Beruflich Dokumente
Kultur Dokumente
net/publication/11620616
CITATIONS READS
1,242 523
6 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Klaus Lehnertz on 26 April 2019.
061907-2
INDICATIONS OF NONLINEAR DETERMINISTIC AND . . . PHYSICAL REVIEW E 64 061907
N * /2
2 iS共 i 兲
m j ⫽
N*
兺
i⫽1 S共 i兲
共2兲
n
1
F ⫽
n 兺 兩 m j ⫺m 兩 ,
j⫽1
共4兲
esis is that the time series is compatible with a Gaussian 3. Estimate of an effective correlation dimension
linear stochastic and stationary process measured by a static Using delay coordinates xi ⫽(x i ,x i⫹ , . . . ,x i⫹(m⫺1) )
and monotonic, possibly nonlinear, function. Since disconti- 关46兴 and the correlation sum
nuities between the end and beginning of a time series are
known to cause spurious spectral frequency components, N⫺1 N⫺1
2
segments of 4396 samples were at first cut out of the record- C 共 ,N 兲 ⫽
共 N⫺T 兲共 N⫺T⫺1 兲 兺 兺
i⫽0 j⫽i⫹T
⌰ 共 ⫺储xi ⫺x j 储兲
ings. Within these longer intervals, the beginning of each of
the final segments of N⫽4096 samples was then chosen in 共5兲
such a way that the amplitude difference of the last and first
关where ⌰ is the Heaviside step function, ⌰(a)⫽0᭙a⭐0
data points was within the range of amplitude differences of
consecutive data points, and the slopes at the end and begin- and ⌰(a)⫽1᭙a⬎0兴 the correlation dimension D 2 is de-
ning of the time series had the same sign. This algorithm fined 关45兴 by
avoids the use of window functions for a calculation of the
power spectrum. A comparable technique was applied in Ref. D 2 ⫽ lim lim d 共 ,N 兲 , 共6兲
N→⬁ →0
关18兴.
061907-3
RALPH G. ANDRZEJAK et al. PHYSICAL REVIEW E 64 061907
was chosen to match the resolution of the analog-to-digital parameter values reported were obtained from preanalysis
converter, and divided into 128 intervals. A ‘‘quasiscaling with regard to an optimum differentiation between nonlinear
region’’ 关 l , u 兴 is defined by deterministic and linear stochastic model systems. In order to
meet the requirements of the respective statistics, we chose
u ⫽max兵 兩 d 共 ,N 兲 兩 m⫽1 ⬎0.975其 , 共8兲 all parameters independently for both P and D 2,e f f .
l ⫽min兵 兩 兩 d 共 u ,N 兲 兩 m⫽25⫺d 共 ,N 兲 兩 m⫽25兩 5. Levels of null hypothesis testing - Statistical methods
⭐0.05d 共 ,N 兲 兩 m⫽25其 . 共9兲 All following steps of analysis were carried out for P and
D 2 , e f f . In the following, M stands for any of the two mea-
If u and l existed, and the number N r of values in sures. The null hypothesis was tested on two levels: on one
关 l , u 兴 was greater than or equal to 5, the estimate level for every individual EEG segment, and on a second
u level for each of the five sets of EEG segments. On the
1
D 2,e f f ⫽
Nr 兺
⫽ l
d 共 ,N 兲 兩 m⫽25 共10兲 individual level the null was rejected if M EEG ranged outside
the distribution calculated from all of its surrogates
兵 M Sur i⫽1, . . . ,s 其 . For a given set and measure, let R min be the
was computed. If no quasiscaling existed or if D 2,e f f ⭓7.2 number of rejections of the null hypothesis which were
⬇2 log N 共cf. Ref. 关53兴兲, D 2,e f f was set to an arbitrary value caused by the fact that M EEG was smaller than the minimum
of D u ⫽10. of 兵 M Sur i⫽1, . . . ,s 其 . To rate a given number of rejections of the
null hypothesis, the probability p min to find R min or less re-
4. Nonlinear prediction error jections on a set of n ⫽100 time series just by chance was
For each reference point xi (i⫽1, . . . ,N⫺m ) in a re- calculated as
constructed state space 共embedding dimension m⫽6 and
time delay ⫽8 sampling times兲 a fixed number (k⫽5) of
冉 冊冉 冊 冉 冊
R min k n⫺k
n 1 1
nearest neighbors 兵 x j 其 j⫽1, . . . ,k was used to perform an p min⫽ 兺
k⫽0 k s⫹1
1⫺
s⫹1
. 共15兲
H-step prediction:
k
1 In complete analogy, we calculated the probability p max to
xg
i⫹H ⫽
k 兺 xj⫹H .
