1. Medical Hypotheses (1998) 51, 367-376
© Harcourt Brace & Co. Ltd 1998
Fractal organization of the pointwise correlation
dimension of the heart rate
E. NAHSHONI, E. ADLER*, S. LANIADO*, G. KEREN*
Department E, The Gehah Psychiatric Hospital, Petah-Tiqva, and Sackler School of Medicine, Tel Aviv,
Israeb *Department of Cardiology, Tel Aviv Medical Center, Tel Aviv, and Sackler School of Medicine,
Tel Aviv, Israel. Correspondence to: E. Nahshoni, POB 102, 49100 Petah-Tiqva, Israel
(Phone: +972 3 9258258; Fax: + 972 3 9241041)
Abstract - - Objective: To depict and quantify the degree of organization of the heart rate
variability (HRV) in normal subjects. Methods: A modified algorithm was created to estimate
series of "point-dimensions" (PD2) from interbeat (R-R) interval series of 10 healthy subjects
(21-56 years). Our innovation is twofold: (i) we quantified instances of low-dimensional chaos,
random fluctuations, and those for which our method failed to provide either (due to poor
statistics); (ii) consecutive subepochs of PD2s underwent a relative dispersion (RD) analysis,
yielding an index (D) which quantifies the dynamical organization of the heart rate generator.
Results: The mean values of PD2 series varied between 4.58 and 5.88 (mean +_SD=
5.21 +_0.41, n = 10). For group 1 (21-30 years, n = 6) we found an averaged PD2 of 5.49 _+0.27,
while for group 2 (47-56 years, n = 4) PD2 averaged 4.79 +_.0.17. The RD analysis performed
for subepochs of PD2s yielded both instances obeying fractal scaling (D < 1.5) and
stochasticity (D > 1.5). The average D for group 1 was 1.39 + 0.04 (14 subepochs) and for
group 2, 1.20 _+0.008 (8 subepochs). Paired t-test and Hartley F-max test for comparison
between D values and homogeneity of variance between the two groups were performed,
yielding P-values 0.004 and 0.02, respectively.
Conclusions: The complexity of the HRV seems to be modulated by a non-random fractal
mechanism of a 'hyperchaotic' system, i.e. it can be hypothesized to contain more than one
attractor. Also, our results support the 'chaos hypothesis' put forth recently, namely, the
complexity of the cardiovascular dynamics is reduced with aging. The index of relative
dispersion of the dimensional complexity has to be tested in various clinico-pathological
settings, in order to corroborate its value as a potential new physiological measure.
Introduction frequency and phases of biological oscillators, or to
the coupling of various regulatory feedback loops,
Physiological systems have long been recognized to thus engaging nonlinear mechanisms for elucidation
display complex temporal fluctuations, even during of the dynamics. Although, as in the physical sciences,
'steady state' conditions. Attempts were made to attri- solutions have resulted in 'linearizations', only during
bute them to random influences, which perturb the the last decade has a natural link been drawn between
Received 28 April 1997
Accepted 12 June 1997
367

2. 368 MEDICAL HYPOTHESES'
the mathematico-physical field of nonlinear dynamics theses, which motivated ongoing research efforts
and physiology. This has triggered an ongoing trend meant to quantify the dynamical characteristics of
of 'paradigm shift' in the medical sciences and in the heart rate dynamics under the assumption that
biological thinking in general. it evolves on a low-dimensional 'strange attractor'.
Since the advent of digital processing, the heart rate These attempts were based mainly on dimensional
became the most accessible and reliable signal for analysis, which resulted in correlation dimensions
analysis among cardiovascular variables. The heart (interpreted as a static measure of the number of
rate variability (HRV) is traditionally assessed using independent variables necessary to specify the state of
frequency (spectral analysis) and time (standard de- the system under study), ranging between 3.6 and 5.2
viations, interval occurrence histograms, etc.) domain in normal subjects (17). This was supported later, by
techniques. Using such techniques a complex coupling introducing another measure of deterministic chaos,
with respiration, baroreceptors, the nervous system, i.e. the largest Lyapunov exponent which yielded a
body temperature, metabolic rate, hormones, sleep finite positive value, thus demonstrating the property
cycles, etc. was revealed. For example, spectral analysis, of sensitivity to initial conditions, which is the hall-
which exposed activity bands in the frequency mark of chaotic behavior (18). But later estimates
domain comprising thermoregulation (~0.05Hz), of the correlation dimension were found to be much
baroreflex control of peripheral resistance (~ 0.1 Hz) higher (-8.5) than previously reported, thus pre-
and respiratory control (~ 0.2 Hz), was combined cluding firm conclusions as to the true nature of the
with pharmacological blockade to attribute the lower- heart rate generator (19).
