Exp Brain Res (2006) 174: 256269

DOI 10.1007/s00221-006-0441-7

R ES E AR C H A RT I C L E

M. Roerdink M. De Haart A. Daertshofer

S. F. Donker A. C. H. Geurts P. J. Beek

Dynamical structure of center-of-pressure trajectories

in patients recovering from stroke

Received: 31 August 2005 / Accepted: 12 March 2006 / Published online: 10 May 2006

Springer-Verlag 2006

Abstract In a recent study, De Haart et al. (Arch Phys

Med Rehabil 85:886895, 2004) investigated the recovery
of balance in stroke patients using traditional analyses of
center-of-pressure (COP) trajectories to assess the eects
of health status, rehabilitation, and task conditions like
standing with eyes open or closed and standing while
performing a cognitive dual task. To unravel the underlying control processes, we reanalyzed these data in terms
of stochastic dynamics using more advanced analyses.
Dimensionality, local stability, regularity, and scaling
behavior of COP trajectories were determined and
compared with shued and phase-randomized surrogate
data. The presence of long-range correlations discarded
the possibility that the COP trajectories were purely
random. Compared to the healthy controls, the COP
M. Roerdink (&) A. Daertshofer S. F. Donker
A. C. H. Geurts P. J. Beek
Faculty of Human Movement Sciences,
Institute for Fundamental and Clinical
Human Movement Sciences,
Vrije Universiteit, Van der Boechorststraat 9,
1081 BT Amsterdam, The Netherlands
E-mail: [email protected]
Tel.: +31-20-5988516
Fax: +31-20-5988529
M. De Haart
Department of Rehabilitation, Amsterdam Medical Centre,
University of Amsterdam,
Amsterdam, The Netherlands
M. De Haart A. C. H. Geurts
Department of Research, Development, and Education,
St. Maartenskliniek,
Nijmegen, The Netherlands
S. F. Donker
Department of Otorhinolaryngology,
Vrije Universiteit Medical Centre,
Amsterdam, The Netherlands
A. C. H. Geurts
Department of Rehabilitation Medicine,
University Medical Centre, St. Radboud,
Nijmegen, The Netherlands

trajectories of the stroke patients were characterized by

increased dimensionality and instability, but greater
regularity in the frontal plane. These ndings were taken
to imply that the stroke patients actively (i.e., cognitively)
coped with the stroke-induced impairment of posture, as
reected in the increased regularity and decreased local
stability, by recruiting additional control processes (i.e.,
more degrees of freedom) and/or by tightening the
present control structure while releasing non-essential
degrees of freedom from postural control. In the course
of rehabilitation, dimensionality stayed fairly constant,
whereas local stability increased and regularity decreased. The progressively less regular COP trajectories
were interpreted to indicate a reduction of cognitive
involvement in postural control as recovery from stroke
progressed. Consistent with this interpretation, the dual
task condition resulted in less regular COP trajectories of
greater dimensionality, reecting a task-related decrease
of active, cognitive contributions to postural control. In
comparison with conventional posturography, our results show a clear surplus value of dynamical measures in
studying postural control.
Keywords Motor control Posture Non-linear
dynamics Stroke Rehabilitation

Introduction
In quiet standing, the position of the center of mass
varies continuously, resulting in changes in the forces
exerted by the human body on the support surface and in
the corresponding ground reaction forces. This postural
sway can be studied by using force platforms that measure the displacement of the application point of the
ground reaction force, that is, the center of pressure
(COP). The time-evolution of the resulting COP trajectories is often viewed, usually implicitly, as a manifestation of random uctuations in the postural control
system. This view underlies the application of conventional averaging techniques in posturography that aim to

257

identify scalar values, like the mean COP velocity, by

averaging out the assumed noisy or random character of
postural sway. Such descriptive statistical values have
been shown to change with pathologies and aging and to
vary over a range of sensory and cognitive conditions.
There are ample indications in the literature, however,
that more detailed analyses of COP time-evolutions may
provide further insight into postural control.
In this context, the nature of postural sway has been
characterized both as deterministic chaotic (e.g., Newell
et al. 1993; Yamada 1995; Pascolo et al. 2005) and stochastic (e.g., Collins and De Luca 1993; Newell et al.
1997; Delignie`res et al. 2003). Chaotic time series appear
random and unpredictable but arise from deterministic
non-linear processes, whereas stochastic time series are
governed by chance alone (e.g., Brownian motion, see
below) or by a combination of deterministic and random
processes (e.g., biased random walk). Although conceptually dierent, these approaches both focus on the
dynamical structure of COP trajectories, which may
contain information about the postural control exerted.
The smoothness of the COP trajectories hints at strong
deterministic components in the stochastic postural
sway dynamics and comparatively weak inuences of
noise (see also Collins and De Luca 1993, 1995; Riley
et al. 1999; Riley and Turvey 2002). Using linear systems
theory and corresponding identication techniques,
Kiemel et al. (2002) similarly concluded that COP trajectories reect a mixture of deterministic and stochastic
components. However, this more traditional approach
tends to downplay the importance of the repeatedly
demonstrated non-linear character of the COP dynamics. Hence, in order to identify the stochastic dynamics
of postural sway, COP analyses that account for the
non-linear and stochastic temporal evolution of postural
sway appear appropriate. While it has already been
demonstrated that analyses borrowed from the eld of
dynamical systems can be meaningfully applied to
measured COP trajectories (e.g., Newell et al. 1993;
Yamada 1995; Pascolo et al. 2005), we expected that
they would also be valuable in evaluating the eects of
health status, rehabilitation, and task manipulations in
clinical studies (cf. Raymakers et al. 2005).
From this expectation, we reanalyzed the COP data
that were collected and analyzed by De Haart et al.
(2004) in an encompassing longitudinal study of the
recovery of standing balance in stroke patients. Those
data are particularly interesting from a clinical point of
view because they cover over 30 patients whose postural
control was assessed at ve stages, during a 3 month
recovery period under both sensory (eyes-open versus
eyes-closed) and cognitive (standing with eyes open while
performing an arithmetic dual task) manipulations. With
conventional posturographic measures (e.g., mean position, sway amplitude, and sway velocity), De Haart et al.
(2004) found that, compared to age- and gender-matched
healthy elderly controls, stroke patients exhibited a severe weight-bearing asymmetry that was accompanied by
increased sway amplitude and sway velocity, especially in

the frontal plane. With follow-up assessments, the

weight-bearing asymmetry reduced, as did the sway
amplitude and velocity, suggesting that posture stabilized
in the course of rehabilitation. Under the arithmetic dual
task, COP velocity in the sagittal plane was greater,
which was interpreted as an indication that patients
stiened up to some degree, perhaps as a result of an
increased alertness. Finally, without vision, a larger COP
amplitude and velocity were observed.
We reanalyzed the data of De Haart et al. (2004) using
a set of complementary measures borrowed from
dynamical systems theory to evaluate our expectation
that they would reveal new aspects of the data that were
not disclosed by means of the conventional measures
mentioned above. Since a plethora of measures may be
used to analyze the temporal structure underlying COP
trajectories, we restricted the analysis to four: the correlation dimension, the largest Lyapunov exponent, the
sample entropy, and the Hurst exponent. This particular
choice was motivated from the consideration that these
measures are all dened operationally in terms of readily
interpretable and complementary (i.e., deterministic and
stochastic) properties of motor control, namely, number
of degrees of freedom, local stability, regularity, and
scaling behavior. More specically, the correlation
dimension (e.g., Grassberger and Procaccia 1983) measures the dimensionality of the COP time series (e.g.,
Newell 1998; Yamada 1995) and may provide an estimate of the number of (active) dynamical degrees of
freedom involved in postural control (e.g., Kay 1988).
Lyapunov exponents quantify the convergence or
divergence of nearby points in the postural control state
space, with the largest Lyapunov exponent characterizing a systems local stability (Wolf et al. 1985; Rosenstein
et al. 1993; Yamada 1995), that is, the sensitivity of the
postural control system to local perturbations. Sample
entropy (Richman and Moorman 2000) quanties the
regularity (or predictability) of a time series, while scaling
factors like the Hurst exponent [e.g., determined via
detrended uctuation analysis (DFA), Peng et al. 1994,
1995] quantify the extent to which a recorded COP time
series exhibits long-range correlations. To date, these
dynamical measures have rarely been used in clinical
studies of upright standing (see, e.g., Newell et al. 1993;
Pascolo et al. 2005, for exceptions), in spite of their
capacity to provide new insights, as widely demonstrated
in, for instance, physiology and the study of dynamical
diseases (e.g., Goldberger et al. 1996, 1997, 2002; Glass
2001; Lipsitz 2002; Kyriazis 2003).
With our choice of measures, we abstained from
using other techniques that have recently been applied in
the study of postural control, such as recurrence quantication analysis (RQA, Riley et al. 1999), rambling
trembling component analysis (Zatsiorsky and Duarte
1999), sway-density analysis (Baratto et al. 2002), and
stabilogram diusion analysis (Collins and De Luca
1993, 1995). Although the various output measures of
RQA are conceptually related to some of the measures
used in the present study, especially those indexing

