1: \documentclass[aps,twocolumn,showpacs]{revtex4}
2: \usepackage{epsfig}
3: %\newcommand{\be}{\begin{equation}}
4: %\newcommand{\ee}{\end{equation}}
5:
6: \begin{document}
7:
8: \title{Event synchronization: a simple and fast method to measure synchronicity and
9: time delay patterns.}
10:
11: \author{R. Quian Quiroga$^{1, 2}$}
12: \altaffiliation{Corresponding author. E-mail: rodri@vis.caltech.edu}
13: \author{T. Kreuz$^{1, 3}$ and P. Grassberger$^1$}
14: \affiliation{$^1$ John von Neumann Institute for Computing, Forschungszentrum J\"ulich,
15: D-52425 J\"ulich, Germany\\
16: $^2$ Sloan-Swartz Center for Theoretical Neurobiology, Div. of Biology, 139-74,\\
17: California Institute of Technology, Pasadena, 91125, CA\\
18: $^3$ Department of Epileptology, University of Bonn, Sigmund-Freud Str. 25,
19: D-53105 Bonn, Germany}
20:
21: \date{\today}
22:
23: \begin{abstract}
24:
25: We propose a simple method to measure synchronization and time delay patterns between
26: signals. It is based on the relative timings of events in the time series, defined
27: e.g. as local maxima. The degree of synchronization is obtained from the number
28: of quasi-simultaneous appearances of events, and the delay is calculated from the
29: precedence of events in one signal with respect to the other.
30: Moreover, we can easily visualize the time evolution of the delay and synchronization
31: level with an excellent resolution.
32:
33: We apply the algorithm to short rat EEG signals, some of them containing spikes. We also
34: apply it to an intracranial human EEG recording containing an epileptic seizure, and
35: we propose that the method might be useful for the detection of foci and for seizure
36: prediction. It can be easily extended to other types of data and it is very simple
37: and fast, thus being suitable for on-line implementations.
38:
39: \end{abstract}
40:
41: \pacs{05.45.Tp; 05.45.Xt; 87.90.+y; 87.19.Nn}
42: \maketitle
43:
44: \section{Introduction}
45:
46: In recent years, several measures of synchronization have been proposed and applied
47: successfully to different types of data.
48: Among these studies we can distinguish two main approaches:
49: 1) One based on similarities of trajectories in phase space (constructed e.g. by
50: time-delay embedding) \cite{schiff,quyen,arnhold,quian,quian1};
51: 2) One that measures phase differences between the signals, where the phases are
52: defined either from a Hilbert \cite{rosemblum,tass,florian} or from a wavelet
53: transform \cite{lachaux,rodriguez} (as shown in \cite{quian1}, these
54: two apparently different phases are indeed closely related).
55:
56: These new methods compete in popularity with standard measures such as the
57: cross-correlation, the coherence function, mutual information, and also with simple
58: visual inspection of the recordings.
59: Cross-correlation and coherence are clearly the measures most used so far.
60: In contrast to them, all new measures are {\it non-linear} in the sense that they
61: depend also on properties beyond second moments. In addition, some of them have
62: the advantage of being asymmetric, eventually being able to show driver-response
63: relationships \cite{arnhold,quian}.
64:
65: Among others, synchronization measures have been used for the study of
66: electroencephalogram (EEG) signals. Applications include prediction and localization
67: of epileptic activity \cite{quyen,arnhold,florian}, phase-locking between different
68: recording sites upon visual stimulation \cite{lachaux,rodriguez}, resonance between EEG and
69: muscle activity in Parkinson patients \cite{tass}, desynchronization upon lesions in the
70: thalamic reticular nucleus in rats \cite{sleepwake}, synchronization in motoneurons
71: within the spinal cord \cite{schiff}, etc.
72:
73: In the present paper we present a very simple algorithm that can be used for any
74: time series in which we can define {\it events}. These can be spikes in single-neuron
75: recordings, epileptiform spikes in EEGs, heart beats, stock market crashes, etc. In
76: principle, when dealing with signals of different character, the events
77: could be defined differently in each time series, since their common cause might
78: manifest itself differently in each series.