band fluctuations (0.04--0.15 Hz) to the joint influence Recently, other modified measures of dynamical
of the sympathetic and parasympathetic arms of the complexity, mostly suited for nonstationary, noisy,
autonomic nervous system, while the higher fre- and limited record length signals, have been intro-
quency band (0.15-0.4 Hz) was shown to be purely duced. Among them is the estimation of the pointwise
parasympathetically mediated. The spectral signature correlation dimension (PD2), which provides more
of HRV was also related to various physiological information about the temporo-spatial evolution of
and pathophysiological settings, such as standing, the dominant complexity of the heartbeat (20,21).
hemorrhage and hypotension, which enhance the low This technique was applied to a very limited number
frequency fluctuations, while exercise and standing of subjects, from which no firm conclusions could
decrease the respiratory fluctuations. From the clinico- be drawn, except for one clinical study which corre-
pathological viewpoint, patients with heart failure lated a reduced dynamical complexity hours before
have diminished power spectrum at frequencies above the occurrence of lethal arrhythmias in high-risk
~ 0.02 Hz (1-7). The other arm of traditional analysis, patients (22).
the time domain analysis, has related decreased HRV In the light of the open questions and computa-
in diabetes mellitus, ischemic heart disease, conges- tional restrictions in this growing field of research,
tive heart failure (8), and also associated an increased we addressed the issue of the irregular nature of the
mortality in patients after acute myocardial infarction HRV in 10 healthy subjects. We computed correlation
(9). dimensions and the series of pointwise dimensions.
Taken together, these techniques have several short- We also introduced a modified version of the point-
comings. For example, spectral analysis is a method wise dimension algorithm, which, we believe, can
mostly suited for linear systems, while physiological depict both instances of low-dimensional chaos and
systems are inherently nonlinear. Also, time domain stochasticity. The complex relation between them
analysis, which is basically an averaging technique, was investigated using fractal techniques. The physio-
overlooks the dynamical nature. Thus, it appears that logical and clinical importance of the measure we
these techniques are often insufficient to characterize introduced is still unknown.
the complex behavior of the heart rate generator.
Since the last decade, nonlinear methods of analysis,
based on the paradigm of deterministic chaos (10), Methods
have permeated the realm of biomedical signal
analysis (11). This was motivated by the observation
Subjects
of an inverse power-law scaling (also called 1/f Ten volunteers, aged 21-56 (6 males, 4 females)
spectrum), which some chaotic systems may display, without symptoms or history of heart disease and
and by its association to the fractal concept (mani- under no medication, were recruited for the study.
fested by self-similarity over multiple orders of Their surface ECG, which showed no signs of patho-
temporal magnitude) (12-16). In the case of heart rate logy, was recorded at rest in a supine position, during
dynamics, these observations heralded new hypo- quite spontaneous breathing (~ 15 breaths/min) for

3. FRACTAL ORGANIZATION IN HEART RATE 369
20 rain. All recordings were done between 10 and series were time-delayed for successive embedding
12 a.m., and each subject was allowed to adjust dimensions (from m = 1 to m = 16). Within a given
comfortably for 10 min in a supine position before the embedding dimension, the distance (r) of each point
data were collected. They all gave informed consent to every other point was calculated. Their absolute
to the protocol. values were rank-ordered from the smallest to the
largest, and the range from the smallest to the largest
Data aquisition value was broken up into discrete intervals. Then, the
number of times a distance fell within an interval was
The ECG signals were continuously recorded using a counted. A cumulative histogram was then formed
laptop-based HIPEC ANALIZER HA-200/AH system by summing the number of instances for which a
(Aerotel - - computerized systems, Israel) with a distance was less than or equal to the upper boundary
sampling rate of 1000 Hz, and 16 bit signal resolu- of the interval. This is the correlation integral C(r).
