Fractals, Vol. 11, No.

2 (2003) 195–204
c World Scientific Publishing Company

DISCRIMINATION BY MULTIFRACTAL SPECTRUM

ESTIMATION OF HUMAN HEARTBEAT
INTERVAL DYNAMICS

M. MEYER∗ and O. STIEDL

Max Planck Institute for Experimental Medicine
Hermann-Rein Str. 3, D-37075 Goettingen, Germany
∗ meyer@exmda2.mpiem.gwdg.de

B. KERMAN
Infodynamics, Ancaster, Ontario, L9G 3M8, Canada

Received May 16, 2002

Accepted October 9, 2002

Abstract
The complexity of the cardiac rhythm is demonstrated to exhibit self-affine multifractal vari-
ability. The dynamics of heartbeat interval time series was analyzed by application of the
multifractal formalism based on the Cramer theory of large deviations. The continuous mul-
tifractal large deviation spectrum uncovers the nonlinear fractal properties in the dynamics of
heart rate and presents a useful diagnostic framework for discrimination and classification of
patients with cardiac disease, e.g. congestive heart failure. The characteristic multifractal spec-
tral pattern in heart transplant recipients or chronic heart disease highlights the importance of
neuroautonomic control mechanisms regulating the fractal dynamics of the cardiac rhythm.

Keywords: Cardiac Disease; Heart Rate Variability; Large Deviation Spectrum; Long-range
Dependence; Multifractals.

1. INTRODUCTION control expected to result from both intrinsic and

The beat-to-beat variation in the heart rate
of man and other mammals is generated by
a complex process and displays inhomogeneous,
non-stationary extremely irregular temporal orga-
nization. The physiological mechanisms of cardiac
extrinsic factors operating at different time scales
or resolution have not been identified clearly. Early
studies have attributed the complex dynamics of the
cardiac rhythm to be the result of an autonomous
low-dimensional deterministically chaotic system
but this concept has remained enigmatic. 1−5 Recent

195
196 M. Meyer et al.
studies have identified scale invariance or self-
affinity in the fluctuations of cardiac rate which are
characterized as 1/f-noise (strictly, f −B noise, with
a Fourier spectral density given in terms of the fre-
quency f by the power law f −B and B ∼ 1) and
quantified by a global Hurst-Hölder roughness ex-
ponent (H).6−13
Implicit in these studies was the assumption that
the cardiac interbeat time series was fractionally ho-
mogeneous, i.e. unifractal or uniscaling, and could
appropriately be described by a single H-coefficient.
Further generalization of the analysis beyond the
property of uniscaling leads to multiscaling or mul-
tifractals allowing the exponent H to depend on
time and to be chosen from an infinity of possi-
ble distinct values over the interval H min > 0 to
Hmax < ∞. Multifractals are therefore character-
ized by a plethora of scaling relations or power laws
with correspondingly many exponents. Multifrac-
tals, originally introduced by the seminal papers
of Frisch and Parisi14 and of Halsey et al.15 and
the work of Mandelbrot16 exemplify extreme vari-
ability and belong to the broad and unified notion
of self-affine fractal variation. Evidence for multi-
fractality in human heartbeat dynamics has been
suggested earlier8,9 and the multifractal formalism
based on the Legendre spectrum (see below) has re-
cently been invoked in the analysis of the variability
of human heart rate17−22 and posture.23
2. METHODS
2.1 Multifractal Spectrum
Biological signals typically (i) exhibit significant
long-range dependence (LRD), but display short-
term correlations and scaling behavior inconsistent
with strict self-similarity; (ii) the scaling behavior
of moments is a non-trivial (nonlinear) function of
the moment order as the signal is aggregated; and
(iii) the increments in the signal are inherently pos-
itive and hence non-Gaussian. Signals with these
properties fall naturally into the class of multifrac-
tal processes. Hence, multifractal analysis retrieves
positive measures or distributions exhibiting self-
similarity but non-homogeneous scaling.
The multifractal analysis of irregular but oth-
erwise arbitrary one-dimensional signals yields a
multifractal spectrum (f (α)) that may be viewed
as a measure of spikyness or global characteriza-
tion of the singularity structure of the data. It
provides information as to which singularities occur
in the signal and which are dominant. 24−26 More
precisely, f (α) is a graph where the abscissa repre-
sents the Hölder or regularity exponent (α) in the
signal and the ordinate is the fractal co-dimension
which measures the extent by which a given sin-
gularity is encountered (coarse-grained multifractal
spectrum). A Hölder exponent within the interval
[0, 1] means that the signal is continuous but not
differentiable at a given point considered. The
lower the exponent, the more irregular is the sig-
nal. More generally, the multifractal formalism,
assessing the singularity spectrum of the function
fg (α) yields information about the statistical be-
havior of the probability of finding a point with a
given Hölder exponent in the signal under changes
of resolution. The multifractal spectrum is a one-
dimensional curve where abscissa represents the
Hölder exponents actually present in the signal and
ordinates are related to the “amount” of points
where a given singularity is encountered.
2.2 Large Deviation Multifractal
Spectrum
Consider a positive continuous-time process X(t)
defined on [0, 1] where X varies considerably and
is nowhere differentiable, i.e. it cannot be approxi-
mated locally by a linear function. The strength of
spikyness, also referred to as the degree of Hölder
continuity, is characterized by a scaling exponent,
α(t):
α(t) := lim αnkn (1)
kn 2−n→t
with
1
αnkn := − log2 |X((kn + 1)2−n ) − X(kn 2−n )| (2)
n
and kn = 0, . . . , 2n − 1.
The smaller α(t), the larger the increments of X
at t and the more spiky X is at time t. For the unit
interval t ∈ [0, 1], the frequency of occurrence of a
given strength α is obtained by the coarse-grained
multifractal spectrum:
1
fg (α) = lim lim log2 Nn (α, ε) (3)
ε→0 n→∞ n
with
Nn (α) = #{αnkn ∈ (α − ε, α + ε)} . (4)
The function fg (α) takes values between 0 and 1
and is typically concave shaped like a ∩. The
 Multifractal Spectrum Estimation of Human Heartbeat Interval Dynamics 197