258
Table 1 Characteristics of the 33 stroke patients at the baseline assessment
Age (years)
Time post-stroke (weeks)
Type of stroke (infarction/haematoma)
Hemisphere of stroke (left/right)
Functional Ambulation Categoriesa
Lower-limb motor selectivity (Brunnstrom stage)b

61.2 (SD 13.0; range 2782)

9.8 (SD 5.2; range 3.324.1)
26/7
10/23
2 (range 14)
IV (range IIVI)

a
Values are median scores (range). Functional Ambulation Categories are dened as follows: 0, Non-functional (unable): patient cannot
walk or requires help of two or more people; 1, Dependent (level 2): patient requires rm continuous support from one person who helps
carrying weight and with balance; 2, Dependent (level 1): patient needs continuous or intermittent support of one person to help with
balance and coordination; 3, Dependent (supervision): patient requires verbal supervision or standby help from one person without
physical contact; 4, Independent (on level ground): patient can walk independently on level ground, but requires help on stairs, slopes, or
uneven surfaces; 5, Independent: patient can walk independently anywhere (cf. Collen et al. 1990; Wade 1992)
b
Values are median scores (range). Brunnstrom stages are dened as follows: I, accid paralysis; II, increased muscle tone without active
movement; III, increased muscle tone with active movements mainly in rigid extension synergy; IV, increased muscle tone with alternating
gross movements in extension and exion synergies; V, muscle tone normalization with some degree of selective muscle control (i.e.,
combined active knee extension and foot dorsiexion against some resistance); and VI, normal muscle tone and control (cf. Brunnstrom
1966; Fugl-Meyer et al. 1975)

stability and regularity, they are not identical and

cannot be simply mapped onto each other. Unlike RQA,
rambling-trembling component analysis, sway-density
analysis, and stabilogram diusion analysis have been
constructed specically to examine postural sway
dynamics, but the aim of our study was to evaluate
whether generally motivated measures derived from
dynamical systems theory would have surplus value in
the analysis of COP trajectories compared to standard
posturographic measures.

open while performing an arithmetic dual task (DT),

and eyes closed (EC)] were performed1. In each condition, the participants were asked to stand as still and
symmetrically as possible (i.e., barefoot, arms alongside
the trunk (if possible), feet placed with the heels 8.4 cm
apart, with the toes pointing outward at a 9 angle from
the sagittal midline), while their COP was recorded for
30 s at a sample frequency of 60 Hz. The healthy elderly
underwent the same procedure (see Nienhuis et al. 2001).
We refer to the original study by De Haart et al. (2004)
for more detailed information about the participants,
equipment, and the procedures.

Methods
Posturographic data analysis
Participants and procedure
We reanalyzed the COP data from 33 stroke patients
(mean age 61.2 years, SD 13.0 years) out of an inception
cohort of 37 stroke patients (viz. data from four patients
were excluded due to missing values) and 22 healthy
elderly (mean age 63.9 years, SD 9.3 years) that was
used by De Haart et al. (2004).
In this study, stroke patients receiving standard
rehabilitation training were evaluated ve times over a
12 week period from the moment that each patient was
able to stand without assistance for at least 30 s [i.e.,
grade 4 according to the standing balance scale described by Bohannon (1995)], the so-called baseline
assessment, as well as 2, 4, 8, and 12 weeks after that
moment. Stroke type (infarction/haematoma), location
(left/right hemisphere), and the time post-stroke of the
33 included patients are provided in Table 1, together
with an indication of the level of recovery at the baseline
assessment, as quantied by the clinical evaluation of the
patients (1) walking skills, as rated according to the sixpoint (range 05) Functional Ambulation Categories (cf.
Collen et al. 1990; Wade 1992), and (2) lower-limb
motor selectivity, as scored according to the six motor
stages dened by Brunnstrom (cf. Brunnstrom 1966;
Fugl-Meyer et al. 1975).
During each assessment, two posturographic trials
for three quiet-standing conditions [eyes open (EO), eyes

Prior to all the analyses, each signals mean was subtracted. Time series were sampled by means of t ti, with
i=1, 2, 3, ..., N and N indicating the total number of
samples in the COP time series. The signals are denoted as
x(ti), with x=ML (i.e., medio-lateral)=AP (i.e., anterioposterior) COP displacements. We used the conventional
standard deviation rx to quantify the variability of the
postural sway dynamics x(ti). Besides the mean and variance of the time series, which, by denition, ignore the
temporal structure of the COP trajectories, we assessed
the COP dynamics by calculating its correlation dimension, largest Lyapunov exponent, sample entropy, and
scaling behavior. In the following, we explain how those

In the original study of De Haart et al. (2004), each balance

assessment consisted of two consecutive test series, incorporating
four quiet-standing tasks and one weight-shifting task, presented in a
xed sequence. This sequence was repeated in reverse order to control for time eects. Between the DT and EC condition, participants
conducted a trial while looking at a vertical black bar, which served
as a visual midline reference. The weight-shifting task was performed
twice after (and preceding) the rst (second) EC condition. A 1-min
rest was given after each balance test, whereas a longer pause was
allowed between the two test series. The arithmetic task in the DT
condition consisted of a (varying) verbal sequence of eight singledigit additions (e.g., 7+4=11 or 3+5=7) equally timed over the
30-s period. The participants were instructed to verbally indicate
the correctness of each summation by good or fault response.

259

1996), that is, the time at which the original and delayed
signals were maximally independent. As shown in Fig. 1,
the modied correlation sum Cm of the COP time series
x(ti) was nally computed for each embedding dimension

Dimension Analysis

10

C (r )
m b

original
shuffled
1

10

Cm r

m
2

10

C (r )
m

10

2d + 1

10

10

10

original
shuffled

10

10

10

2d +1=m
m

5
0

W X
N





1 NX
H r

X i
X j

Np i1 jiW

in which m refers to the embedding dimension and Hz

denotes the unit step function: Hz =1 if z0, Hz =0
otherwise. Xi is given as [x(ti), x(ti+s), ... ,
x(ti+(m
1)s)], |Xi
Xj| is the Euclidean distance between Xi and Xj, and r is a distance on a log scale. The
cut-o parameter3 W was dened as twice the rst
minimum in the mutual information function. Np is the
number of pairs i, j such that |i-j|W, where Np=(NW+1)(N-W)/2 (Theiler and Rapp 1996). Importantly,
for small distances r, the function Cm(r) behaves as a
power law, i.e.,
Cm r / rD2