tion. The ECG records were transferred to a personal C(r) was then plotted as a function of r on a log-log
computer for off-line analysis which started with a representation, resulting in a sigmoid-shaped curve
quality control procedure: visual inspection, baseline (in this case implying chaotic dynamics). The slope
shift evaluation and a 'moving average' (four points over the largest linear range (if there is one) was
averaging in succession along the record) for signal to measured, using linear regression (with a regression
noise ratio improvement. Then the interbeat intervals coefficient R 2 > 0.98). In this scaling range the local
(R-R) were computed using an algorithm developed exponent is constant and ~ d InC(r)/d In(r). Then, the
in our laboratory, with which an R wave threshold embedding dimension was advanced and its corre-
detection was combined with first derivative and QRS sponding slope was calculated. These slopes were
width considerations, for an accurate R wave detection. then plotted versus the embedding dimension, looking
for a saturation region, i.e. a region in which the
Attractor reconstruction slopes no longer grow. This plateau region was then
Usually the experimentalist is confronted with in- considered as the correlation dimension (D2), and its
ability to gain access to m simultaneous recordings value was calculated with a weighted average tech-
necessary to describe the system's trajectory in nique (each value in the plateau region was weighted
m-dimensional phase space. Thus, only one scalar by the variance of its underlying slope calculation).
observable can be monitored as a function of time. This process was also performed for randomized
Fortunately, it has been shown that certain properties versions of the R - R series (with similar mean and
of the dynamics are feasible through the method of variance) in order to provide confidence limits for our
time delays using Taken's theorem, as follows (23). calculations.
Consider a single time series regularly spaced in
time: xi = x%), i = 1..... N. Then a time lag "~is intro- Pointwise dimension of R-R intervals
duced, such that m-dimensional vectors are created:
xi= [x(ti), x(ti + "c)..... x(ti+ (m-1)x)]. This process is The 'point' estimate of the correlation dimension
termed embedding, and m is called the embedding (PD2) begins with the time lag calculation, followed
dimension. Through this reconstruction a phase by the embedding procedure, as described before.
space is spanned and dynamical and metric measures Then, starting with the initial point in the series, its
(Lyapunov exponents, dimensions) may be accessible. local correlation integral is calculated, i.e. the dis-
tances are taken with respect to this point and ranked-
ordered as usual, for each embedding dimension
Correlation dimension of R-R intervals
(m = 1..... 16). The slopes for each m were evaluated
The correlation dimension was calculated using the using a linear regression (R2> 0.98), and a slope
method of Grassberger and Procaccia (G-P) as follows values, corresponding to m = 8 ..... 16 were stored in
(24). First, for each R - R interval series, the normal- a file. The algorithm steps to the next point in the
ized autocorrelation function given by: series, and the whole procedure is repeated until the
whole file is exhausted. Then comes the procedure
~g('~) = {(I/N) £[R-R)i - < (R-R) >][(R-R)i+x -
that we call slope convergence, which calculates the
< (R-R) >] }/{(I/N) Z[R-R)i - < (R-R) >]2}
slope of the 9 slopes versus the embedding dimensions
where (m = 8 ..... 16) using linear regression. Our innova-
tion was to subject this to the imposition of three
< (R-R) > = (l/N) Z (R-R)/
conditions as follows: Ill if the slope was less than
was constructed, and its first zero crossing was calcu- 0.5 and larger than -0.23, we considered this as good
lated to provide the time lag (x) in beats. Then, the convergence and the PD2 could be estimated using

4. 370 MEDICAL HYPOTHESES
the weighted average technique as described before; dimensions was needed. Homogeneity of variance
(ii) if the slope was equal or larger than 0.5, we was tested by the Hartley F-max test. Statistical
considered it as if no saturation existed, and at significance was assumed if the null hypothesis could
this point (or time), the system probably manifested be rejected at the 0.05 probability level.
a random fluctuation. In order to incorporate such
a behavior into the sequence of PD2s, we decided,
quite arbitrarily, to take the average of the two highest Results
slope estimates, as the point-dimension, when such
a condition appears; (iii) if the slope was equal to Thc correlation dimension (D2) of R-R intervals
or less than --0.23, we considered it as if no slope varied from 3.29 to 5.16, with an overall mean of
convergence existed, and the dimensional estimate 4.01 ± 0.54 (Table 1). Fig. la illustrates one of the
at this point was excluded, possibly because of poor series of R-R intervals. This corresponding normal-
statistics. ized autocorrelation function is shown in Fig. lb. The
The results of the PD2 series were 'assigned' first zero crossing (x), in this case was equal to 6
according to the three conditions mentioned above. beats. The correlation integral (C(r)) for embedding
This provided us with the ability to discriminate the dimensions (m = 2,4,6,9,12,16) is shown in Fig. 2a,
points which manifested low-dimensional chaos and while the calculated slopes in the linear regions of
random behavior, from those for which a dimensional the log-C(r) representation, versus the embedding
estimate could not be achieved. From the above dimension, is shown in Fig. 2b. Note the convergence
output files we extracted sequences of dimensional towards a dimensional value of 4. Randomized ver-
subepochs, which were then suited for the relative sions of the R-R intervals have demonstrated, as
dispersion analysis. expected, non-convergence (Fig. 2c).