smaller fg (α), the “fewer” points t exhibit α(t) ' α. The transform τ ∗ in Eq. (6) is the Legendre trans-
If ᾱ denotes the value of α(t) assumed by “most” form of τ . While fg (α) is hard to assess directly on
points t, then fg (ᾱ) = 1, i.e. for a significant num- experimental signals, because of the double limit in
ber of kn = 0, . . . , 2n − 1, the increments of size Eq. (3), estimation of τ (q) is easier and f (α) is con-
n
4nk [X] := |X((kn + 1)2−n ) − X(kn 2−n )| = 2−nαk ' veniently determined by the Legendre multifractal
2−nᾱ and hence X varies smoothly with respect to spectrum fl (α):
the scale of measurement ε. fg (α) is the rate func-
tion, also referred to as partition function, τ (q), of fl (α) := τ ∗ (α) = inf (qα − τ (q)) . (7)
the Large Deviation Principle (LDP). 26−33 q∈IR
The multifractal large deviation spectrum f g (α)
is closely related to the partition function τ (q) 31,34 fl (α) may provide less information than f g (α) since
defined as the back-transformation yields
 n

2X−1
1
τ (q) = lim log 2 IE  (4nkn [X])q  . (5) fg (α) ≤ fg (α)∗∗ = fl (α) (8)
n→∞ −n
k =0 n

Assuming that τ (q) exists and is differentiable for
all real q, application of the LDP Gärtner-Ellis 28,31
theorem shows that the double limit of Eq. (3) exists
for all α, and, moreover
fg (α) ≤ τ ∗ (α) := inf (qα − τ (q)).
q
where fg (α)∗∗ is the concave hull of fg .26,35 For a
generalized analytical estimate of f (α) on a typical
path of X, the multifractal formalism 34,36−38 yields
(6) fg (α) ≤ fl (α) ≤ τ ∗ (α) . (9)

(a) (b)
6 0.02
Weierstrass Function Multinomial Measure
H(t)=t Base: 3
N=8192 Weights: 0.1,0.3,0.6
4 0.015 N=6561

2 0.01

0 0.005

−2 0
0 0.2 0.4 0.6 0.8 1 0 2000 4000 6000
t
(c) (d)
Legendre Legendre
Large Deviation Large Deviation
1 1 Theoretical

0.8 0.8
f(α)

f(α)

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 0 0.5 1 1.5 2 2.5
α α

Fig. 1 Multifractal spectra of classical functions and measures. (a) Generalized Weierstrass function. (b) Trinomial measure.
(c) and (d) Legendre spectrum (fl (α)) and large deviation spectrum (fg (α)) of (a) and (b).
198 M. Meyer et al.