79: %For instance, they could be maxima in one MEG channel, but minima in another.
80: This {\it event synchronization} (ES) does not require the notion of phase. It cannot
81: distinguish between different forms
82: of $m:n$ lockings \cite{rosemblum,tass}, but it can tell which of the two time series
83: leads the other. And, above all, it is very simple conceptually and easy to implement.
84: Due to that, it can be used on-line and can show rapid changes of synchronization
85: patterns.
86:
87: \section{Event synchronization and delay asymmetry}
88:
89: Given two simultaneously measured discrete univariate time series $x_n$ and $y_n$,
90: $n=1,\ldots,N$, we first define suitable {\it events} and event times $t^x_i$ and
91: $t^y_j$ $(i = 1, \ldots, m_x;
92: j = 1, \ldots, m_y)$. In the signals to be analyzed in this paper, these events will be
93: simply local maxima, subject to some further conditions. If the signals are synchronized,
94: many events will appear more or less simultaneously. Essentially, we count the fraction
95: of event pairs matching in time, and we count how often each time series leads in these
96: matches. Similar concepts were used in \cite{pijn}.
97:
98: Let us first assume that there is a well defined characteristic event rate in each
99: time series. Counter examples include strong chirps and onsets of epileptic seizures
100: where event rates change rapidly. Such cases will be treated below.
101: Allowing a time lag $\pm\tau$ between two `synchronous' events (which should be
102: smaller than half the minimum inter-event distance, to avoid double counting),
103: let us denote by $c^{\tau} (x|y)$ the number of times an event appears in
104: $x$ shortly after it appears in $y$, i.e:
105: \begin{equation}
106: c^{\tau} (x|y) = \sum_{i=1}^{m_x} \sum_{j=1}^{m_y} J_{ij}^\tau \label{eq:cxy}
107: \end{equation}
108: with
109: \begin{equation}
110: J_{ij}^\tau = \cases{ 1 & if ~~$0 < t_i^x - t_j^y \leq \tau$ \cr
111: 1/2 & if ~~$ t^x_i = t^y_j $ \cr
112: 0 & else} \label{eq:j}
113: \end{equation}
114: and analogously for $c^{\tau} (y|x)$. Next, we define the symmetrical and anti-symmetrical
115: combinations\begin{equation}
116: Q_{\tau} = \frac {c^{\tau}(y|x) + c^{\tau}(x|y)} {\sqrt{m_x m_y}} \ , \quad
117: q_{\tau} = \frac {c^{\tau}(y|x) - c^{\tau}(x|y)} {\sqrt{m_x m_y}} \ , \label{eq:qxy}
118: \end{equation}
119: which measure the synchronization of the events and their delay behavior, respectively.
120: They are normalized to $0 \leq Q_{\tau} \leq 1$ and $-1 \leq q_{\tau} \leq 1$. We
121: have $Q_{\tau} =1$ if and only if the events of the signals are fully synchronized. In
122: addition, if the events in $x$ always precede those in $y$, then $q_{\tau} = 1$.
123:
124: In cases where we want to avoid a global time scale $\tau$ since event rates change
125: during the recording, we use a local definition $\tau_{ij}$ for each event pair $(ij)$.
126: More precisely, we define
127: \begin{equation}
128: \tau_{ij} = \min \{ t^x_{i+1} - t^x_i, t^x_i - t^x_{i-1},
129: t^y_{j+1} - t^y_j, t^y_j - t^y_{j-1}\} \,/2\;. \label{eq:tauij}
130: \end{equation}
131: We then define $J_{ij}$ as in Eq.(\ref{eq:j}) with $\tau$ replaced by $\tau_{ij}$,
132: and $c(x|y)$ as in Eq.(\ref{eq:cxy}) with $J_{ij}^\tau$ replaced by $J_{ij}$. The
133: factor $1/2$ in the definition of $\tau_{ij}$ avoids double counting if, e.g., two
134: events in $x$ are close to the same event in $y$. Of course, one could also make
135: other choices, e.g. by taking $\tau_{ij}$ smaller than in Eq.(\ref{eq:tauij}) or by
136: using $\tau'_{ij} = \min\{\tau, \tau_{ij}\}$. As in the definition of events, an
137: optimal choice of $\tau$ depends on the problem. In the
138: following we shall suppress the dependence on $\tau$, understanding that all formulas
139: apply for both variants.