A sequence of pointwise dimensions (PD2s) versus
the reference point is shown in Fig. 3a. Note three
Dispersion analysis regions in the dimensional complexity plot, i.e. high
values (PD2 > 6), low-dimensional region (3 > PD2 < 6)
There are three basic methods of dispersion analysis
and zero-valued reference points, corresponding to
that can be applied to temporal observations (25). One
non-convergence due to poor statistics. This can be
of them, adopted in our study for each sequence of
seen from the histogram (Fig. 3b) showing the distrib-
calculated pointwise dimensions, is called relative
ution of the rounded dimensional values, including
dispersion (RD) analysis. Our intention was to try and
the points corresponding to stochasticity and to non-
see if the temporal evolution of PD2 series obeys any
calculability at both ends of the figure. For the subject
scaling properties. Thus for each subject, this simple
shown in the figure the average PD2 was 5.37 ± 0.93.
algorithm goes as follows: first, the mean, standard
In Fig. 4, four subepochs, each comprising ~150 PD2
deviation (SD), and RD% (= 100 x SD/mean) of the
values (corresponding to an average of about 2.5
original PD2 series were calculated. Then, pairs of
minutes'-record-length each) are shown. In Fig. 5 the
adjacent PD2s were averaged and their RD% values
logarithmic plot of the RD(%) versus the interval
were calculated, thus doubling the interval length.
Recursive pairing with doubling of each previous
interval length was carried out while its correspond-
Table 1 Correlation dimension (D2) of 10 healthy
ing RD% was calculated. This was done until the subjects at rest
whole record was exhausted. By plotting the RD%
against the interval length on a logarithmic scale, the Gender Age (years) HR ± SD Correlation dimension
slope was estimated using a least-squares linear fit. (beat/rain) (D2 ± ZkD2)
The fractal dimension (D) could thus be extracted
from the slope (slope --- l-D). In order to confirm the M 21 69.1 ± 5.6 4.51 ± 0.13
temporal organization of the PD2 series, randomized F 25 65.5 ± 3.0 3.58± 0.07
versions based on similar statistical characteristics M 26 65.2 ± 2.5 3.93 ± 0.09
F 28 67.6± 3.5 5.16 ± 0.02
(number of points, mean and standard deviation) were M 30 57.6 ± 4.6 3.99 ± 0.16
generated, and their RD analysis was also performed. M 30 54.2±2.3 4.50±0.19
M 47 71.4± 3.9 3.87 ± 0.21
M 56 60.9 ± 3.0 3.29 ± 0.19
F 56 64.7 ± 2.3 3.48 ± 0.21
Statistical analysis
F 56 66.8 ± 2.6 3.77 ± 0.03
All data are expressed as mean ± SD. A paired t-test mean ± SD 37.5 ± 13.7 64.3 ± 5.0 4.01 ± 0.54
was performed when comparison between fractal

5. FRACTALORGANIZA~ONINHEARTRATE 371
1.25
0.75
e-
•- 0.5-
I
e¢-
0.25
0
1 500 1000
Beat number
0.7
0.6
0.5
0.4--
0.3--
0.2
0.1-
0
-0.1 .'v , , .... 'IW ,v,' ' V V_
-0.2 m
100 200 300
b T
Fig. 1 (a) R - R intervals for one of the subjects (1296 intervals, 20 min). (b) The normalized
autocorrelation function of R - R intervals shown in (a). The first zero crossing was found to be 6 beats.
length (measured in beat number) for one of the 850. The overall mean values of the PD2 series varied
subepochs is shown. Its slope provides the fractal between 4.5 and 5.88 (mean = 5.21 ± 0.41, n = 10),
dimension of the dimensional complexity at a parti- but the mean PD2s of the various subepochs were
cular subepoch. smaller than the overall average, at least during one
Table 2 summarizes the results of the fractal subepoch for each subject. We divided the subjects
dimensions (D) of the subepochs of series of PD2s. into two groups according to their age. For group 1
The shortest subepoch consisted of 80 consecutive (21-30 years) the average PD2 varied between 5.19
dimensional values, while the longest consisted of and 5.88 (mean = 5.49 ± 0.27, n = 6), while the

6. 372 MEDICAL HYPOTHESES
0
-1,
-2.