2.3 Continuous Large Deviation spaced between the minimum and maximum val-
Multifractal Spectrum ues of the coarse-grain Hölder exponents at size η.
c,ε η(α) yields the large deviations from the “most
fg,η
The computation of the large deviation spectrum frequent” singularity exponent and thus displays in-
(α, fg (α)) is based on kernel-density estimation of formation about the occurrence of rare events such
the coarse-grain Hölder exponents α nkn . The estima- as bursts (small α).
tor of fg (α) is optimized for (i) implicit dependency Figure 1(a) and (b) shows the estimation of
between precision ε and resolution n, (ii) spatial c,ε η(α) and f (α) for the synthesized general-
fg,η l
adaptation of precision ε, and (iii) stable resolution ized Weierstrass function with smoothly increas-
within the range nmin ≤ n ≤ nmax , yielding a new ing regularity (h(t) = t) and a trinomial deter-
definition, the continuous large deviation multifrac- ministic measure40–42 at resolution 8 (38 = 6561
tal spectrum, fg,ηc,ε η(α).39 For the estimation of αn ,
kn intervals) constructed iteratively through a cas-
n
2 dyadic intervals that partition [0, 1] are consid- cade structure (c) and (d). The estimate of
ered for each resolution n. For each interval, α nkn fl (α) agrees reasonably well with the kernel es-
is computed which is the logarithm of a quantity timate of fg,η c,ε η(α) for the classical Weierstrass

measuring the variation of the signal in the interval function and just presents a concave approxima-
(the maximum minus the minimum of the signal in tion to fg,η c,ε η(α). For the trinomial measure,
the dyadic interval) divided by the logarithm of the both fl (α) and fg,η c,ε η(α) match quite well with

size of the interval. Interval discretization is linearly the theoretical spectrum. Contrary to the more

(a)
1400
RR−interval

1200
1000
800
600
400
5 10 15 20 25 30 35 40
(b) Time (min)
1
α

0.5

0
500 1000 1500 2000 2500 3000
(c) Number of points
1
α

0.5

0
500 1000 1500 2000 2500 3000
(d) Number of points
1.2
1
0.8
fc,εη(α)