140:
141: To obtain time resolved variants of $Q$ and $q$ we simply modify eq.(\ref{eq:cxy}) to
142: \begin{equation}
143: c_n(x|y) = \sum_i \sum_j J_{ij}\, \Theta(n-t_i^x)
144: \label{eq:cxyt}
145: \end{equation}
146: with $n = 1, \ldots, N$ and $\Theta$ the step function (i.e. $\Theta (x) = 0$ for
147: $x \leq 0$ and $\Theta(x) = 1$ for $x > 0$). Similarly, $c_n(y|x)$ is obtained by
148: exchanging $x$ and $y$. Then, we define the time resolved anti-symmetric combination
149: $q(n) = c_n(y|x) - c_n(x|y)$ which can be seen as a random
150: walk that takes one step up every time an event in $x$ precedes one in $y$
151: and one step down if vice versa. If an event occurs simultaneously in both signals
152: or if it appears only in one of them, the random walker does not move.
153: Exchanging $x$ and $y$ just reverses the walk.
154: For non-synchronized signals, we expect to obtain a random walk with the typical
155: diffusion behavior. With delayed synchronization we will have a bias going up
156: (down) if $x$ precedes (follows) $y$.
157: We should remark that such a bias clearly shows the presence of a time delay of one
158: signal with respect to the other, but does not necessarily prove a driver-response
159: relationship, although it might suggest it. In fact, internal delay loops of one of
160: the systems can fool the interpretation. Also, the two signals might be driven by a
161: common hidden source and the bias just indicates different delays.
162:
163: The time course of the strength of ES can be obtained from $Q(n) = c_n(y|x) + c_n(x|y)$.
164: If an event is found both in $x$ and $y$ within the window $\tau$ (resp. $\tau_{ij}$),
165: $Q(n)$ increases one step, otherwise it does not change. Of course, $Q(n)$ will
166: also not change if there are no new events at all. The synchronization level at time
167: $n$, averaged over the last $\Delta n$ time steps, is thus obtained as
168: \begin{equation}
169: Q'(n) = {Q(n)-Q(n-\Delta n)\over \sqrt{\Delta n_x\Delta n_y}}\;,
170: \end{equation}
171: where $\Delta n_x$ and $\Delta n_y$ are the numbers of events in the interval
172: $[n-\Delta n,n]$.
173: Similarly, we can also define instantaneous delay asymmetries $q'(n)$.
174:
175: \section{Applications}
176:
177: Let us now apply these concepts to two sets of intracranial EEG recordings, one from
178: rats and the other from an epileptic patient.
179:
180: \subsection{Rat EEGs}
181:
182: The five pairs of rat EEG signals were
183: obtained from electrodes placed on the left and right frontal cortex of male
184: adult WAG/Rij rats (a genetic animal model of human epilepsy) \cite{giles}. They
185: were referenced to an electrode placed in the cerebellum, filtered between 1-100
186: Hz and digitized at 200 Hz. In Fig.~\ref{fig:examples} we show these signals \cite{www}.
187: The first pair (example A in Fig.~\ref{fig:examples}) is a normal EEG, all others
188: contain spike discharges (not to be confused with spikes in single neuron recordings)
189: which are the landmark of epileptic activity. They arise from abnormal synchronization
190: in an epileptic brain even when there are no seizures. A localized appearance of spikes
191: can indeed delimit a zone with abnormal activity (though this will not necessarily
192: be the epileptic focus). Furthermore, time delays between them can identify the
193: electrode closest to the epileptic focus, especially at the onset of seizures.