Fig. 1 Dimensionality analysis. Top panel: correlation sum Cm

plotted for each embedding dimension m for original (thick black
lines) and shued surrogate (thinner gray lines) COP time series
(ML, EO, stroke patient). The dotted horizontal lines indicate the
boundaries between which the slopes dm were estimated (Kaplan
et al. 1991; see text for further details). Lower panel: a plot of the
slopes dm for the original (black solid circles) and shued surrogate
(gray open circles) correlation sums for increasing embedding
dimension m. Note that at the sixth embedding dimension, the
slope of the original COP time series satises the condition
m>2dm+1 (i.e., the corresponding dimension estimate D2 is 2.26),
whereas for the shued surrogate data, Cm(r) scales with rm

where D2 is the correlation dimension that is approximated via the linear region of the slope of the loglog
display of Cm(r) as a function of r. Notice that for very
small r, estimating D2 becomes dicult because usually
the number of data points (samples) becomes very small,
yielding inaccuracies similar to the case in which r becomes comparable to the attractor size. Hence, the slope
was determined for an interval between ra, i.e., the distance r capturing 0.5% of the pairs of points
(Cm(ra)=0.005) and rb, i.e., the distance r capturing 75%
of the pairs of points (Cm(rb)=0.75) (see also Fig. 1),
upper panel4. The slopes were determined per embedding dimension m according to:

calculations were performed in such a manner that the

analyses can be replicated if desired.

dm

10 11 12

log10 Cm rb
log10 Cm ra
log10 rb
log10 ra

Correlation dimension
To calculate the correlation dimension of the time series,
we followed Grassberger and Procaccia (1983) by conducting a phase space reconstruction of the COP
dynamics (Takens 1981). To this end, the measured COP
time series was embedded in a high-dimensional phase
space by time-delaying the original time series as x(ti),
x(ti+s), x(ti+2s), ... (see, for instance, Abarbanel 1996;
Packard et al. 1980). The embedding delay s corresponded to the rst minimum of the mutual information
function2 (Fraser and Swinney 1986; see also Abarbanel
2
For our data, the rst minimum of the mutual information occurred
at 11 (mean 11.14, SE 0.15) and 10 (mean 9.91, SE 0.20) data samples
for ML and AP sway components of the stroke patients, respectively.
These minima did not change with rehabilitation or condition
(P>0.05). For the healthy controls, the rst minimum of the mutual
information occurred at 11 data samples for both ML (mean 11.10,
SE 0.27) and AP (mean 10.54, SE 0.21) sway components. Again, no
change with condition was found (P>0.05). Note that the choice of
the time delay s was not based on these group averages but was
determined independently for each trial.

3
Correlations between consecutively sampled points can produce
spurious indications of low-dimensional structure. With the introduction of the cut-o parameter W>1, it is possible to minimize
these correlations (Grassberger 1986; Theiler 1986). Therefore, all
pairs of points that are closer together in time than some cut-o W
were excluded. W=1 returns the standard Grassberger and Procaccia (1983) formula.
4
The dimension is often calculated by looking at the slope of the
most linear segments of Cm(r), requiring a means of evaluating a
score for each plausible linear segment (i.e., based on the length of
the segment or the goodness of t to a line). The optimal linear
segment is chosen. In this way, these techniques emphasize the
possible existence of strange attractors. A drawback of such
methods is that the length scale chosen can depend discontinuously
on the underlying signal, because a small change in the signal can
change the relative ranks of the candidate linear segments and
thereby change the calculated dimension substantially (Kaplan
et al. 1991). Because the applied dimension analysis in this study
did not involve examination of the linear scaling of Cm(r), it would
be incorrect to interpret the estimated dimension D2 as the
dimension of the attractor. Similarly, it would be incorrect to infer
from this analysis that an attractor must exist.

260

We estimated the dimension D2 by looking for cases in

which the slopes dm saturated with increasing embedding
dimension m, with m>2dm+1 (Fig. 1, lower panel).
Finally, whenever m>2dm+1 was fullled, the estimate
D2 could be considered reliable.
Largest Lyapunov exponent
We quantied the mean exponential divergence d(t) at
time t of initially close state-space trajectories by means
of dt / Cekmax t (e.g., Rosenstein et al. 1993). In this
form, the exponent kmax reects a systems local stability, that is, its sensitivity to small and local perturbations
and is referred to as the largest Lyapunov exponent.
Briey, if kmax is negative, then any perturbation will
exponentially damp out and initially close trajectories
will stay close. In contrast, if kmax is positive, the d(t)s
will diverge, i.e., the distance between trajectories will
increase exponentially. Positive kmax values indicate the
presence of chaos, provided that the system stays in nite vicinity, implying that some attractor exists. In the
latter case, nearby points diverge exponentially (local
instability), causing a lack of predictability.
To calculate kmax, for each Xi of embedding dimension m (m>2dm+1), the nearest neighbor was identied
as the point closest to Xi with a temporal separation
larger than twice the rst minimum in the mutual
information function (see above). Next, distances between neighboring trajectories in state space were calculated as a function of time (i.e., jDt=3 s) and averaged
over all original pairs of nearest neighbors i. The kmax
were nally estimated from the slopes of
yj

N
1 1X
ln di j
Dt N i1

after tting a range from jDt=0 to 0.75 s (Rosenstein

et al. 1993).
Sample entropy
In addition to the aforementioned measures, we used
more statistical descriptions to characterize not only
the deterministic structure of the COP dynamics, but
also its stochastic features. In general, entropy- or
information-related measures are rst candidates when
analyzing a systems regularity or ordering. Since we
investigated time series rather than arbitrary statistical
ensembles, we used the sample entropy (Richman and
Moorman 2000), which builds on the conditional
probability that a signal of length N will repeat itself
for M points, provided that it already repeated itself
for M-1 points (within a tolerance range r and without
allowing self-matches): the higher the entropy, the
lower the time series regularity.
In view of the limited length of the to-be-analyzed
time series, we chose the window parameter as
M=3note that the choice of r is limited since too

small a tolerance yields low condence, while the discriminative capacity drops when increasing the tolerance
range (Pincus 1991; Pincus and Goldberger 1994). Recently, Lake et al. (2002) proposed to use the maximal
value of the relative errors of sample entropy and conditional probability for a range of values of r and M to
optimize the choice of r and M while enhancing the
eciency of the entropy estimate by penalizing conditional probabilities near 0 and near 1. Here, the sample
entropy analysis parameters r and M were selected based
on minima in the gray-scaled relative error map
(Fig. 2)cf. Lake et al. (2002) for further details.5
Scaling: detrended uctuation analysis
The Hurst exponent H quanties the leading order of
the temporal change of a time series correlation (or least
squared displacements). Formally one may write
r2Dx Dt / Dt2H   

i.e., the variance r of the displacement Dx changes in

time Dt according to a power law (with exponent 2H).
For the simplest diusion process, the free Brownian
motion or ordinary random walk, the Hurst exponent is
H=0.5. Other values in the range 0<H<1 are typically
referred to as fractional Brownian motion (Mandelbrot
and van Ness 1968): H>0.5 implies persistence, i.e., the
trajectory tends to continue in its current direction and
thus produces enhanced diusion; H<0.5 implies antipersistence, i.e., the trajectory tends to return from
where it came and thus suppresses diusion.
Importantly, COP trajectories are always bounded
within the area of support, whereas (fractional)
Brownian motion is, in principle, unbounded. Indeed,
Eq. (5) does not hold for diusion processes that remain
limited in that the corresponding variance saturates beyond a critical time interval (see Delignie`res et al. 2003).
Thus, in order to estimate the Hurst exponent using Eq.
(5), some preprocessing needs to be applied. The basic
idea is that if the original signal is bounded, then the
integrated time series is unbounded and exhibits quantiable scaling properties (Delignie`res et al. 2003; see
also Eke et al. 2000, 2002). This integration is the rst
step in the so-called DFA (Peng et al. 1994, 1995), in
which, subsequently, this integrated time series is divided
into non-overlapping intervals. Within each interval, the
time series is linearly detrended to remove trivial correlations (Fig. 3, center panel) and, nally, the root mean
square uctuations Fn of the residual are displayed as
function of the interval size n (Fig. 3, lower panel). In
the presence of a power law, the according loglog
representation yields a scaling exponent a that relates to
the Hurst exponent as H=(2a1)/4.
5
In agreement with, e.g., Lake et al. (2002) and Richman and
Moorman (2000), time series were normalized to unit variance.
Sample entropy software was obtained from PhysioNet (Goldberger et al. 2000).