0.6 αmode 0.39

g,η

αmin 0.10
0.4
α 0.92
max
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
α

Fig. 2 Multifractality in heartbeat interval time series. (a) Successive heartbeat intervals (msec) as function of time (∼ 40
minutes) from a healthy subject in resting supine posture. The cardiac time series exhibits wild irregular fluctuations and
non-stationary behavior. (b) The coarse-grain Hölder exponent (α) distributed over the interval [0, 1], denoting the range
from irregularity to regularity, reflects the heterogeneity that is present in the cardiac time series irrespective of the fact that
the subject was in steady-state resting supine conditions. The broad range of coarse-grain Hölder exponents associated with
the signal reflects the complex temporal organization of singularities characterized by the range of irregular fluctuating Hölder
exponents. (c) Color coding of (b). The color spectrum goes from blue to cyan to yellow to orange and to red. Blue indicates
small values of α while yellow-orange indicates large values of α. (d) The continuous large deviation multifractal spectrum,
c,ε
fg,η η(α), of the Hölder coefficients (α) shows a smooth concave function supported by αmin to αmax with mode, αmode [0.39, 1],
indicated by filled circle. The extracted parameters, αmode , αmin and αmax , respectively, are used to quantify the spectra for
the purpose of discrimination and classification of heartbeat interval time series of normal subjects and patients with cardiac
disease.
 Multifractal Spectrum Estimation of Human Heartbeat Interval Dynamics 199
convenient and generally more robust (though at
the expense of severe loss of information) Legen-
dre spectrum, fl (α), which is based on the Legen-
dre transformation19 of the Rényi exponents (τ (q))
of the qth moments of the measure and is always
shaped like a ∩ (concave and thus continuous, and
almost everywhere differentiable), the continuous
large deviation multifractal spectrum, f g,ηc,ε η(α), is
superior in that it does not need to be concave and
hence is more appropriate in a general setting. Fig-
ure 2(a) shows a highly irregular cardiac interval
time series recorded over ∼ 40 minutes (∼ 3000
beats) of a healthy human subject resting quietly
in supine posture. The multifractal estimation of
the coarse-grain Hölder exponents α nkn [Fig. 2(b)
and (c)] is used to determine the continuous large
c,ε η(α) [Fig. 2(d)].
deviation multifractal spectrum f g,η
3. EXPERIMENTAL PROCEDURES
The cardiac interval time series (length ∼ 10 5
beats) of healthy subjects (n = 9, seven males,
two females; mean age ± SD, 33.7 ± 4.6 years)
and patients with congestive heart failure (n =
11, nine males, two females; mean age ± SD,
56.4 ± 7.2 years) were obtained by 24-hour Holter
monitoring. The electrocardiographic database was
obtained using a dual-channel ambulatory ECG
recorder (Model R6, Custo Med, München, Ger-
many) provided with RAM-memory and advanced
data compression technology (sampling rate 500 Hz,
resolution 2 msec). Additional short-term steady-
state episodes (40 minutes) in healthy control
subjects, in patients with non-sustained ventricu-
lar tachycardia, and in heart transplant recipients
(< two years after transplantation) were recorded
by a custom-made dual-channel battery-powered
miniature ECG recorder amplifier interfaced to a
PC by a fiber-optical link (sampling rate 1200 Hz,
resolution ∼ 0.8 msec).
The digitized ECG was automatically analyzed
using an adaptive QRS-template pattern-matching
algorithm to obtain a discrete cardiac time series
or function, x(t) = ti+1 − ti , i.e. the time in-
terval between successive R-wave maxima of the
ECG. Ectopic beats or outliers were identified by
fitting a third order autoregressive model to the in-
terbeat interval data points using multiples (3.5)
of the interquartile distance as detection threshold
and corrected by linear-spline interpolation. Pre-
processing was reviewed by a senior cardiologist
(M.M.). All computations were performed within
the MATLAB environment. The C-LAB routine
for calculation of continuous large deviation mul-
tifractal spectra was obtained from J. L. Véhel at
http://www.inria.fr/fractales.
4. RESULTS
4.1 Cardiac Dynamics in Health
and Disease
c,ε η(α), is a
For all healthy subjects, the spectrum, f g,η
smooth concave function (group mean ± SD, α mode
0.36 ± 0.02; αmin 0.09 ± 0.01; αmax 0.50 ± 0.03) over
a broad range of Hölder exponents α [Fig. 3(a)].
The broad range spectrum indicates that heart rate
of healthy subjects exhibits multifractal dynamics,
i.e. the normal cardiac rhythm displays self-affine
multifractal variability. In contrast [Fig. 3(b)],
the spectrum of heart rate dynamics from patients
with congestive heart failure consistently displays
a marked departure from concavity showing a
pronounced trough on the increasing part of the
spectrum and αmode is shifted to larger Hölder ex-
ponent values (group mean ± SD, αmode 0.45±0.03;
αmin 0.09 ± 0.02; αmax 0.53 ± 0.04). The shift
of the αmode to larger values and the “removal”
of low singularity strength (0.1 ≤ α ≤ 0.4) ren-
der “smoothing” and indicate that the cardiac time
series in chronic cardiac disease is more regular
while the degree of multifractality given by the
αmax −αmin difference is not materially different be-
tween the two groups. The pattern of the spectrum
estimate in cardiac patients appears to uncover a
superposition of two “basic” spectra reminiscent of
those which are observed by synthesis of theoretical
multinomial measures, e.g. mixing or sum of two
binomial measures. This can be taken as an indica-
tion for (at least) two different phenomena underly-
ing perturbation of the dynamics in a pathological
condition — congestive heart failure. Less preva-
lent parasympathetically-mediated cardiac control
whereby the heart operates in a sympathetically-
dominated regime signifies the pathophysiology of
advanced chronic heart disease.
4.2 Circadian Cardiac Dynamics
In order to analyze the circadian stability of the
c,ε η(α) spectrum and to assess the impact of
fg,η
200 M. Meyer et al.
(a) (b)
1.2 1.2