j⫽1
共11兲 find R max or less rejections that were caused by the fact that
M EEG exceeded the surrogates’ maximum.
The difference between the actual xi⫹H and predicted trans- On the set level one surrogate per EEG segment was used.
lation xg In accordance with Ref. 关3兴, values of 兵 M EEG 其 and 兵 M SUR 1 其
i⫹H is the local prediction error
were used as paired observations for a nonparametric Wil-
i,xg
i⫹H
⫽ 兩 xi⫹H ⫺xg
i⫹H 兩 . 共12兲 coxon signed rank test. In order to ensure that the choice of
D u did not influence results of the null hypothesis testing on
the set level, the Wilcoxon test was carried out with different
The local prediction error for the mean of the time series x̄ is values of D u .
i,x̄⫽ 兩 xi⫹H ⫺x̄兩 , 共13兲
III. RESULTS
where x̄ is an m-dimensional vector that carries the mean in
Figure 4 depicts results of exemplary EEGs. Table I and
each of its component. Finally the nonlinear prediction error
Fig. 5 summarize the results of all EEG segments and sets.
共P兲 is calculated from
On the individual level, sets E and D exhibited highest and
second highest R min values for both P and D 2,e f f . For all
R共 i,xg )
P⫽
i⫹H
, 共14兲 sets, the use of P led to higher R min values than the use of
R共 i,x̄兲 D 2,e f f . This was particularly true for sets B and C, for which
the strongest discordance between the two measures was
where R indicates the root mean square. For a nearest neigh- found: while for P both R min values were significant, they
bor search for every reference point xi , a Theiler window were both nonsignificant for D 2,e f f . Concordance was found
(T⫽25 sampling times兲 was applied to its own trajectory only for set A: here none of the two measures led to signifi-
segment and to neighboring trajectory segments: In a first cant R min values. None of the R max values was significant.
step, vectors with indices 兵 i⫺T, . . . ,i⫺1,i⫹1, . . . ,i⫹T 其 Segments with rejection of the null hypothesis for both P and
were discarded. In a second step, out of a group of nearest D 2,e f f were found in sets D and E only.
neighbors passing the first step but closer to each other in For the vast majority of segments D 2,e f f could not be
time than T, only the one nearest to xi was included. The calculated, and was therefore set to D u 共cf. Table I兲. For sets
second step is important to ensure that information for the P A and B, this was the case for all EEG segments and almost
is gathered from multiple adjacent trajectory segments rather all surrogates. Consequently, the null hypothesis on the set
than from only one trajectory segment 共cf. Refs. 关48,49兴兲. level could not be tested using D 2,e f f . For set C, only a few
The prediction horizon H was set to 65 sampling times. The segments resulted in D 2,e f f ⬍D u , whereas most segments re-
061907-4
INDICATIONS OF NONLINEAR DETERMINISTIC AND . . . PHYSICAL REVIEW E 64 061907
FIG. 4. Results for the exemplary EEG times series depicted in Fig. 3. From top to bottom: results for EEG-a to EEG-e. Left column:
P EEG and P SUR 1, . . . ,s vs prediction horizon H. Circles mark P EEG . Right column: d(,N) EEG and d(,N) SUR 1, . . . ,s vs the logarithm of the
hypersphere radius for an embedding dimension m⫽25. Diamonds mark d(,N) EEG . In addition, d(,N) EEG for embedding dimension
m⫽1 is given. Asterisks mark cases for which the null hypothesis was rejected. Note that for EEG-b there is a limited range of the
hypersphere radius for which d(,N) EEG was below the surrogates minimum. Hence the null hypothesis could be rejected. Neither the
EEG nor the surrogate time series, however, resulted in a finite value of D 2,e f f , so that the null was not rejected for D 2,e f f . Only for
examples EEG-d and EEG-e values of D 2,e f f could be estimated: 5.5 and 4.4, respectively.