194:
195: \begin{figure}
196: \begin{center}
197: \epsfig{file=fig1.eps, width=7.6cm,angle=0}
198: \end{center}
199: \caption{Five pairs of rat EEG signals from right and left cortical intracranial
200: electrodes. For a better visualization, left signals are plotted with an offset.}
201: \label{fig:examples}
202: \end{figure}
203:
204: Several measures of synchronization were recently applied to the first three cases
205: of Fig.~\ref{fig:examples} \cite{quian1}. Since spike trains lasted usually about
206: 5 seconds, the challenge was to try the different measures in these short
207: epochs. Surprisingly, nearly all the measures gave qualitatively similar results
208: hard to be guessed beforehand. These examples and two additional cases (D and E),
209: also containing spikes, will be further analyzed in this paper.
210:
211: For the example A it is difficult, due to its random-like appearance, to
212: visually estimate its level of synchronization and any delay of one electrode with
213: respect to the other. However, we can already observe some patterns appearing
214: nearly simultaneously in both the left and right channels, thus showing some
215: degree of interdependence.
216: The spike-wave trains in the other examples in principle suggests a high level of
217: synchronization. However, as already shown in \cite{quian1}, the spikes of example
218: C appear with a varying time lag between right and left channels and are therefore
219: much less synchronized than those in B.
220: This is of course not easily seen by visual inspection of Fig.~\ref{fig:examples},
221: but will be clear from the following analysis.
222:
223: Events were defined as local maxima fulfilling the following additional conditions:
224: \begin{enumerate}
225: \item $x(t_i) > x(t_{i+k})$, for $k = -K+1,\ldots, 0, \ldots,K-1$
226: \item $x(t_i) > x(t_{i \pm K}) + h $
227: \end{enumerate}
228: and the same for $y$. We took $K = 3$ and $h = 0.1$. Other choices gave very
229: similar results.
230:
231: Since the rate of events is more or less constant, we used a fixed $\tau$. The
232: choice $\tau=2$ gave a good discrimination between the five cases.
233: All results shown below were compared to those obtained with surrogate pairs
234: which were defined by shifting the left channel signals 500 data points (2.5 sec)
235: to the right, with periodic boundary conditions.
236: Our test hypothesis is that without changing the individual properties
237: of each signal, after a large enough shifting synchronization should reach a
238: background `zero' level. The usefulness of such surrogates was discussed in
239: in more detail in \cite{quian1}.
240:
241: \begin{table}
242: \begin{center}
243: \begin{tabular}[c]{c c c c c c}
244: {\bf Example} & \vline & {\bf $Q_{\tau = 2}$} & {\bf $q_{\tau = 2}$} &
245: {\bf $Q_{\tau = 2}^{^{\rm surr}}$} & {\bf $q_{\tau = 2}^{^{\rm surr}}$} \\
246: \hline
247: {\bf A} & \vline & 0.57 & ~0.15 & 0.24 & -0.01 \\
248: {\bf B} & \vline & 0.80 & -0.29 & 0.29 & ~0.01 \\
249: {\bf C} & \vline & 0.48 & -0.20 & 0.13 & -0.01 \\
250: {\bf D} & \vline & 0.93 & -0.59 & 0.41 & ~0.04 \\
251: {\bf E} & \vline & 0.90 & -0.13 & 0.46 & ~0.03 \\
252: \end{tabular}
253: \vspace{0.5cm}
254: \caption{Time averaged event synchronization $Q$ and delay $q$ for the five examples
255: of Fig.~\ref{fig:examples}. Positive values of $q$ indicate that events in the left
256: side lags behind the right one. Surrogate values of synchronization were obtained by
257: shifting left channel signals by 500 data points.}
258: \label{tab:q}
259: \end{center}
260: \end{table}
261:
262: For the five EEG signals of Fig.~\ref{fig:examples}, we show the values of $Q_{\tau =2}$
263: and $q_{\tau =2}$ in Table~\ref{tab:q}, both for the original signals and the
264: `time-shifted' surrogates. We observe that synchronization levels rank
265: ${\rm D > E > B > A > C}$.
266: This is in agreement with the analysis of examples A, B and C done in
267: \cite{quian1} with several other measures of synchronization. Note that even example A
268: is ranked consistently with the other measures, although it does not contain obvious
269: events such as the spikes of the other examples.