1 1

0.8 0.8

fc,εη(α)

fc,εη(α)
0.6 0.6
g,η

g,η
0.4 0.4

α 0.36 ± 0.02 α 0.45 ± 0.03

mode mode

0.2 0.2

Healthy Heart Failure

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
α α

Fig. 3 Multifractality in health and cardiac disease. Group averages ± SD of the continuous large deviation multifractal
c,ε
spectrum fg,η η(α) versus Hölder exponents (α) in healthy subjects (a) and patients (b) with congestive heart failure. The
broad shape of the multifractal spectra exemplifies multifractal behavior in the cardiac interbeat rhythm which is different
between the healthy group and the heart-failure group. The spectra of the cardiac-failure group display a markedly left-sided
bi-modal shape and shift of αmode towards higher values.

extrinsic factors such as physical activity, the spec- (spontaneous versus paced breathing over a range
tra were calculated for one hour sub-epochs of of six to 48 breaths per minute) did not reveal any
the full-length 24-hour database. While the heart systematic changes in the fg,ηc,ε η(α) spectrum (data

rate of a healthy subject typically shows a circa-
dian rhythm with shorter interbeat intervals (higher
heart rates) during day-time episodes and longer in-
tervals (lower heart rates) during night-time epochs
[Fig. 4(a)], cardiac failure patients demonstrate
little, if any, circadian changes [Fig. 4(b)]. In-
deed, absence of a circadian rhythm in the clin-
ical setting is interpreted as evidence for chronic
heart failure. The multifractal spectrum of the
cardiac time series does not undergo appreciable
circadian changes both in healthy subjects and
cardiac patients [Figs. 4(c) and (d)] and α mode re-
mains essentially unchanged in the course of the
day [Figs. 4(e) and (f)]. Thus, the physiological
mechanisms controlling the dynamics of heart rate
appear to be independent of regulatory circadian
factors and the form of fg,ηc,ε η(α) is solely a function
of intrinsic factors affecting cardiac activity.
The analysis of short-term data from healthy
subjects recorded in steady-state conditions (∼ 40
minutes) comparing different postures (supine ver-
sus sitting) and the effects of breathing frequency
not shown). These findings along with the absence
of circadian changes suggest that the multifractal
spectrum of cardiac dynamics is unlikely to re-
sult from extrinsic factors such as physical activity,
posture or respiration. This feature of the multi-
fractal spectrum facilitates the comparison of car-
diac dynamics across species, e.g. humans versus
mice, humans and mice exhibiting identical spectra
irrespective of the fact that heart rate in mice is
about eight times higher than in humans (data not
shown).
4.3 Discrimination of Healthy
and Diseased Cardiac
Autonomic Function
c,ε η(α) spectrum clearly
The calculation of the fg,η
discriminates healthy subjects from heart-failure
patients when based on a single quantity: α mode
(Fig. 5). Further analysis of the database by
randomly selecting subsets of varying data lengths
 Multifractal Spectrum Estimation of Human Heartbeat Interval Dynamics 201

(a) (b)
1400 1400

1200 Healthy 1200 Cardiac failure

RR−interval

RR−interval
1000 1000

800 800

600 600

400 400
5 10 15 20 5 10 15 20
(c) Hours (d) Hours
1.2 1.2

1 1

0.8 0.8
fg,ηη(α)

fg,ηη(α)
0.6 0.6
c,ε

c,ε
0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
(e) α (f) α
1 1

0.8 0.8
Mean 0.42 ± 0.02 Mean 0.50 ± 0.05
0.6 0.6
mode

mode
0.4 0.4
α

0.2 α 0.2

0 0
5 10 15 20 5 10 15 20
Hours Hours

Fig. 4 Circadian heart rate dynamics. (a) and (b) Mean RR-interval ± SD of one hour subepochs versus time of the day
in a healthy subject and a patient with congestive heart failure. The heartbeat interval time series of normal subjects exhibit
a circadian rhythm with lengthening of RR-intervals (lower heart rate) during night-time episodes which is almost eliminated
c,ε
in chronic cardiac failure. (c) and (d) Group average ± SD of α versus fg,η η(α) of one hour subepochs. The multifractal
spectra display a characteristic difference of shape between health and cardiac failure but exhibit a remarkable stability over
a circadian 24-hour cycle. (e) and (f ) The mode of the multifractal spectra is lower in normal subjects and higher in cardiac
patients but would not undergo systematic circadian-related changes.