061907-5
RALPH G. ANDRZEJAK et al. PHYSICAL REVIEW E 64 061907
TABLE I. Results for P and D 2,e f f on the individual level 共rows 1 and 2 and 6 and 7兲 and on the set level
共rows 3 and 4 and 8 and 9兲. Values of p max are not listed, since all were nonsignificant. Values in row 5
denote the number of segments for which a value of D 2,e f f could be computed for the EEGs and the
surrogates, respectively. The last row contains the number of segments for which the null was rejected for
both measures. n.s.: values of p⬎0.05 are assumed to be nonsignificant. x: no Wilcoxon test was carried out
for sets A and B 共see the body text兲.
A B C D E
061907-6
INDICATIONS OF NONLINEAR DETERMINISTIC AND . . . PHYSICAL REVIEW E 64 061907
tervals, and the anticipation of seizures by analysis of multi- age any further statistical analysis of the set level which as-
channel EEGs recorded from epilepsy patients 关31,36 –39兴. sumes the sets to be homogeneous.
The results of these studies, as well as the present one, sug- Doubtlessly, sets under investigation exhibited distinct
gest that the pathological epileptic process imposes certain linear properties described by power spectra and amplitude
constraints on neuronal dynamics. Even in the absence of distributions. However, since these properties were shared by
seizure activity these constraints appear to be reflected by the surrogates, they cannot account for results given here. It
nonlinear deterministic traits in the EEG, as suggested by is important, however, to point out that a rejection of the
results found for set D. Most prominently, however, this phe- surrogates’ null hypothesis is only a necessary but not suffi-
nomenon was observed during seizure activity 共set E兲. This cient criterion for nonlinearity. Surrogates that were used are
finding is in correspondence with previous studies on intra stationary by construction. Therefore, even if a simple sta-
cranial 关29兴 and extracranial 关9,32兴 recordings of epileptic tionarity inclusion criterion was used to select the EEG seg-
seizures. Theiler 关32兴 even a priori assumed seizure activity ments, nonstationarity as a cause of ‘‘false’’ positive rejec-
to be ‘‘undoubtedly nonlinear.’’ tions cannot be excluded. Since this stationarity test is based
When comparing extracranial and intracranial recording on linear properties of the time series, it can neither be
locations, it is important to note that the latter integrate po- highly sensitive nor highly specific for nonstationarities of
tentials over a much smaller steradian, i.e., fewer neurons nonlinear dynamical systems. Nonlinear techniques calcu-
contribute to the measured potentials, which furthermore are lated in a reconstructed state space such as the one proposed
less filtered as compared to extracranial recording locations. in Ref. 关54兴 could help to overcome this shortcoming. Fur-
This might explain differences in results from sets A and C. thermore, Kugiumtzis 关55兴 showed that although the iterative
An additional and not contradictory explanation might be amplitude adjusted surrogates are more consistent in repre-
given by the following. Set C was measured from brain re- senting the given null hypothesis than older techniques 关56兴,
gions which were proven to be nonepileptogenic but which the remaining mismatch of linear correlations of the original
may nevertheless participate in secondary, nonautonomous time series can still be relevant, and can cause false rejec-
epileptic processes initiated by the epileptogenic zone. tions of the null. Since these and other shortcomings weaken
The significant numbers of rejections R min on the indi- the probative force of the applied method, we only use the
vidual level allow two different interpretations. On the one term indication rather than evidence of nonlinear determin-
hand, the sets could be assumed to have a certain distribution istic structures.