261
0.01
1

Detrended Fluctuation Analysis

10

0.03

0.04

ML

0.02

0.05

M 4

10

0.06
5

0.07

0.08

10

20

time

original
shuffled

0.2

0.3

0.4

0.5

0.6

0.7

0.1

Fig. 2 Visual selection of the optimal tolerance r for estimating the

sample entropy of 66 randomly chosen, normalized ML COP time
series (i.e., rML=1) of healthy adults, based on the median value of
the relative error. As can be seen from this gray-scaled map, a
window length of M=1 allows a condent estimation of entropy
for a wide range of r. For M2, however, optimum values of r are
clearly evident and lie generally between 0.02 and 0.1. The relative
error rises very steeply for r<0.02 and for larger r values, stressing
the fact that many combinations of M and r are intolerable. Based
on inspection of the gray-scaled visual relative error plot at the
M=3 trace, r=0.035 was the optimal choice for the sample
entropy parameter r (lowest relative error). Similarly, r=0.035 and
r=0.045 were selected as optimal parameter values for ML and AP
sample entropy estimates, respectively. These r values were the
same for the stroke patients and the healthy controls

10

20

time

30

= 1.36
= 0.50
2

log10 Fn

0.1

0.09

ML

1000

30

Detrended uctuation analysis was applied to all x(ti)

time series using logarithmically spaced interval lengths
n from 10 to N/2 samples6 (N=1,800). The slope of the
double-logarithmic plot of Fn versus n was estimated
between n=0.26 s and n=8.66 s. The largest values of n
were disregarded, since these points were based on only
a few data segments, rendering the estimate unreliable.
Surrogate data
In order to guarantee the validity of the non-linear
analyses applied here, we exploited surrogate data by
means of both time- and phase-randomized COP trajectories (Theiler et al. 1992). We rst generated surrogate data by randomly selecting samples of x(ti), i.e., we
preserved the datas statistics (mean, variance, etc.), but
destroyed the datas temporal correlation by shuing
its temporal ordering (cf. Fig. 3, upper panel). Notice
that, due to the absence of temporal correlations, shufed surrogate data have a huge dimension (converging
to innity), a high sample entropy, and a Hurst exponent close to zero. In addition, we randomized the datas
phase after Fourier transformation. In general,
6
Notice that a multivariate extension of the detrended uctuation
analysis algorithm yields identical results when applied to the
embedded time series (see above) since we assumed stationarity.

log

1
10

Fig. 3 Detrended uctuation analysis. Upper panel: ML COP time

series xML of a stroke patient with eyes open (black line) and its
shued surrogate counterpart (gray line). Middle panel: integrated
time series XML. The vertical dashed lines indicate an interval of
length n=5 s, the dotted straight line segments represent the trend
estimated in each interval by a least-squares t. Lower panel:
double-logarithmic plot of uctuation Fn versus interval length n.
The slope a relates to the scaling exponent H according to
H=(2a
1)/4. Note that for the shued data H=0, indicating an
uncorrelated random process, whereas for the COP time series
H=0.43, indicating long-range correlations

randomizing the phase does not alter the spectral power

distribution and thus preserves the datas auto-correlation function (cf. Kantz and Schreiber 2004). Hence,
Hurst exponents of phase-randomized and original data
should match, while, as for the shued surrogates,
phase-randomization is expected to result in high correlation dimension and sample entropy.
Statistics
All the dependent variables for the stroke patients were
analyzed using a repeated measures analysis of variance
(ANOVA) with within-subject factors rehabilitation (ve

262
1.0

6.0
4.0
2.0
0

0 2 4

12

follow-up [weeks]

ns

P < 0.05
P < 0.05

0.8

ML max

ns ns

ML sample entropy

ML

[mm]

8.0

0.6
0.4

ns

2.0

1.0

0.2
0

0 2 4

12

follow-up [weeks]

0 2 4

12

follow-up [weeks]

Fig. 4 Rehabilitation eects for rML, ML sample entropy and ML

kmax values. Error bars denote the 95% condence intervals for
dierence after Bonferroni adjustment for multiple comparisons.
rML and kmax decreased signicantly with follow-up assessments

[except when explicitly stated so (ns)]. Sample entropy was

signicantly higher after 8 and 12 weeks when compared to
baseline (0 weeks)

follow-up levels: baseline (week 0) and 2, 4, 8, and

12 weeks later) and condition (three levels: EO, DT, and
EC). In the case of signicant main or interaction eects,
post-hoc analysis was performed by means of t-tests
using the standard Bonferroni correction to maintain the
family-wise error rate at 5%. The eect of health status
was examined by means of a repeated measures ANOVA with health status (two levels: stroke patients (averaged over follow-up assessments) and healthy controls)
as between-subject factor and condition as within-subject
factor. Post-hoc t-testing with Bonferroni correction was
performed for signicant eects of condition, whereas for
signicant interaction eects, separate repeated measures ANOVAs with condition as within-subject factor
were performed for the two groups. For main and
interaction eects of rehabilitation, condition, and health
status, Cohens f was used as a measure of eect size;
large eect sizes were operationalized by convention as
f>0.4 (Cohen 1988). To quantify the eects of randomization, we used a repeated measures ANOVA with
randomization as within-subject factor (e.g., three levels:
original COP data, shued surrogates, and phase-randomized surrogates), for stroke patients and healthy
elderly, separately.

Scaling exponents7 (H), dimension estimates (D2), and

sample entropy in the AP direction were not aected
signicantly by rehabilitation (Table 2).

Results
The results (F, P, and f values) of the rehabilitation by
condition ANOVA and the health status by condition
ANOVA for the dependent COP variables are presented
in Tables 2 and 3, respectively.
Rehabilitation
With follow-up assessments, COP variability (rAP and
rML) decreased signicantly and local stability improved as evidenced by a signicantly reduced kmax in
both ML and AP directions. In addition, COP regularity decreased in the ML direction as indexed by a
signicantly increased sample entropy (Fig. 4; Table 2).

Condition
Figure 5 displays condition eects for stroke patients (see
also Table 2). The COP variability was signicantly
greater for EC as compared to both EO and DT. The
COP regularity was clearly lower (signicantly higher
sample entropy) for DT than for EO and EC. The ML
scaling exponents (H) were signicantly lower for EC
than for EO and DT. In contrast, AP scaling exponents
were signicantly higher for EO than for the two other
conditions. The estimated dimension was signicantly
larger for DT than for EO. The largest Lyapunov
exponent kmax was signicantly larger for EC (i.e., locally less stable) than for EO and DT. No signicant
rehabilitation condition interaction was observed for
either of the COP measures. For the healthy elderly,
qualitatively similar condition eects were observed.
Health status
Stroke patients diered signicantly from healthy elderly
in all COP measures except for the ML scaling exponents (see also Table 3 and Fig. 6)8. COP variability
7
To avoid false or spurious conclusions, Hurst exponents were also
determined by means of a rescaled range analysis (Hurst 1965;
Rangarajan and Ding 2000; Delignie`res et al. 2003; cf. Wing et al.
2004 for a related power spectral approach), yielding slightly higher
estimates of the diusion process than the DFA. To compare these
two methods, the pair-wise two-tailed Pearson correlation coecient between the scaling exponent based on the rescaled range
analysis (H HR/S) and the detrended uctuation analysis
(H HDFA) was determined for all the trials of the stroke patients
(N=990). For both the AP and ML scaling estimates, the correlation analysis showed a good agreement between HR/S and HDFA
(r=0.918, P<0.01 and r=0.895, P<0.01, respectively).
8
The signicant results reported in Table 3 were all preserved when
the averaged post-stroke values were replaced by the earliest poststroke values.