reveals that statistical discrimination of healthy and
diseased individuals may be achieved on the basis of
any arbitrarily selected segment of 8192 data points
which corresponds to an ∼ three-hour record.
The physiological mechanisms underlying the
multifractal variability in the cardiac rhythm have
not been identified. Intrinsic instability, extrin-
sic stochastic perturbations and a changing en-
vironment act together to produce the intriguing
irregular patterns. In healthy subjects, an impor-
tant mechanism is control of sinus node activity
mediated by the autonomous nervous system. A
straightforward approach as to the significance of
autonomic cardiac control for the multifractal dy-
namics of heart rate is facilitated by studies in
recipients of a cardiac transplant. The database
of healthy control subjects, heart transplant re-
cipients (< two years after transplantation), and
patients with non-sustained ventricular tachycar-
dia was obtained from 40-minute records (Fig. 6).
c,ε η(α) spectrum in heart transplant recipi-
The fg,η
ents [Fig. 6(b)] shows a preservation of α mode but
cutoff of the upper-range Hölder exponents indi-
cating breakdown of multifractality-strength which
is given by the αmax − αmin difference. Thus, the
transplanted denervated allograft in transplant re-
cipients which is effectively deprived of its neu-
roautonomic control is operating over a narrowed
multifractal regime [Fig. 6(b)]. In contrast, patients
with episodes of ventricular tachycardia [Fig. 6(c)]
demonstrate enhanced broadening of the spec-
trum and a bimodal shape similar to that ob-
served in congestive heart failure [cf. Figs. 3(b)
and 4(b)]. These findings emphasize the differ-
ential effects of absence (in cardiac transplanta-
tion) or degradation (in ventricular tachycardia) of
autonomous nervous system influences on the mul-
tifractal spectrum of the cardiac rhythm. However,
202 M. Meyer et al.
(a) (b)
0.55 0.55

0.5 Heart failure 0.5

0.45 0.45

mode

mode
α

α
0.4 0.4

0.35 0.35

Healthy

0.3 0.3
0.3 0.35 0.4 0.45 0.5 0.55
α −α Healthy Heart failure
max min

Fig. 5 Stratification of healthy subjects and patients with congestive heart failure by continuous large deviation multifractal
spectrum estimation. (a) Mode of spectrum (αmode ) versus degree of multifractality (αmax −αmin ) based on full-length 24-hour
time series records. The multifractal approach clearly discriminates healthy from heart-failure subjects. (b) Overall medians
of αmode presented in box-plot format for healthy and cardiac failure subjects. The statistical significance is easily evident by
non-overlapping notches. +, denotes outlier. The results demonstrate that a single quantity, α mode , may be used successfully
for discretization of normal subjects and patients with cardiac disease.

(a) (b) (c)

1.2 1.2 1.2

Cardiac Ventricular

Healthy transplantation tachycardia

1 1 1

0.8 0.8 0.8

η(α)