-3
to
t- -4
-5
-6
-7 1'5 ' ' 2~0 . . . . 2~5 . . . .
Inr
Emb 9 --K--- Emb 12 ~ Emb 16
5 8
C C
0 O
r
r 4 r
? ?6
i 3 . i
0 O
n n 4
d
i
2 ?
m m
e e 2
n 1 [I
0 O
n 0 n
. . . . . . . . . . . . . . 0 . . . . . . . . 110 111 I i I4 I
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 2 3 4 5 6 7 8 9 1 13 1 1 16
b Embedding dimension c Embedding dimension
Fig. 2 The correlation integral C(r) versus r on a logarithmicplot. Seen are embeddingdimensions m = 2, 4, 6, 9, 12, 16. (b) The slope
of the scaling region as a function of the embeddingdimension (m), for a healthy subject (26 years). Note the saturation towards a
correlation dimension (D2) estimate of ~4. (c) For a randomlygenerated version of R-R intervals, the slope estimates of In C(r) versus
lnr, as a functionof the embeddingdimension, do not saturate.
relative dispersion analysis of their consecutive PD2 Discussion
series yielded both instances of fractal scaling
(D < 1.5) and stochasticity (D > 1.5). The averaged The concept of fractals, first coined by B. Mandelbrot
fractal dimension for this group was 1.39 4-0.04 (26), and its association with chaos theory, heralded
(14 subepochs). In group 2 (47-56 years), the PD2 novel insights into the realm of structural and
mean values ranged between 4.58 and 5.03 (mean = dynamical variability in the medical sciences (27)
4.79 ± 0.17, n = 4). Note that in group 2 the fractal and biology in general (11). During the last decade,
estimates ranged between 1.09 and 1.33 (mean = the intimate connection between deterministic chaos
1.20 ± 0.008, 8 subepochs), i.e. never exceeded 1.5. and fractal geometry has stimulated ongoing research
The t-test and the F-test showed statistical signi- efforts to quantify the dynamical aspects of the heart
ficance when the means and variances of the fractal rate generator. Babloyantz and Destexhe were the first
dimensions of the PD2 subepochs were compared to quantify its dynamical measures using chaos theory
(P values: 0.004 and 0.02, respectively). Note that the techniques (17). Their results (correlation dimensions,
overall results of the RD analysis were indicative Kolmogorov entropies and the largest Lyapunov
of fractal scaling (D < 1.5) in about 80% of all exponent), were supportive to the contention that the
subepochs tested. heart rate generator evolves on a low-dimensional

7. FRACTAL ORGANIZATION IN HEART RATE 373
10
I N of points
round(D2)-
alp-)0.50
0 N o! p-
4z m
0
| round(O2)- I N of p- 0
i
I round(D2}-
round(D2)-
2 H o f p-
3 H of
rouncl(D2)m 4 N o f
roun4(D2)- S N of
roarl4(D2)- 6 N o f
p-
pm 246
p- Se2
pm 146
0
"k
roun4(D2)- 7 X o f p - 64
round(D2)- 8 g of p,, ].
round(D2)- 9 N o f p- 0
round(D2)-20 N of p- 0
i
roun4(D2)-22 N of p- 0
round(D2)-12 N of p- 0
round(D2)-13 N of p- 0
round(D2)-14 N of p- 0
round(D2)-25 N of p- 0
round(D2)-26 H of p" 0
round(D2)-17 N of p,, 0
round(D2)-28 N of p- 0
round(D2)-lg N Of p- 0
round(D2)-20 N of p- @
H of exclu sip<-0.23 2v •
a
0 250 500 750 1000 1250
Fig. 3 (a) Serial pointwise dimensions (PD2s) as a function of the beat number (nref) for one of the subjects. The zero-valued PD2s are
only 'sign' of the instances for which a dimensional estimate could not be derived. (b) A histogram showing the distribution of the beat
number as a function of the rounded dimensional estimates. At the two extremes of the diagram we note the number of points for which a
random fluctuation is supposed to take place (slope > 0.5), and on the other side the number of point for which an estimate could not be
found (slope < -0.23).
