average. To examine the influence of this process on the assessments of pulse rate and spectral components, the effects of the frequency of VE on the agreement between values calculated from the pulse interval by PFDM and ECG R-R interval were analyzed.
In this study we proposed the PFDM technique as a method for demodulating instantaneous pulse rate from pulse wave signal and demonstrated its usefulness for assessing spectral components of PRV. The simulation studies revealed that (1) the PFDM provides reliable measurement of instantaneous pulse rate even if it fluctuates rapidly and over a wide range, (2) it is robust to the variations of pulse height and baseline trend, and (3) it preserves the frequency characteristics of the source modulator function, a feature necessary for spectral analysis. The analysis of actual pulse wave revealed that (1) minute-to-minute pulse rate assessed by the PFDM agreed perfectly with minute-to-minute heart rate measured by ECG, (2) the amplitudes of LF and HF components of PRV show overnight trends quite similar to those of HRV, and (3) the mean values of LF and HF amplitudes during night show good agreement between PRV and HRV. These observations indicate that the PFDM provides a reliable assessment of pulse rate and PRV and suggest that this technology makes pulse wave a potentially useful source signal for assessing cardiovascular variability and dynamics.
It is noteworthy that the pulse wave was sampled at a frequency of 20 Hz, while the ECG was sampled at 125 Hz. Nevertheless, we observed surprisingly good agreement not only between pulse rate and heart rate but also between the spectral components of PRV and HRV. This indicates that the accuracy of the PFDM analysis of pulse interval is not directly dependent on the time resolution of data. Indeed, the simulation studies revealed that the estimation errors of pulse interval by the PFDM are within ± 2 ms despite the fact that the sampling interval is 50 ms (20 Hz). This seems attributable to the fact that only PFDM, but not R-R interval measurement, utilizes the periodicity of signals; i.e., the PFDM analyzed pulse wave with assuming it as a cosine function with slowly changing amplitude and phase. Even for ECG, a periodicity analysis method can estimate instantaneous heart rate from signals sampled at a low frequency (5 Hz) .
Although measurement of R-R intervals of ECG is the standard for HRV analysis, this method has practical limitations . Recordings of ECG require attachment of multiple electrodes and cables, which limit applications of HRV analysis in public or home health cares. Also, R-R intervals are intervals of event series. An appropriate interpolation is necessary to estimate the underlying modulator function that is the putative subject of HRV analysis. For this purpose, several interpolation methods with differing mathematic features have been proposed; but convincing physiological reasoning for selecting a method is lacking.
In contrast to ECG, recording of pulse wave requires only a single sensor, which allowed development of devices that can be used in daily life even everyday. Also, smooth waveform of pulse wave requires a lower sampling frequency (10–20 Hz) and its sinusoidal feature is advantageous to the PFDM as mentioned above. Furthermore, by the use of the PFDM, the signal putatively modulating the pulse interval is directly extracted as a continuous function. The signal can be used directly for spectral analysis without interpolation processes.
An apparent limitation of the PFDM of pulse wave is inability to assess the type of arrhythmias. A substantial part of such arrhythmia, however, can be detected by using appropriate criteria such as those using deviation of intervals from the local mean. In fact, the PFDM was able to detect persistent atrial fibrillation in this study. Also, even for the segments including frequent VE, the agreement for the pulse rate and LF amplitude are unaffected, although the HF amplitude was overestimated for the segments including VE>~10% compared with the HF amplitude obtained by R-R interval analysis. Interestingly, however, the exclusion of arrhythmic data has been reported to result in an underestimation of the HF amplitude for HRV analysis with R-R interval . It should be also noted that the optimal criteria for excluding abnormal beats are not uniform but subject to change depending on age and other conditions that could affect the magnitude of physiological PRV. This issue, however, is not specific to PFDM but common to PRV and HRV assessment at least in detection of atrial ectopic beats and heart blocks.
Another limitation of the PFDM may be caused by the physiological differences between pulse interval and R-R interval. Theoretically, the variability of pulse rate is the sum of the variability existing in R-R interval, pre-ejection period and pulse wave velocity. Constant et al  suggested that, in the standing position, respiratory fluctuation of pulse wave velocity might be important cause of respiratory pulse rate variation. Although the present study indicates usefulness of the PFDM for assessing PRV and good agreement between PRV and HRV during night sleep, its usefulness as a surrogate of HRV for assessing autonomic functions and mortality risk need to be examined de novo.
The PFDM of pulse wave signal provides a reliable assessment of PRV. Given the popularity of pulse wave equipments, this technology may open new ways to studies of long-term assessment of cardiovascular variability and dynamics among general populations.
This work was supported in part by the Research Grant from Suzuken Memorial Foundation (2003), by the Grant-in-Aid for Scientific Research (C) from the Japanese Ministry of Education, Culture, Sports, Science and Technology (15590765) and by the Research Grant (14A-9-09) for Nervous and Mental Disorders from the Japanese Ministry of Health, Labor and Welfare (2002–2004).