0.6 0.6 0.6

g,η
fc,ε

0.4 0.4 0.4

n = 17 n = 21 n=8

0.2 0.2 0.2

αmode 0.47 ± 0.03 αmode 0.43 ± 0.06 αmode 0.51 ± 0.06

α 0.17 ± 0.03 α 0.18 ± 0.08 α 0.09 ± 0.01

min min min

αmax 0.65 ± 0.06 αmax 0.48 ± 0.07 αmax 0.84 ± 0.04

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1

α α α

Fig. 6 Multifractality and neuroautonomic cardiac control. Multifractal spectra from normal subjects (a), heart transplant
recipients (b), and patients with non-sustained ventricular tachycardia (c). The different patterns of the multifractal spectra
exemplify the importance of neuroautonomic cardiac control in generating the broad-range multifractal spectrum of healthy
cardiac dynamics. Cardiac disease is associated with a characteristic pathological shape of the multifractal spectrum.
 Multifractal Spectrum Estimation of Human Heartbeat Interval Dynamics 203
the precise pathology of neurocardiac control mech-
anisms that is uncovered by changes in shape of the
multifractal spectrum remains to be determined.
5. CONCLUDING REMARKS
The multifractal properties of heartbeat dynam-
ics elicited by control mechanisms that regulate
the cardiac rhythm may be associated with the
behavior of dynamical systems typically operating
near a critical point of phase transition, i.e. far
from equilibrium.43,44 Multifractality or multiscal-
ing in cardiac time series signifies the interac-
tion of coupled mixed-feedback systems operating
over a wide range of time scales. These proper-
ties help provide responsiveness to environmental
stimuli (elasticity) while preserving a relative insen-
sitivity to errors (error tolerance). 45 From a prac-
tical point of view, the continuous large deviation
multifractal spectrum of heartbeat interval time se-
ries is expected to provide a useful diagnostic frame-
work for discretization and classification of patients
with cardiac disease and may be applied to other
biological time series data.
ACKNOWLEDGMENTS
This work was supported by BMFT Genomic
Networks — Genetic Susceptibility to Heart Failure
and Predictors of Therapeutic Response.
REFERENCES
1. J. Lefebvre et al., “Predictability of Normal Heart
Rhythms and Deterministic Chaos,” Chaos 3, 267
(1993).
2. Y. Yamamoto et al., “Operation Everest II: An
Indication of Deterministic Chaos in Human Heart
Rate Variability at Simulated Altitude,” Biol.
Cybern. 69, 205 (1993).
3. J. K. Kanters, N. H. Holstein-Rathlou and
E. Agner, “Lack of Evidence for Low-Dimensional
Chaos in Heart Rate Variability,” J. Cardiovasc.
Electrophysiol. 5, 128 (1994).
4. G. Sugihara, W. Allan, D. Sobel and K. D. Allan,
“Nonlinear Control of Heart Rate Variability in
Human Infants,” Proc. Natl. Acad. Sci. USA 93,
2608 (1996).
5. C.-S. Poon and C. K. Merrill, “Decrease of Cardiac
Chaos in Congestive Heart Failure,” Nature 389, 492
(1997).
6. C.-K. Peng, S. Havlin, H. E. Stanley and A. L.
Goldberger, “Quantification of Scaling Exponents
and Crossover Phenomena in Nonstationary Time
Series,” Chaos 5, 82 (1995).
7. C.-K. Peng et al., “Fractal Mechanisms and Heart
Rate Dynamics: Long-Range Correlations and Their
Breakdown with Disease,” J. Electrocardiol. 28, 59
(1996).
8. M. Meyer et al., “Is the Heart Preadapted to
Hypoxia? Evidence from Fractal Dynamics of
Heartbeat Interval Fluctuations at High Altitude
(5.050 m),” Int. Phys. Behav. Sci. 33, 9 (1998).
9. M. Meyer et al., “Stability of Heartbeat Interval
Distributions in Chronic High Altitude Hypoxia,”
Int. Phys. Behav. Sci. 33, 344 (1998).
10. P. Ch. Ivanov et al., “Scaling Behavior of Heartbeat
Intervals Obtained by Wavelet-Based Time-Series
Analysis,” Nature 383, 323 (1996).
11. P. Ch. Ivanov et al., “Scaling and Universality
of Heart Rate Variability Distributions,” Physica
A249, 587 (1998).
12. S. Thurner, M. C. Feurstein and M. C. Teich,
“Multiresolution Wavelet Analysis of Heartbeat
Intervals Discriminates Healthy Patients from Those
with Cardiac Pathology,” Phys. Rev. Lett. 80, 1544
(1988).
13. L. A. N. Amaral, A. L. Goldberger, P. Ch. Ivanov
and H. E. Stanley, “Scale-Independent Measures and
Pathologic Cardiac Dynamics,” Phys. Rev. Lett. 81,
2388 (1998).
14. U. Frisch and G. Parisi, “Fully Developed Turbu-
lence and Intermittency,” in Turbulence and Pre-