chaotic attractor. Later, other groups provided sup- periodic behavior was seen in normal subjects, with
portive evidence to this hypothesis (18,28), although an increase in complexity during sleep (21). Recent
recently Kanters et al found weak evidence in favor studies found the dimensional complexity during
(19). Thus, the existence of low dimensional chaos in experimental myocardial infarction in pigs to decline
cardiac activity, at least at the whole organ level of significantly prior to the occurrence of ventricular
activity, is still an open question. Most of the estima- fibrillation (31). This has motivated Skinner et al
tions were based on dimensional estimations of the to evaluate lethal arryhythmias in various groups of
widely used Grassbeger-Procaccia algorithm. Imple- high-risk patients. It was found (with high specificity
mentation of this algorithm needs several precondi- and sensitivity) that the dimensional complexity
tions to be observed (29,30): an adequate choice of is reduced hours before the occurrence of lethal
embedding dimension, a suitable choice of the time arrhythmias (22).
delay needed to span the attractor, low level of noise Our results support the contention of low-
present in the system, stationarity, and the data set dimensional chaos, as proposed by others. The corre-
should not be too short. Some of these requirements lation dimension was found to vary between 3.29
are not attainable in biology and in physiology in and 5.16, with a mean of 4.01 + 0.54 (n = 10). The
particular. Moreover, the G-P algorithm provides a average pointwise dimension ranged between 4.58
dimensional estimate which averages out possible and 5.88, with a mean of 5.21 + 0.41 (n = 10). Further-
relevant dynamical features. Recently, the intro- more, we noticed that our subjects could be divided
duction of the pointwise dimension algorithm, which into two groups according to age as follows: group 1
provides series of 'point' dimensions, has provided (21-30) years, n = 6) had a higher average, 5.49 + 0.27,
some solutions to the limitations of the G-P algo- than group 2 (47-56 years, n = 4 ) at 4.79+ 0.17.
rithm, namely, non-stationarity and record length. Our innovation in this study was twofold. First, we
This method was implemented for heart transplant included in the dimensional complexity algorithm
recipients and the dimensional complexity was found means to include both instances of low-dimensional
to oscillate almost periodically (20). Also, a roughly chaos and stochastic bursts in a sequential manner,

8. 374 MEDICAL HYPOTHESES
10 2O
g
tJL, JJl|
RF'
Jd l
I0 10
i1. D ill
I'r
,I
Fig. 4 Four subepochs of sequential PD2s from the series shown in Fig. 3a. Each subepoch comprises about
150 PD2s.
i.e. as a function of the beat number. Second, this cant decline in the complexity of the cardiovascular
enabled us to apply a fractal technique (relative system (blood pressure and heart rate). Such findings
dispersion analysis) to explore the different subepochs may reflect the breakdown and decoupling of inte-
of dimension series for scale independence. We found grated physiologic regulatory systems with aging
that the older group manifested fractal scaling and may signal an impairment in the cardiovascular
(D < 1.5) in all subepochs tested. As for the younger ability to adapt efficiently to internal and external
group, only in 64% of tested subepochs did we find perturbations. This is contradictory to the sacred
fractal scaling (D < 1.5), while the rest was indicative principle of 'homeostasis', which was developed by
of a random control (D > 1.5). Moreover, the differ- Walter Cannon, and postulates that with disease and
ences in the averages and variances of the fractal aging the body is less able to maintain a constant
dimensions between the two groups were found to steady state, as a result of breakdown of its regulatory
be statistically significant. This is in contention with systems. Our findings support the chaos hypothesis of
results from other chaos-derived techniques imple- a 'homeokinetic' principle in physiology, namely,
mented by Kaplan et al on old versus young subjects physiological systems in young healthy subjects tend
(32). They found that the older group showed signifi- to fluctuate between a set of metastable states, thus

9. FRACTAL ORGANIZATIONIN HEART RATE 375
of the heart rate variability, which can be quantified
20 using fractal techniques, seems to contain 'mixing' of
both chaotic and random fluctuations. The nature
10
of such a behavior is not yet understood, but one
5
may hypothesize that an increase in the dimensional
complexity (D2 can be thought of as a measure
--,.. of independent variables necessary to describe the
2
system), may correspond to recruitment of several
1 subsystems influencing the heart rate generator, or
to the activation of more independent control loops.
0.50 A reduced complexity, on the other hand, may mani-
0.20
fest deactivation of control loops, or maybe increased
self-organization of some of these systems. Also, the
abrupt changes in the dimensional complexity may
0.10 represent shifts between different attractors of the
system. Such hypotheses, may better be resolved by
140 74 37 19 10 5 3
comparing the fractal dimensions of the dimensional
Fig. 5 Plot of RD% (relative dispersion) versus interval length on
complexity (and other measures of nonlinear tech-
a logarithmic scale. The fractal dimension (D) is derived from the niques) under different physiological and clinico-
slope (slope = l-D). pathological settings.