of some distinct dynamical property not included in the null It has become a common point of view that values of an
hypothesis. Due to the limited sensitivity of the applied test estimate of the correlation dimension calculated from an un-
statistics for this very property, the individual null will be known dynamical system cannot be taken as a true estimate
rejected only for segments in one tail of this distribution. of a number of degrees of freedom. Particularly, our results
Differences of R min values between sets could then be ex- give no indication of chaos in the underlying dynamical sys-
plained by different centers and widths of these distributions. tem brain. In a way this is to be expected, since a purely
Given the respective calculation parameters used in our chaotic behavior of brain functions would consequently im-
study, both higher values and a higher intergroup variability ply that our brains’ behavior is changed dramatically by each
of R min were found for P as compared to D 2,e f f . A compari- and every arbitrary small input.
son between Refs. 关4兴, and 关8兴, as well as data reported in In accordance with other studies, our results show that an
Ref. 关18兴 show analogous results. This could be interpreted application of NTSA measures to EEG dynamics offers in-
as a higher degree of sensitivity of P for indications of dy- sights into the dynamical nature and variability of the system
namical structures in EEG time series. The contrary conclu- brain. As a feedback for NTSA, complicated properties of
sion, i.e., a lower level of specificity, cannot of course be EEG dynamics can further motivate one to improve existing
ruled out. methods and develop new methods 关57兴.
On the other hand, the sets could consist of different sub-
ACKNOWLEDGMENTS
sets, each containing segments of different dynamical nature.
This idea was followed in Ref. 关8兴, where Stam et al. con- We are grateful to Wieland Burr, Thomas Kreuz, Thomas
cluded that their ‘‘study underscores the heterogeneous na- Schreiber, and Guido Widman for useful discussions. This
ture 关 . . . 兴 of the alpha rhythm from a dynamical point of work was supported by the Deutsche Forschungsgemein-
view.’’ This interpretation of different subsets would discour- schaft.
关1兴 E. Ott, T. Sauer, and J. A. Yorke, Coping with Chaos 共Wiley, J. Psychophysiol 26, 237 共1997兲.
New York, 1994兲. 关5兴 W.S. Pritchard, D.W. Duke, and K.K. Krieble, Psychophysiol-
关2兴 H. Kantz and T. Schreiber, Nonlinear Time Series Analysis ogy 32, 486 共1995兲.
共Cambridge University Press, Cambridge, England, 1997兲. 关6兴 M. Palus, Biol. Cybern. 75, 389 共1996兲.
关3兴 S.A.R.B. Rombouts, R.W.M. Keunen, and C.J. Stam, Phys. 关7兴 A.C.K. Soong and C.I.J.M. Stuart, Biol. Cybern. 62, 55
Lett. A 202, 352 共1995兲. 共1989兲.
关4兴 F.H. Lopes da Silva, J.P. Pijn, D. Velis, and P.C.G. Nijssen, Int. 关8兴 C.J. Stam, J.P.M. Pijn, P. Suffczynski, and F.H. Lopes da Silva,
061907-7
RALPH G. ANDRZEJAK et al. PHYSICAL REVIEW E 64 061907
Clin. Neurophysiol. 110, 1801 共1999兲. 关33兴 A. Babloyantz and A. Destexhe, Proc. Natl. Acad. Sci. U.S.A.
关9兴 H. Jing and M. Takigawa, Biol. Cybern. 83, 391 共2000兲. 83, 3513 共1986兲.