dictability in Geophysical Dynamics and Climate
Dynamics, eds. M. Ghil et al. (North-Holland,
Amsterdam, 1985), pp. 84–88.
15. T. C. Halsey et al., “Fractal Measures and Their
Singularities: The Characterization of Strange Sets,”
Phys. Rev. A33, 1141 (1986).
16. B. B. Mandelbrot, Multifractals and 1/f Noise
(Springer, New York, 1998).
17. Z. R. Struzik, “Revealing Local Variability Prop-
erties of Human Heartbeat Intervals with Local
Effective Hölder Exponent,” Fractals 9, 77 (2001).
18. Z. R. Struzik and J. van Wijngaarden, “Cumulative
Effective Hölder Exponent Based Indicator for Real-
Time Fetal Heartbeat Analysis During Labor,” in
Emergent Nature: Fractals 2002, ed. M. M. Novak
(World Scientific, 2002), pp. 45–55.
19. P. Ch. Ivanov et al., “Multifractality in Human
Heartbeat Dynamics,” Nature 399, 461 (1999).
20. P. Ch. Ivanov et al., “From 1/f Noise to Multifractal
Cascades in Heartbeat Dynamics,” Chaos 11, 641
(2001).
21. A. L. Goldberger et al., “Fractal Dynamics in Phys-
iology: Alterations with Disease and Aging,” PNAS
99, 2466 (2002).
22. P. Ch. Ivanov, A. L. Goldberger and H. E.
Stanley, “Fractal and Multifractal Approaches in
204 M. Meyer et al.
Physiology,” in The Science of Disasters, eds.
A. Bunde et al. (Springer, New York, 2002),
pp. 218–257.
23. Y. Shimizu et al., “Multifractal Spectra as a Mea-
sure of Complexity in Human Posture,” Fractals 10,
103 (2002).
24. K. J. Falconer, Fractal Geometry: Mathematical
Foundations and Applications (Wiley, New York,
1990).
25. S. Jaffard, “Multifractal Formalism for Functions: I.
Results Valid for All Functions, II. Self-Similar Func-
tions,” SIAM J. Math. Anal. 28, 944 (1997).
26. J. L. Véhel and R. Vojak, “Multifractal Analysis
of Choquet Quantities: Preliminary Results,” Adv.
Appl. Math. 20, 1 (1998).
27. H. Cramèr, in Actualites Scientifiques et Indus-
trielles, No. 736, Colloque consacre a la theorie des
probabilities (Hermann, 1938), pp. 5–23.
28. R. S. Ellis, Entropy, Large Deviations, and Statistical
Mechanics (Springer, New York, 1985).
29. J. A. Bucklew, Large Deviation Techniques in Deci-
sion, Simulation, and Estimation (Wiley, New York,
1990).
30. A. Dembo and O. Zeitouni, Large Deviations Tech-
niques and Applications (Jones and Bartlett, 1993).
31. R. Riedi, “An Improved Multifractal Formalism and
Self-Similar Measures,” J. Math. Anal. Appl. 189,
462 (1995).
32. F. den Hollander, Large Deviations (American Math-
ematical Society, 2000).
33. D. Veniziano, “Large Deviations of Multifractal
Measures,” Fractals 10, 117 (2002).
34. R. H. Riedi, M. S. Crouse and V. J. Ribiero, “A Mul-
tifractal Wavelet Model with Application to Network
Traffic,” IEEE Trans. Info. Theory 45, 992 (1999).
35. G. Brown, G. Michon and J. Peyrière, “On the
Multifractal Analysis of Measures,” J. Stat. Phys.
66, 775 (1992).
36. J. F. Muzy, E. Bacry and A. Arneodo, “The Multi-
fractal Formalism Revisited with Wavelets,” Int. J.
Bifurc. Chaos 4, 245 (1994).
37. A. Arneodo, E. Bacry and J. F. Muzy, “The Ther-
modynamics of Fractals Revisited with Wavelets,”
Physica A213, 232 (1995).
38. A.-L. Barabási and H. E. Stanley, Fractal Concepts
in Surface Growth (Cambridge University Press,
Cambridge, 1995).
39. C. Canus, J. L. Véhel and C. Tricot, “Continu-
ous Large Deviation Multifractal Spectrum: Defini-
tion and Estimation,” in Fractals and Beyond, ed.
M. M. Novak (World Scientific, Singapore, 1998),
pp. 117–128.
40. B. B. Mandelbrot, “Limit Lognormal Multifractal
Measures,” Physica A163, 306 (1990).
41. R. Holley and E. Waymire, “Multifractal Dimensions
and Scaling Exponents for Strongly Bounded Ran-
dom Cascades,” Ann. Appl. Prob. 2, 819 (1992).
42. R. H. Riedi and I. Scheuring, “Conditional and Rel-
ative Multifractal Spectra,” Fractals 1, 153 (1997).
43. C. Meneveau and K. R. Streenivasan, “Simple
Multifractal Cascade Model for Fully Developed
Turbulence,” Phys. Rev. Lett. 59, 1424 (1987).
44. P. Ch. Ivanov, A. L. N. Amaral, A. L. Goldberger
and H. E. Stanley, “Stochastic Feedback and the
Regulation of Biological Rhythms,” Europhys. Lett.
43, 363 (1998).
45. A. Réka, H. Jeong and A.-L. Barabási, “Error and
Attack Tolerance of Complex Networks,” Nature
406, 378 (2000).