263
Table 2 Main and interaction eects of rehabilitation (ve levels) and condition (three levels) on standard deviation r, Hurst exponent H,
dimension estimate D2, sample entropy, and largest Lyapunov exponent kmax for ML and AP COP components of 33 stroke patients
Rehabilitation
F a(4,

128)

Standard deviation r
ML
24.29
AP
11.97
Hurst exponent H
ML
0.65
AP
0.89
Dimension estimate D2
ML
0.32
AP
0.36
Sample entropy
ML
5.70
AP
0.66
Lyapunov exponent kmax
ML
24.35
AP
26.95

Interaction

Condition
P

F a(2,

F a(8,

<0.001
<0.001

0.87
0.61

5.80
15.93

<0.01
<0.001

0.43
0.71

ns
ns

0.14
0.17

8.88
8.94

<0.001
<0.001

ns
ns

0.10
0.21

5.35
3.96

<0.005
ns

0.42
0.14

<0.001
<0.001

0.87
1.02

1.67
1.42

ns
ns

0.23
0.21

0.68
0.53

1.57
0.55

ns
ns

0.22
0.13

<0.01
<0.05

0.41
0.39

0.73
1.17

ns
ns

0.15
0.06

4.76
17.75

<0.05
<0.001

0.38
0.75

1.14
1.31

ns
ns

0.19
0.20

8.30
5.94

<0.005
<0.01

0.51
0.48

0.95
0.66

ns
ns

0.17
0.16

64)

256)

ns not signicant
a
In case the sphericity assumption was violated, the number of degrees of freedom was adjusted using the HuynhFeldt method. Missing
values arose for the dimension estimate D2 and the Lyapunov exponent kmax due to the m>2dm+1 criterion for six stroke patients for the
AP COP component only (nine missing values in total). No missing value analysis was performed, resulting in a loss of six patients in the
rehabilitation condition repeated measures ANOVA

was signicantly larger for stroke patients than for the

healthy controls. COP regularity in the ML direction was
higher (signicantly lower sample entropy) for stroke
patients, whereas COP regularity in the AP direction was
signicantly higher for the healthy controls. For the
stroke patients, the scaling exponents for AP were signicantly lower, while the estimated dimension and kmax
were signicantly higher in both directions when compared to the healthy controls. There were two signicant
health status condition interactions (Table 3), but these
could not be interpreted as none of the post-hoc comparisons was signicant.

Randomization
As expected, all Hurst exponents vanished for the
shued surrogates (see above and Fig. 3, lower panel)
and H did not dier from the original value after phaserandomization (P>0.05; see Fig. 7) for both the stroke
patients and the healthy elderly. This implies that strong
correlations were present in the time evolution of the
COP trajectories, suggesting a deterministic dynamical
structure. Furthermore, ML and AP sample entropy
estimates for the shued and phase-randomized surrogate time series were both signicantly higher than their
original counterparts for the stroke patients and the
healthy elderly alike (Fig. 7). Higher entropy values
imply that more information is required to describe
the surrogate data due to the applied time- and phaserandomization, which again suggests a deterministic
component in the original COP trajectories.
The dimension analysis revealed that Cm(r) of shufed surrogates scaled with rm rather than rD2 (see Fig. 1,

lower panel), reecting the presence of very high (or

innite) dimensional noise in the surrogate data that was
absent in the original data. Irrespective of the health
status and movement direction, the D2 and kmax of
phase-randomized surrogates were signicantly higher
than their original counterparts (Fig. 7). In line with the
previous results, this again indicates the presence of a
pronounced deterministic component.

Discussion
In the present study, we reanalyzed the COP data of De
Haart et al. (2004) using dynamical rather than conventional measures to examine whether this would lead
to novel insights into the changes in postural control as a
function of (a) health status (stroke patients versus
healthy elderly), (b) rehabilitation (follow-up assessments), and (c) task conditions (EO, DT, EC). In the
following, we systematically discuss to what extent new
ndings and insights were obtained vis-a`-vis those reported by De Haart et al. (2004), using the three aforementioned independent variables as entry points. We
then conclude with a general evaluation of the signicance of studying postural sway dynamics.
Health status
As expected, we replicated the nding of De Haart et al.
(2004) that stroke induced substantial dierences in the
global characteristics of the COP trajectories in comparison with the healthy controls (see also De Haart
et al. 2004). Postural sway variability (i.e., deviation
from the mean) was larger in patients than in healthy

264
EO
DT
EC

Condition
*

8.0

6.0
4.0
2.0

ML
sample entropy

1.0

AP

0.5

ML

AP

0.4

0.2

ML

AP

3.5

D2

2.5

*
*

1.5
0.5

ML

AP

max

2.0
1.0

ML

AP

Fig. 5 Condition eects of ML and AP COP variability r, sample

entropy, Hurst exponent H, dimension estimate D2, and largest
Lyapunov exponent kmax. Eyes open (EO: white bars), dual task
(DT: gray bars), and eyes closed (EC: black bars) conditions for
stroke patients (healthy adults showed qualitatively similar results)
are depicted. Error bars denote the 95% condence intervals. Stars
placed at an error bar of a specic condition denote a signicant
dierence from the other two conditions, whereas stars placed
above connecting lines represent signicant dierences between the
two connected conditions (see Table 2 for F, P, and f values for
condition eects)

elderly. Analyses of surrogate data revealed, however,

that the observed increase in variability was not due to
an increase in noise. Unlike their shued and phaserandomized counterparts, the original data exhibited
dynamical features other than pure randomness (see
Fig. 7), suggesting the presence of a correlated
dynamical structure in the recorded COP trajectories.
Importantly, the structure (i.e., scaling behavior) of the
correlations in the COP trajectories was preserved
qualitatively in the stroke patients despite their
impairment. That is, stroke did not lead to a breakdown
of long-range correlations, as has been found for other

pathologies (e.g., Goldberger et al. 1996, 1997, 2002).

Our scaling analysis revealed Hurst exponents that
varied between white noise and Brownian motion (i.e.,
0<H<0.5), indicating that the COP trajectories of both
the healthy elderly and the stroke patients exhibited an
anti-persistent behavior: a decreasing trend is followed
by an increasing future trend (cf. Delignie`res et al. 2003,
see also Collins and De Luca 1993; Frank et al. 2001).
By denition, those results regarding the correlated
dynamical structure in the recorded COP trajectories
could only be obtained by applying dynamical methods
of data analysis.
Marked dierences between patients and healthy
controls were found in the three other dynamical
measures (Fig. 6). Dimensionality was increased in AP
direction due to stroke (Fig. 6). This increased dimensionality may simply be interpreted as a result of increased noise, which would be consistent with the
increased dimensionality and sample entropy after
phase-randomization and shuing (Fig. 7). However,
the increased dimensionality may also be interpreted as
a change in postural control. One option is to interpret
the D2 values literally in terms of chaos theory, which
would imply that the data were chaotic because D2
always saturated at a non-integer value, and that the
dimension of the chaotic COP attractor was greater for
the patients than for the healthy controls. However,
claims about the presence of deterministic chaos have
to be made with great caution because distinguishing
between chaos and stochasticity on the basis of nite
datasets might be dicult. Hence, we focused on the
dierences between groups and across conditions rather
than the possibly chaotic nature of the data. Accordingly, the dimensionality ndings may be interpreted to
imply that the stroke patients recruited additional
control processes (degrees of freedom), for instance to
compensate for the reduced ecacy of ankle mechanisms on their paretic side (De Haart et al. 2004; Geurts et al. 2005). Alternatively, it could also be the case
that the already present control structure (dened over
essential degrees of freedom) was tightened so that
(non-essential) degrees of freedom were released from
control, resulting in greater dimensionality due to
greater expression of noise along uncontrolled dimensions (i.e., the notion of uncontrolled manifold; cf.
Schoner 1995; Scholz and Schoner 1999). Unfortunately, at the level of the COP variable, i.e., an output
variable integrating many subsystems, it seems hard, if
not impossible, to unambiguously relate the observed
increased dimensionality to an increased noise level (be
it directly or indirectly via an increasing number of
released (or uncontrolled) degrees of freedom) versus
an increased number of recruited control processes. As
will become apparent in the following, however, the
stability analysis and the regularity analysis provided
clues to tentatively resolve this impasse.
The largest Lyapunov exponents (kmax) of the COP
trajectories were signicantly greater for the stroke
patients than the healthy elderly, demonstrating a

265
Table 3 Main and interaction eects of health status (between-subject factor: two levels) and condition (within-subject factor: three levels)
on standard deviation r, Hurst exponent H, dimension estimate D2, sample entropy, and largest Lyapunov exponent kmax for ML and AP
COP components of 33 stroke patients and 22 healthy controls
Interaction