- Malik M, Camm AJ, Eds: Dynamic Electrocardiography. New York: Blackwell Futura; 2004.Google Scholar
- Pomeranz B, Macaulay RJ, Caudill MA, Kutz I, Adam D, Gordon D, et al.: Assessment of autonomic function in humans by heart rate spectral analysis. Am J Physiol 1985, 248: H151-H153.Google Scholar
- Pagani M, Lombardi F, Guzzetti S, Rimoldi O, Furlan R, Pizzinelli P, et al.: Power spectral analysis of heart rate and arterial pressure variabilities as a marker of sympatho-vagal interaction in man and conscious dog. Circ Res 1986, 59: 178–193.View ArticleGoogle Scholar
- La Rovere MT, Pinna GD, Hohnloser SH, Marcus FI, Mortara A, Nohara R, et al.: Baroreflex sensitivity and heart rate variability in the identification of patients at risk for life-threatening arrhythmias: implications for clinical trials. Circulation 2001, 103: 2072–2077.View ArticleGoogle Scholar
- Kleiger RE, Miller JP, Bigger JT Jr, Moss AJ: Decreased heart rate variability and its association with increased mortality after acute myocardial infarction. Am J Cardiol 1987, 59: 256–262. 10.1016/0002-9149(87)90795-8View ArticleGoogle Scholar
- Tsuji H, Venditti FJ Jr, Manders ES, Evans JC, Larson MG, Feldman CL, et al.: Reduced heart rate variability and mortality risk in an elderly cohort. The Framingham Heart Study. Circulation 1994, 90: 878–883.View ArticleGoogle Scholar
- Garde AH, Laursen B, Jorgensen AH, Jensen BR: Effects of mental and physical demands on heart rate variability during computer work. Eur J Appl Physiol 2002, 87: 456–461. 10.1007/s00421-002-0656-7View ArticleGoogle Scholar
- Kristal-Boneh E, Raifel M, Froom P, Ribak J: Heart rate variability in health and disease. Scand J Work Environ Health 1995, 21: 85–95.View ArticleGoogle Scholar
- Pichot V, Bourin E, Roche F, Garet M, Gaspoz JM, Duverney D, et al.: Quantification of cumulated physical fatigue at the workplace. Pflugers Arch 2002, 445: 267–272. 10.1007/s00424-002-0917-7View ArticleGoogle Scholar
- Bonnet MH, Arand DL: Heart rate variability in insomniacs and matched normal sleepers. Psychosom Med 1998, 60: 610–615.View ArticleGoogle Scholar
- Hayano J, Yasuma F: Hypothesis: respiratory sinus arrhythmia is an intrinsic resting function of cardiopulmonary system. Cardiovasc Res 2003, 58: 1–9. 10.1016/S0008-6363(02)00851-9View ArticleGoogle Scholar
- Dayawansa S, Umeno K, Takakura H, Hori E, Tabuchi E, Nagashima Y, et al.: Autonomic responses during inhalation of natural fragrance of Cedrol in humans. Auton Neurosci 2003, 108: 79–86. 10.1016/j.autneu.2003.08.002View ArticleGoogle Scholar
- Umemura M, Honda K: Influence of music on heart rate variability and comfort – a consideration through comparison of music and noise. J Hum Ergol (Tokyo) 1998, 27: 30–38.Google Scholar
- Constant I, Laude D, Murat I, Elghozi JL: Pulse rate variability is not a surrogate for heart rate variability. Clin Sci (Lond) 1999, 97: 391–397.View ArticleGoogle Scholar
- Bloomfield P: Complex demodulation. In Fourier Analysis of Time Series: An Introduction. Edited by: Bloomfield P. New York: Wiley; 1976:118–150.Google Scholar
- Hayano J, Taylor JA, Mukai S, Okada A, Watanabe Y, Takata K, et al.: Assessment of frequency shifts in R-R interval variability and respiration with complex demodulation. J Appl Physiol 1994, 77: 2879–2888.Google Scholar
- Hayano J, Taylor JA, Yamada A, Mukai S, Hori R, Asakawa T, et al.: Continuous assessment of hemodynamic control by complex demodulation of cardiovascular variability. Am J Physiol 1993, 264: H1229-H1238.Google Scholar
- Lippman N, Stein KM, Lerman BB: Comparison of methods for removal of ectopy in measurement of heart rate variability. Am J Physiol 1994, 267: H411-H418.Google Scholar
- Bland JM, Altman DG: Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986, 1: 307–310.View ArticleGoogle Scholar
- Shrout PE, Fleiss JL: Intraclass correlations: uses in assessing rater reliability. Psychological Bulletin 1979, 86: 420–428. 10.1037//0033-2909.86.2.420View ArticleGoogle Scholar
- Saul JP, Albrecht P, Berger RD, Cohen RJ: Analysis of long term heart rate variability: method, 1/f scaling and implications. Comp Cardiol 1987, 14: 419–422.Google Scholar
- Barros AK, Ohnishi N: Heart instantaneous frequency (HIF): an alternative approach to extract heart rate variability. IEEE Trans Biomed Eng 2001, 48: 850–855. 10.1109/10.936361View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.