关10兴 J. Jeong, M.S. Kim, and S.Y. Kim, Phys. Rev. E 60, 831 关34兴 F. Mormann, K. Lehnertz, P. David, and C.E. Elger, Physica D
共1999兲. 144, 358 共2000兲.
关11兴 A. Babloyantz, J.M. Salazar, and G. Nicolis, Phys. Lett. A 111, 关35兴 G.W. Frank, T. Lookman, M.A.H. Nerenberg, C. Essex, J. Le-
152 共1985兲. mieux, and W. Blume, Physica D 46, 427 共1990兲.
关12兴 P. Achermann, R. Hartmann, A. Gunzinger, W. Guggenbuhl, 关36兴 K. Lehnertz and C.E. Elger, Phys. Rev. Lett. 80, 5019 共1998兲.
and A.A. Borbely, Electroencephalogr. Clin. Neurophysiol. 90, 关37兴 M. Le Van Quyen, J. Martinerie, V. Navarro, P. Boon, M.
384 共1994兲. D’Have, C. Adam, B. Renault, F. Varela, and M. Baulac, Lan-
关13兴 J. Fell, J. Röschke, and C. Schäffner, Biol. Cybern. 75, 85 cet 357, 183 共2001兲.
关38兴 J. Martinerie, C. Adam, M. Le Van Quyen, M. Baulac, S.
共1996兲.
Clemenceau, B. Renault, and F.J. Varela, Nat. Med. 4, 1173
关14兴 C.J. Stam, T.C.A.M. van Woerkom, and W.S. Pritchard, Elec-
共1998兲.
troencephalogr. Clin. Neurophysiol. 99, 214 共1996兲.
关39兴 K. Lehnertz, J. Arnhold, P. Grassberger, and C. E. Elger, Chaos
关15兴 J. Theiler and P. Rapp, Electroencephalogr. Clin. Neuro-
in Brain? 共World Scientific, Singapore, 2000兲.
physiol. 98, 213 共1996兲. 关40兴 L.D. Iasemidis, J.C. Sackellares, H.P. Zaveri, and W.J. Will-
关16兴 B. Jelles, J.H. van Birgelen, J.P.J. Slaets, R.E.M. Hekster, E.J. iams, Brain Topogr. 2, 187 共1990兲.
Jonkman, and C.J. Stam, J. Clin. Neurophysiol. 110, 1159 关41兴 J. Arnhold, K. Lehnertz, P. Grassberger, and C.E. Elger,
共1999兲. Physica D 134, 419 共1999兲.
关17兴 P.E. Rapp, T.R. Bashore, J.A. Martinerie, A.M. Albano, I.D. 关42兴 F.H. Lopes da Silva, Electroencephalogr. Clin. Neurophysiol.
Zimmermann, and A.I. Mees, Brain Topogr. 2, 99 共1989兲. 79, 81 共1991兲.
关18兴 C.L. Ehlers, J. Havstad, D. Prichard, and J. Theiler, J. Neuro- 关43兴 H. O. Lüders and I. Awad, in Epilepsy Surgery, edited by H. O.
sci. 18, 7474 共1998兲. Lüders 共Raven Press, New York, 1991兲, pp. 51– 61.
关19兴 G. Widman, T. Schreiber, B. Rehberg, A. Hoeft, and C.E. El- 关44兴 A. S. Weigend and N. A. Gershenfeld, Time Series Prediction:
ger, Phys. Rev. E 62, 4898 共2000兲. Forecasting the Future and Understanding the Past, Proceed-
关20兴 J. Jeong, T.H. Chae, S.Y. Kim, and S.H. Han, J. Clin. Neuro- ings of the Santa Fe Institute Studies in the Science of Com-
physiol. 18, 58 共2001兲. plexity Vol. XV 共Addison -Wesley, Reading, 1993兲.
关21兴 V. Müller, W. Lutzenberger, F. Pulvermüller, B. Mohr, and N. 关45兴 P. Grassberger and I. Procaccia, Phys. Rev. Lett. 50, 346
Birbaumer, Exp. Brain Res. 137, 103 共2001兲. 共1983兲.