Health status

Condition

F a(1, 53)

a
F(2,

Standard deviation r
ML
63.35
AP
20.62
Hurst exponent H
ML
0.02
AP
15.11
Dimension estimate D2
ML
3.67
AP
24.69
Sample entropy
ML
7.38
AP
9.60
Lyapunov exponent kmax
ML
126.34
AP
37.23

F a(2, 106)

<0.001
<0.001

1.09
0.62

4.40
23.98

<0.05
<0.001

0.29
0.67

ns
<0.001

0.02
0.53

8.27
12.63

<0.001
<0.001

=0.061
<0.001

0.26
0.68

5.90
8.68

<0.01
<0.005

0.37
0.43

<0.001
<0.001

2.77
0.83

3.34
0.44

<0.05
ns

0.25
0.09

0.40
0.49

0.03
1.04

ns
ns

0.02
0.14

<0.005
<0.001

0.33
0.41

1.13
0.51

ns
ns

0.15
0.10

0.15
7.88

ns
<0.005

0.05
0.38

1.37
0.40

ns
ns

0.16
0.08

14.36
30.53

<0.001
<0.001

0.52
0.76

0.29
5.89

ns
<0.005

0.07
0.33

106)

ns not signicant
In case the sphericity assumption was violated, the number of degrees of freedom was adjusted using the HuynhFeldt method. Due to
the averaging of conditions with rehabilitation, no missing values for the factor health status were present (i.e., 33 stroke patients versus 22
healthy adults)
a

decreased local stability or a deteriorated neuromuscular

control. Interestingly, Buzzi et al. (2003) found similar
dynamical signatures (i.e., larger dimensionality and
decreased local stability) in the variability of joint
kinematics of gait with aging, which they attributed to
deciencies in the ability to actively control joint motion. Elderly walkers were unable to compensate for the
natural stride-to-stride variations, which could increase
the risk of falling (Buzzi et al. 2003). Similarly, stroke
patients may have a reduced ability to compensate for
small (internal and/or external) perturbations, forcing
them to actively adjust postural control either by
recruiting more degrees of freedom or by tightening the
control of essential variables while releasing non-essential degrees of freedom. Thus, in all likelihood, the observed increase in dimensionality has functional
signicance and is not simply a reection of an increased
noise level.
Further evidence for this interpretation was found in
the analysis of the regularity of the COP trajectories. In
the stroke patients, the postural sway was more regular
in the frontal plane (i.e., lower sample entropy), whereas
it was more regular in the sagittal plane in the healthy
controls (Fig. 6). Furthermore, the stroke patients
showed greater sway variability in the frontal plane,
whereas the healthy controls showed greater sway variability in the sagittal plane (Fig. 6). Hence, the direction
with the largest sway variability also showed the greatest
regularity in the COP movements, suggesting that postural sway was more tightly controlled along this
direction. Notice that the studied rehabilitation cohort
was characterized by severe impairments in frontal plane
balance (De Haart et al. 2004; see also Paillex and So
2005). Furthermore, Brown et al. (2002) provided

evidence for increased attention demands for quiet

standing in stroke patients. Possibly, therefore, stroke
patients actively (i.e., by means of increased cognitive
control) compensated for the loss of accurate sensory
information from the paretic leg and for other strokemediated impairments hampering balance control. The
observed increased regularity of the medio-lateral COP
trajectories in the stroke patients could therefore reect
an elevated cognitive contribution to postural control
(see also below). Moreover, this nding excludes the
possibility that the accompanying increased dimensionality is an expression of increased noise, since that would
have produced higher rather than lower sample entropy,
as was the case for the shued and phase-randomized
surrogates (Fig. 7). All in all, it is fair to conclude at this
stage of the discussion that using dynamical measures
had surplus value in assessing the eects of health status
on postural control, although not all of those eects
could be interpreted yet in a conclusive manner.
Rehabilitation
With follow-up assessments the COP variability decreased signicantly (Fig. 4), as did the corresponding
regularity in the frontal plane (i.e., ML sample entropy
increased, Fig. 4). Initially, the postural sway was very
large and fairly regular, whereas 3 months later it was
smaller and markedly less regular, which is in partial
agreement with the ndings and theoretical perspective
of Goldberger et al. (1996, 1997, 2002). The long-range
correlations (i.e., scaling exponents) in the COP trajectories, however, did not change over the ve follow-up
assessments.

266
patients
controls

8.0

6.0

Health Status
*

4.0
2.0

sample entropy

ML
1.0

AP

*
0.5

ML

AP

ns

0.4
0.2
0

ML

3.5
P = 0.061

AP

D2

2.5
1.5
0.5

ML

AP

max

2.0
1.0

ML

AP

Fig. 6 Health status eects of ML and AP COP variability r,

sample entropy, Hurst exponent H, dimension estimate D2, and
largest Lyapunov exponent kmax. The eect of health status is
visualized by comparing stroke patients (black bars) and healthy
controls (white bars). Error bars denote the 95% condence
intervals. Stars placed at the error bars denote signicant
dierences between patients and controls (see Table 3 for F, P,
and f values for health status eects)

The fact that the dimensionality of the COP trajectories was consistently higher for the stroke patients
than for the healthy controls suggests that the reported
dierences in postural control persisted in time. However, in the course of rehabilitation, postural control
improved in the stroke patients, as evidenced by increased local stability and decreased regularity. Both
these novel ndings are theoretically signicant.
The observed increase in local stability is important
in view of the suggestion of De Haart et al. (2004) that
posture stabilized in the course of rehabilitation. By
assessing postural stability directly by means of the
largest Lyapunov exponent rather than by assuming that
postural stability is inversely related to postural sway
variability (which is not necessarily valid, cf. Newell
et al. 1993), we demonstrated that postural stability
increased in the course of rehabilitation. Thus, by

reanalyzing the data using dynamical measures the

proposition of De Haart et al. (2004) was conrmed
empirically.
The observed decreased regularity with follow-up
assessments is important in that it may reect a reduction in the cognitive component directed to postural
control in the course of rehabilitation, possibly due to
improved multi-sensory integration and progressive
internalization of altered body dynamics (see also Geurts et al. 2005). It could be speculated that, at the
beginning of independent standing, stroke patients actively (i.e., cognitively) recruit additional control processes (e.g., additional strategies, co-contraction), while
in the course of rehabilitation they learn to more automatically exploit the subspaces of controlled and
uncontrolled variables. This possible trade-o between
increased number of control processes and increased
number of uncontrolled degrees of freedom can leave the
observed dimensionality unchanged, despite marked
changes in postural control. It should be noted in this
context that after 3 months of follow-up, the frontal
plane balance was still more regular and less stable in the
stroke patients than in the healthy controls (Fig. 6),
indicating that, congruent with the proposed relation
between COP regularity and the cognitive contribution
to postural control, the cognitive involvement in posture
was still slightly elevated in the stroke patients.
Thus, in assessing the eects of rehabilitation, using
dynamical measures was clearly benecial in that they
allowed for a conrmation of a tentative interpretation
regarding the eect of recovery on postural stability, as
well as a novel, albeit admittedly speculative, interpretation of the relationship between COP regularity and
the cognitive regulation of posture. The cognitive dual
task manipulation may provide a means to further assess
this novel interpretation.
Task conditions
Without visual information, COP variability and kmax
increased signicantly, while no dierence in dimensionality was observed with or without vision. These
ndings agree with the study of Megrot et al. (2002) on
center-of-mass trajectories when standing on an unstable
platform with eyes open and closed. In addition, sample
entropy values did not dier with and without vision
(Fig. 5), indicating that there was no change in postural
sway regularity as a function of the availability of visual
information. The scaling behavior indicated that successive data points were more negatively correlated with
eyes closed than with eyes open. The corresponding increase in COP variability with eyes closed might have
brought the postural control system close to its stability
limits, which might have amplied the negative serial
correlation between points in the COP time series in
order to stay upright. It seems likely that other sensory
systems (such as the vestibulum and muscle and joint
receptors) may be facilitated to a greater extent in order