Currently, we are in the process of obtaining longer
data records from heart transplant recipients, in order
making the system more adaptable to its internal and to gain more insight regarding the value of the fractal
external surroundings (15,33). estimate of the dimensional complexity of the heart
We thus propose that the dimensional complexity rate generator, as a potential new dynamical measure.
T a b l e 2 Subepochs o f pointwise dimension (PD2) series, averaged P D 2 s for each subepoch, fractal dimension (D) for
each subepoch, and averaged P D 2 s for w h o l e records
Gender Age Nref PD2 ± SD average over subepochs D(RD) ± SD PD2 ± SD average over total record
M 21 1-150 5.28 ± 0.88 1.42 + 0.07 5.37 ± 0.93
300--450 5.09 ± 0.09 1.27 + 0.03
570-720 5.61 ± 0.73 1.49 ± 0.06
850-1000 4.35 ± 0.69 1.09 ± 0.04
F 25 1-90 4.45 ± 1.05 1.21 ± 0.09 5.19 ± 1.04
200-350 4.86 + 0.75 1.54 ± 0.06
M 26 130-196 6.01 ± 0.97 1.72 ± 0.11 5.83 ± 1.02
250-850 5.62 ± 0.97 1.19 ± 0.10
F 28 1-300 5.90 ± 0.90 L28 + 0.09 5.34 ± 0.72
301--600 5.03 ± 0.35 1.36 ± 0.12
M 30 1-200 4.91 ± 0.69 1.55 ± 0.09 5.30 ± 0.11
600--995 5.09 ± 1.10 1.13 ± 0.06
M 30 30-110 5.92 ± 0.83 1.57 ± 0.09 5.88 ± 1.08
125-295 5.34 ± 0.82 1.63 ± 0.09
M 47 1-350 3.92 ± 0.51 1.11 ± 0.04 4.69 ± 0,77
400-1250 4.99 ± 0,63 1.21 ± 0.06
M 56 1-300 4.95 ± 0,68 1.18 ± 0.07 5.03 ± 0,71
500-1000 5.11 ± 0.69 1.32 ± 0.05
F 56 100-300 4.63 ± 0.72 1.24 ± 0.07 4.87 ± 0.96
500-750 4.32 ± 1.21 1.09 ± 0.05
F 56 1-80 4.39 ± 0.69 1.33 ± 0.08 4.58 ± 0.64
200-380 4.63 ± 0.79 1.14 ± 0.08
mean ± SD 5.02 ± 0.54 1.32 ± 0.18 5.21 ± 0.41
M, male; F, female; Nref, sequences of consecutive data points' subepochs; PD2 + SD, averaged pointwise dimension + standard deviation;
D(RD), fractal dimension of each subepoch, derived from relative dispersion analysis.

10. 376 MEDICAL HYPOTHESES
References 15. Lipsitz L A. Age-related changes in the 'complexity' of
cardiovascular dynamics: a potential marker of vulnerability to
1. Appel M L, Berger R D, Saul J P, Smith J M, Cohen R J. Beat disease. Chaos 1995; 5(1): 102-109.
to beat variability in cardiovascular variables: noise or music? 16. Kobayashi M, Musha T. 1/f fluctuation of heartbeat period.
J Am Coil Cardiol 1989; 14:1139-1148. IEEE Trans Biomed Eng 1982; 29: 456-457.
2. Malliani A, Pagani M, Lombardi F, Cerutti S. Cardiovascular 17. Babloyantz A, Destexhe A. Is the normal heart a periodic
neural regulation explored in the frequency domain. Circula- oscillator. Biol Cybernetics 1988; 58:203-211.
tion 1992; 84: 482-492. 18. Rigney D R, Mielus J E, Goldberger A L. Is normal sinus
3. Malliani A, Pagani M, Lombardi F. Physiology and clinical rhythm 'chaotic'? measurement of Lyapunov exponents.
implications of variability of cardiovascular parameters with Circulation 1990; 82(4) Suppl. III, abstr.
focus on heart rate and blood pressure. Am J Cardiol 1994; 19. Kanters J K, Holstein-Rathlou N-H, Agner E. Lack of
73: 3C-9C. evidence for low-dimensional chaos in heart rate variability.