关22兴 L. Pezard, R. Jech, and E. Ruzicka, Clin. Neurophysiol. 112, 关46兴 F. Takens, in Dynamical Systems and Turbulence, edited by D.
38 共2001兲. A. Rand and L.-S. Young, Lecture Notes in Mathematics Vol.
关23兴 C.J. Stam, T.C.A.M. van Woerkom, and R.W.M. Keunen, Biol. 898 共Springer-Verlag, Berlin, 1980兲, p. 366 –381.
Cybern. 77, 247 共1997兲. 关47兴 T. Schreiber and A. Schmitz, Phys. Rev. Lett. 77, 635 共1996兲.
关24兴 J.-L. Nandrino, L. Pezard, J. Martinerie, F. El Massioui, B. 关48兴 J. D. Farmer and J. J. Sidorowich, Exploiting Chaos to Predict
Renault, R. Jouvent, J.-F. Allilaire, and D. Wildlöcher, Neu- the Future and Reduce Noise, in Evolution, Learning, and
roReport 5, 528 共1994兲. Cognition, edited by Y.-C. Lee 共World Scientific, Singapore,
关25兴 P. Kirsch, C. Besthorn, S. Klein, J. Rindfleisch, and R. Olbrich, 1988兲, pp. 277–330.
Brain Topogr 2, 187 共1990兲. 关49兴 J. McNames, Proceedings of the International Workshop on
关26兴 K. Lehnertz and C.E. Elger, Electroencephalogr. Clin. Neuro- Advanced Black-Box Techniques for Nonlinear Modeling
physiol. 95, 108 共1995兲. 共Katholieke Universiteit, Leuven, 1998兲, pp. 112–128.
关27兴 B. Weber, K. Lehnertz, C.E. Elger, and H.G. Wieser, Epilepsia 关50兴 P. Grassberger, T. Schreiber, and C. Schaffrath, Int. J. Bifurca-
39, 922 共1998兲. tion Chaos Appl. Sci. Eng. 1, 521 共1991兲.
关28兴 M.C. Casdagli, L.D. Iasemidis, R.S. Savit, R.L. Gilmore, S. 关51兴 G. Widman, K. Lehnertz, P. Jansen, W. Meyer, W. Burr, and
Roper, and J.C. Sackellares, Electroencephalogr. Clin. Neuro- C.E. Elger, Physica D 121, 65 共1998兲.
physiol. 102, 98 共1997兲. 关52兴 J. Theiler, Phys. Rev. A 34, 2427 共1986兲.
关29兴 J.P.M. Pijn, D.N. Velis, M.J. van der Heyden, J. DeGoedde, C. 关53兴 J.-P. Eckmann and D. Ruelle, Physica D 56, 185 共1992兲.
van Veelen, and F.H. Lopes da Silva, Brain Topogr 9, 249 关54兴 T. Schreiber, Phys. Rev. Lett. 78, 843 共1997兲.
共1997兲. 关55兴 D. Kugiumtzis, Phys. Rev. E 60, 2808 共1999兲.
关30兴 R.G. Andrzejak, G. Widman, K. Lehnertz, C. Rieke, P. David, 关56兴 J. Theiler, S. Eubank, A. Longtin, B. Galdrikian, and J.D.
and C.E. Elger, Epilepsy Res. 44, 129 共2001兲. Farmer, Physica D 58, 77 共1992兲.
关31兴 L.D. Iasemidis, P. Pardalos, J.C. Sackellares, and D.-S. Shiau, 关57兴 EEG time series are available under http://www.meb.uni-
J. Combi. Optimization 5, 9 共2001兲. bonn.de/epileptologie/science/physik/eegdata.html Contribu-
关32兴 J. Theiler, Phys. Lett. A 196, 335 共1995兲. tion of further sets of EEG time series is encouraged.
061907-8