267
original
phase-randomized
shuffled

patients controls

1
0
ML AP

ML AP

patients controls

patients controls

2.5

max

0.5
0.4
0.3
0.2
0.1
0

D2

sample entropy

patients controls
4

1.5
0.5

0
ML AP

ML AP

ML AP

ML AP

ML AP

ML AP

Fig. 7 Randomization eects for sample entropy, Hurst exponent

H, dimension estimate D2, and largest Lyapunov exponent kmax
depicted for both the stroke patients and the controls for ML and
AP COP directions. Original COP data (black bars), phaserandomized (gray bars) and shued (white bars) surrogate data
are depicted. Note that, apart from only H for original and phase-

randomized surrogate data (P>0.05), original, phase-randomized

and shued surrogates all diered signicantly from each other
(P<0.005). Note that all Hurst exponents vanished for shued
surrogates, i.e., H=0, whereas for D2 and kmax no shued
surrogates values could be determined due to the m>2dm+1
criterion (cf. Fig. 1, Cm(r) scales with rm)

to compensate for the visual deprivation. The fact that

with eyes closed local stability strongly decreased indicates that vision plays a crucial role in modulating
postural dynamics (e.g., Nienhuis et al. 2001; Megrot
et al. 2002; Geurts et al. 2005).
When the cognitive involvement in postural control
was diminished by introducing an arithmetic dual task,
the COP trajectories became less regular (sample entropy increased, see Fig. 5), which supports the relationship between postural sway regularity and cognitive
contributions to postural control as proposed in the
preceding: larger (smaller) cognitive involvement yields
more (less) regular COP trajectories. De Haart et al.
(2004) reported larger sway velocity related to a shift in
COP median frequency in the DT condition without
accompanying changes in the COPs spectral power (or
variability). We observed increased dimensionality in the
COP trajectories accompanied by less regularity when
performing a dual task as opposed to standing with eyes
open. It could be that the dual task led to a greater
contribution or expression of noise to the COP dynamics
(i.e., either direct or indirect). An increased noise level is
traditionally viewed as detrimental for control, but the
generality of this interpretation has recently been refuted
by demonstrations of the benecial eects of noise in
sensori-motor control (e.g., Collins 1999; Cabrera et al.
2004; see also Wiesenfeld and Moss 1995; Thurner et al.
2002; Collins et al. 2003). Interestingly, under the DT
condition, local stability did not dier from that in the
EO condition, which suggests that the postural control
system could have exploited noise in a like fashion.
In assessing the eects of sensory manipulations (i.e.,
EO versus EC conditions), the use of dynamical measures only yielded marginally novel ndings: with
increasing sensory diculty stability diminished, as is
well-known from previous studies using linear systems
measures (e.g., Peterka 2002). In contrast, in assessing
the eects of a cognitive dual task on postural control,
the use of dynamical measures had surplus value as it led
to the discovery of eects that were absent when using
conventional posturographic measures. Moreover, those
eects were fully consistent with the interpretation of the

eects of health status and rehabilitation on postural

control provided in the preceding: decreased cognitive
involvement in postural control results in less regular
COP trajectories.
Signicance of studying postural sway dynamics
The present study was motivated from the expectation
that, compared to the measures traditionally used in the
study of COP trajectories, dynamical measures would
have surplus value in studying postural control. We
examined this expectation by reanalyzing an encompassing data set on the recovery of postural control
following stroke, which not only included COP measurements while standing upright with eyes open, but
also during sensory and cognitive manipulations. Notwithstanding the fact that the strength of statistical effects found for the dynamical measures were quite
similar to those found for the standard deviation (see
Tables 2, 3), this reanalysis led to several new important
results and discoveries vis-a`-vis the original analysis of
the data using conventional posturographic measures. In
particular, it was established that the datas variability
was temporally structured, that postural stability increased during rehabilitation, and that postural sway
regularity was positively related to the degree of cognitive involvement in postural control.
By combining the observed eects of health status,
rehabilitation, and task conditions on the dynamical
measures of interest, we arrived at a coherent theoretical
interpretation that may be summarized in less scientic
terms as follows. In stroke patients, maintaining balance
is more dicult due to neuromuscular impairments,
resulting in reduced postural stability. To cope with this
reduced stability, postural control is actively (i.e., cognitively) increased, resulting in more regular yet higher
dimensional COP trajectories. In the course of rehabilitation, postural stability improves, allowing the patients
to relax their cognitive involvement in postural control,
which leads to less regular COP trajectories. This process is reminiscent of that of automatization in skill

268

acquisition (cf. Huys and Beek 2002; Harbourne and

Stergiou 2003; Huys et al. 2003, 2004; Milton et al.
2004). In line with this interpretation, the introduction
of a cognitive dual task reduced the cognitive contribution to postural control, resulting in less regular COP
trajectories of larger dimension but similar stability.
Thus, by combining ndings gathered from a complementary set of dynamics-related analyses under various
task conditions and its recovery during rehabilitation
after stroke, we could make readily interpretable inferences about (changes in) the underlying postural control.
All in all, it is fair to conclude that the results of the
present study supported our expectation that the use of
dynamical measures would have surplus value in the
analysis of COP trajectories. The implication of this
overall conclusion is that, in future studies of postural
control, it should be deemed worthwhile to incorporate
both dynamical and conventional measures in the
analysis of COP trajectories.
Acknowledgments This research was conducted while the rst author was working on a grant of the Netherlands Organization for
Health Research and Development (ZonMw grant 1435.0004).

References
Abarbanel HDI (1996) Analysis of observed chaotic data. Springer,
New York
Baratto L, Morasso PG, Re C, Spada G (2002) A new look at
posturographic analysis in the clinical context: sway-density
versus other parameterization techniques. Motor Control
6:246270
Bohannon RW (1995) Standing balance, lower extremity muscle
strength, and walking performance of patients referred for
physical therapy. Percept Mot Skills 80:379385
Brown LA, Sleik RJ, Winder TR (2002) Attentional demands for
static postural control after stroke. Arch Phys Med Rehabil
83:17321735
Brunnstrom S (1966) Motor testing procedures in hemiplegia:
based on sequential recovery stages. Phys Ther 46:357375
Buzzi UH, Stergiou N, Kurz MJ, Hageman PA, Heidel J (2003)
Nonlinear dynamics indicates aging aects variability during
gait. Clin Biomech 18:435443
Cabrera JL, Bormann R, Eurich C, Ohira T, Milton J (2004) Statedependent noise and human balance control. Fluct Noise Lett
4:L107L118
Cohen J (1988) Statistical power analysis for the behavioral sciences. Erlbaum, Hillsdale, NJ
Collen FM, Wade DT, Bradshaw CM (1990) Mobility after stroke:
reliability of measures of impairment and disability. Int Disabil
Stud 12:69
Collins JJ (1999) Fishing for function in noise. Nature 402:241242
Collins JJ, De Luca CJ (1993) Open-loop and closed-loop control
of posture: a random-walk analysis of center-of-pressure trajectories. Exp Brain Res 95:308318
Collins JJ, De Luca CJ (1995) Upright, correlated random walks: a
statistical-biomechanics approach to the human postural control system. Chaos 5:5763
Collins JJ, Priplata AA, Gravelle DC, Niemi J Harry J, Lipsitz LA
(2003) Noise-enhanced human sensorimotor function. IEEE
Eng Med Biol Mag 22:7683
De Haart M, Geurts AC, Huidekoper SC, Fasotti L, van Limbeek
J (2004) Recovery of standing balance in postacute stroke patients: a rehabilitation cohort study. Arch Phys Med Rehabil
85:886895