4. Akselrod S, Gordon D, Ubel F A, Shannon D C, Barger A C, J Cardiovasc Electrophysiol 1994; 5: 591-601.
Cohen R J. Power spectrum analysis of heart rate fluctuations: 20. Zbilut J P, Mayer-Kress G, Giest K. Dimensional analysis of
a quantitative probe of beat-to-beat cardiovascular control. heart rate transplant recipients. Math Biosci 1988; 90: 40-70.
Science 1981; 213: 220-222. 21. Mayer-Kress G, Yates F E, Benton L et al. Dimensional
5. Saul J P, Arai Y, Breger R D, Lilly L S, Colucci W S, Cohen analysis of nonlinear oscillations in brain, heart, and muscle.
R J. Assessment of autonomic regulation in chronic congestive Math Biosci 1988; 90: 155-182.
heart failure by heart rate spectral analysis. Am J Cardiol 1988; 22. Skinner J E, Pratt C M, Vybrial T. A reduction in the correla-
61: 1292-1299. tion dimension heartbeat intervals precedes imminent ventricular
6. Pomeranz B, Macaulay R J B, Caudill M A e t al. Assessment fibrillation in human subjects. Am Heart J 1993; 125: 731-743.
of autonomic function in humans by heart rate spectral 23. Rand D, Young L-S, eds. Detecting strange attractors in fluid
analysis. Am J Physiol 1985; 248 (Heart Circ Physiol 17): turbulence. In: Dynamical Systems and Turbulence. Berlin:
H151-H153. Springer-Verlag, 1981.
7. Madwed J B, Snads K E F, Saul J P, Cojen R J. Spectral 24. Grassberger P, Procaccia I. Measuring the strangeness of
analysis of beat-to-beat variability in HR and ABP during strange attractors. Physica 1983; 9D: 183-208.
hemorrhage and aortic constriction. In: Lown B, Malliani A, 25. Glenny R W, Robertson H T, Yamashiro S, Bassinghwaighte.
Prodocimi M, eds. Neural Mechanisms and Cardiovascular Applications of fractal analysis to physiology. J Appl Physiol
Disease. Fidia Research Series. Padova: Liviana Press, 1986, 1991; 70(6): 2351-2367.
Vol. 5: 291-301. 26. Mandelbrot B B. Thc Fractal Geometry of Nature. San
8. Takase B, Kurita A, Noritake Met al. Heart rate variability in Francisco: W H Freeman, 1982.
patients with diabetes mellitus, ischemic heart disease, and 27. West B J. Fractal Physiology and Chaos in Medicine. New
congestive heart failure. J Electrocardiol 1992; 25: 79-88. Jersey: World Scientific, 1990.
9. Kleiger R E, Miller J P, Bigger J T, Moss A J, and the Multi- 28. Ravelli F, Andolini R. Complex dynamics underlying the
center Post-infarction Research Group. Decreased heart rate human electrocardiogram. Biol Cyb 1992: 67; 57.
variability and its association with increased mortality after 29. Kantz H, Schreiber T. Dimension estimates and physiological
acute myocardial infarction. Am J Cardiol 1987; 59: 256-262. data. Chaos 1995; 5(1): 143-154.
10. Ott E. Chaos in Dynamical Systems. Cambridge: Cambridge 30. Schreiber T, Kantz H. Noise in chaotic data: diagnosis and
University Press, 1993. treatment. Chaos 1995; 5(1): 133-142.
11. Olsen L F, Degn H. Chaos in biological systems. Quarterly 31. Skinner J E, Carpeggiani C, Landisman C E, Fulton K W.
Rev Biophys 1985; 10(2): 165-221. Correlation dimension of hcartbeat intervals is reduced in
12. Goldberger A L, Rigney D R, West B J. Chaos and fractals conscious pigs by myocardial ischemia. Circulation Res 1991;
in human physiology. Sci Am 1990; 262: 42--49. 68: 966-976.
13. Goldberger A L, West B J. Fractals in physiology and 32. Kaplan D T, Furman M 1, Pincus S M, Ryan S M, Lipsitz L A.
medicine. Yale J Biol Med 1987; 60: 421-435. Aging and the complexity of cardiovascular dynamics. Bioph
14. Goldberger A L. Nonlinear dynamics, fractals and chaos: J 1991; 59: 945-949.
applications to cardiac electrophysiology. Ann Biomed Eng 33. Lipsitz L A, Goldberger A L. Loss of 'complexity' and aging.
1990; 18: 195-198. J Am Med Assoc 1992; 267: 1806-1809.