Delignie`res D, Deschamps T, Legros A, Caillou N (2003) A

methodological note on nonlinear time series analysis: is the
open- and closed-loop model of Collins and De Luca (1993) a
statistical artifact? J Mot Behav 35:8696
Frank TD, Daertshofer A, Beek PJ (2001) Multivariate OrnsteinUhlenbeck processes with mean-eld dependent coecients:
application to postural sway. Phys Rev E 63:0011905/116
Fraser AM, Swinney HL (1986) Independent coordinates for
strange attractors from mutual information. Phys Rev A
33:11341140
Fugl-Meyer AR, Jaasko L, Leyman I, Olsson S, Steglind S (1975)
The post-stroke hemiplegic patient 1. A method for evaluation
of physical performance. Scand J Rehabil Med 7:1331
Geurts ACH, de Haart M, van Nes IJW, Duysens J (2005) A review of standing balance recovery from stroke. Gait Posture
22:267281
Glass L (2001) Synchronization and rhythmic processes in physiology. Nature 410:277284
Goldberger AL (1996) Non-linear dynamics for clinicians: chaos
theory, fractals, and complexity at the bedside. Lancet
347:13121314
Goldberger AL (1997) Fractal variability versus pathological
periodicity: complexity loss and stereotypy in disease. Perspect
Biol Med 40:543561
Goldberger AL, Amaral LAN, Glass L, Hausdor JM, Ivanov
PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE
(2000) PhysioBank, PhysioToolkit, and PhysioNet: components
of a new research resource for complex physiologic signals.
Circulation 101:e215e220
Goldberger AL, Amaral LAN, Hausdor JM, Ivanov PC, Peng CK, Stanley HE (2002) Fractal dynamics in physiology: alterations with disease and aging. Proc Natl Acad Sci USA 99:2466
2472
Grassberger P (1986) Do climate attractors exist? Nature 323:609
612
Grassberger P, Procaccia I (1983) Characterization of strange
attractors. Phys Rev Lett 50:346349
Harbourne RT, Stergiou N (2003) Nonlinear analysis of the
development of sitting postural control. Dev Psychobiol
42:368377
Hurst HE (1965) Long-term storage: an experimental study. Constable, London
Huys R, Beek PJ (2002) The coupling between point-of-gaze and
ball movements in three-ball cascade juggling: the eects of
expertise, pattern and tempo. J Sports Sci 20:171186
Huys R, Daertshofer A, Beek PJ (2003) Learning to juggle: on the
assembly of functional subsystems into a task-specic dynamical organization. Biol Cybern 88:302318
Huys R, Daertshofer A, Beek PJ (2004) Multiple time scales and
multiform dynamics in learning to juggle. Motor Control
8:188212
Kantz H, Schreiber T (2004) Nonlinear time series analysis.
Cambridge University Press, Cambridge
Kaplan DT, Furman MI, Pincus SM, Ryan SM, Lipsitz LA,
Goldberger AL (1991) Aging and complexity of cardiovascular
dynamics. Biophys J 59:945949
Kay BA (1988) The dimensionality of movement trajectories and
the degrees of freedom problem: a tutorial. Hum Mov Sci
7:343364
Kiemel T, Oie KS, Jeka JJ (2002) Multisensory fusion and the
stochastic structure of postural sway. Biol Cybern 87:262277
Kyriazis M (2003) Practical applications of chaos theory to the
modulation of human ageing: nature prefers chaos to regularity. Biogerontology 4:7590
Lake DE, Richman JS, Grin MP, Moorman JR (2002) Sample
entropy analysis of neonatal heart rate variability. Am J Physiol
Regul Integr Comp Physiol 283:789797
Lipsitz LA (2002) Dynamics of stability: the physiologic basis of
functional health and frailty. J Gerontol Biol Sci 57A:B115
B125
Mandelbrot BB, van Ness JW (1968) Fractional Brownian motions, fractional noises and applications. SIAM Rev 10:422437

269
Megrot F, Bardy BG, Dietrich G (2002) Dimensionality and the
dynamics of human unstable equilibrium. J Mot Behav 34:323
328
Milton JG, Small SS, Solodkin A (2004) On the road to automatic:
dynamic aspects in the development of expertise. J Clin Neurophysiol 21:134143
Newell KM, van Emmerik REA, Lee D, Sprague RL (1993) On
postural stability and variability. Gait Posture 4:225230
Newell KM, Slobounov SM, Slobounova ES, Molenaar PCM
(1997) Stochastic processes in center-of-pressure proles. Exp
Brain Res 113:158164
Newell KM (1998) Degrees of freedom and the development of
center of pressure proles. In: Newell KM, Molenaar PCM
(eds) Applications of nonlinear dynamics to developmental
process modeling. Erlbaum, Hillsdale, NJ, pp 6384
Nienhuis B, Geurts AC, Duysens J (2001) Are elderly more
dependent on visual information and cognitive guidance in the
control of upright balance? In: Duysens J, Smits-Engelsman
BC, Kingma H (eds) Control of posture and gait. NPI, Maastricht, pp 585588
Packard NH, Crutcheld JP, Farmer JD, Shaw RS (1980) Geometry from time series. Phys Rev Lett 45:712716
Paillex R, So A (2005) Changes in the standing posture of stroke
patients during rehabilitation. Gait Posture 21:403409
Pascolo PB, Marini A, Carniel R, Barazza F (2005) Posture as a
chaotic system and an application to the Parkinsons disease.
Chaos Solitons Fractals 24:13431346
Peng C-K, Buldyrev SV, Havlin S, Simons M, Stanley HE, Goldberger AL (1994) Mosaic organization of DNA nucleotides.
Phys Rev E 49:16851689
Peng C-K, Havlin S, Stanley HE, Goldberger AL (1995) Quantication of scaling exponents and crossover phenomena in
nonstationary heartbeat time series. Chaos 5:8287
Peterka RJ (2002) Sensorimotor integration in human postural
control. J Neurophysiol 88:10971118
Pincus SM, Goldberger AL (1994) Physiological time-series analysis: what does regularity quantify? Am J Physiol Heart Circ
Physiol 266:H1643H1656
Pincus SM (1991) Approximate entropy as a measure of system
complexity. Proc Natl Acad Sci USA 88:22972301
Rangarajan G, Ding M (2000) Integrated approach to the assessment of long-range correlation in time series data. Phys Rev E
61:49915001
Raymakers JA, Samson MM, Verhaar HJJ (2005) The assessment
of body sway and the choice of the stability parameter(s). Gait
Posture 21:4858

Richman JS, Moorman JR (2000) Physiological time-series analysis

using approximate entropy and sample entropy. Am J Physiol
Heart Circ Physiol 278:H2039H2049
Riley MA, Balasubramaniam R, Turvey MT (1999) Recurrence
quantication analysis of postural uctuations. Gait Posture
9:6578
Riley MA, Turvey MT (2002) Variability of determinism in motor
behavior. J Mot Behav 34:99125
Rosenstein MT, Collins JJ, De Luca CJ (1993) A practical method
for calculating largest Lyapunov exponents from small data
sets. Phys D 65:117134
Scholz JP, Schoner G (1999) The uncontrolled manifold concept:
identifying control variables for a functional task. Exp Brain
Res 135:382404
Schoner G (1995) Recent developments and problems in human
movement science and their conceptual implications. Ecol
Psychol 7:291314
Takens F (1981) Detecting strange attractors in turbulence. In:
Rand DA, Young LS (eds) Dynamical systems and turbulence.
Springer, Berlin
Theiler J (1986) Spurious dimension from correlation algorithms
applied to limited time-series data. Phys Rev A 34:24272432
Theiler J, Eubank S, Longtin A, Galdrikian B, Farmer JD (1992)
Testing for nonlinearity in time series: the method of surrogate
data. Phys D 58:7794
Theiler J, Rapp PE (1996) Re-examination of the evidence for lowdimensional, nonlinear structure in the human electroencephalogram. Electroencephalogr Clin Neurophysiol 98:213222
Thurner S, Mittermaier C, Ehrenberger K (2002) Change of complexity patterns in human posture during aging. Audiol Neurootol 7:240248
Wade DT (1992) Measurement in neurological rehabilitation.
Oxford University Press, Oxford
Wiesenfeld K, Moss F (1995) Stochastic resonance and the benets
of noise: from ice ages to craysh and SQUIDs. Nature 373:33
36
Wing A, Daertshofer A, Pressing J (2004) Multiple time scales in
serial production of force: a tutorial on power spectral analysis
of motor variability. Hum Mov Sci 23:569590
Wolf A, Swift JB, Swinney HL, Vastano JA (1985) Determining
Lyapunov exponents from a time series. Phys D 16:285317
Yamada N (1995) Chaotic swaying of the upright posture. Hum
Mov Sci 14:711726
Zatsiorsky VM, Duarte M (1999) Instant equilibrium point and its
migration in standing tasks: rambling and trembling components of the stabilogram. Motor Control 3:2838