270:
271: \begin{figure}
272: \begin{center}
273: \epsfig{file=qs2.eps, width=7.6cm,angle=0}
274: \end{center}
275: \caption{Time resolved event synchronization $Q'_{\tau = 2} (n)$ for the examples
276: of Fig.~\ref{fig:examples}. Blue horizontal lines correspond to the time averages
277: $\pm 1\sigma$ of the surrogate.}
278: \label{fig:qs}
279: \end{figure}
280:
281: All synchronization values are clearly higher than those of their corresponding
282: surrogates (surrogates constructed with other delay values gave similar results). These
283: surrogate values vary a lot for the different examples. This stresses the importance
284: of keeping the individual properties of the signals when constructing surrogates.
285: Except for example A, the values of $q$ show that the signals from the right hemisphere
286: lag behind the left ones ($y$). A closer visual inspection of Fig.~1 at higher
287: resolution shows that this lag is usually 1 data point. The reason of this systematic
288: lag is unclear (it could be an artifact of the data acquisition or a real physiological
289: effect) and it is beyond the scope of this paper.
290:
291: In Fig.~\ref{fig:qs} we show the time evolution of synchronization $Q'(n)$ for the
292: five examples, calculated with a window of $\Delta n = 100$ data points. For most
293: of the time, they are higher than the values calculated from time-shifted surrogates
294: (the light blue horizontal lines indicate time averages $\pm 1$ standard deviation).
295: In examples A, B and C we see abrupt changes of synchronization with time which seem
296: statistically significant.
297: In retrospect they can also be seen in Fig.~\ref{fig:examples} on closer inspection,
298: but they are much less obvious there and could easily be missed. Compared to the
299: first three, examples D and E are more stable in time. Finally, the time resolved
300: ES shows a better resolution than all synchronization measures considered in
301: \cite{quian1}.
302:
303: \begin{figure}
304: \begin{center}
305: \epsfig{file=qa2.eps, width=7.6cm,angle=0}
306: \end{center}
307: \caption{Time delays between the right and left channels (upper plot) and for the surrogates
308: (lower plot). Up (down) shifts mean precedence of the right (left) channel.}
309: \label{fig:qa}
310: \end{figure}
311:
312: Figure \ref{fig:qa} shows the time resolved asymmetry between the right and the
313: left channels (upper plot) and the results from surrogates (lower plot). In all
314: five cases, the bias is in agreement with the $q$ values shown in Table~\ref{tab:q}.
315: The bias in example D is not only the strongest but also the most constant, confirming
316: that D shows the most robust and stationary ES (compare Fig.~\ref{fig:qs}).
317: For the other examples we see regular changes with time. This
318: is of course very difficult to see in the original recordings, and it was also
319: not seen with any of the synchronization measures studied in \cite{quian1}.
320: As expected, for the surrogates we obtain random walks with small and
321: erratic displacements.
322:
323: \subsection{Human EEG}
324:
325: \begin{figure}
326: \begin{center}
327: \epsfig{file=matrix.eps, width=6.7cm,angle=0}
328: \end{center}
329: \caption{Time averaged event synchronization between the contacts on the left and right depth
330: electrodes (TL1-10 and TR1-10, respectively).}
331: \label{fig:matrix}
332: \end{figure}
333:
334: As a second example we analyzed an intracranial EEG recording from an epileptic patient
335: containing 12 min. of pre-seizure and seizure EEG. Data were recorded from 2 needle shaped
336: depth electrodes with 10 contacts each. They were symmetrically placed in the left (contacts
337: TL1 to TL10) and right (contacts TR1 to TR10) temporal lobes, in the entorhinal cortex
338: and hippocampal formation. The EEG was sampled at 173~Hz and band pass filtered
339: between 0.53-40~Hz. For further details on the data we refer to \cite{arnhold}. As in the
340: previous example, event times were defined as local maxima, but using $K=10$ and $h=50$
341: (this large $K$ was needed because the data are more noisy than the rat data, and
342: smaller values would have led to many spurious events).
343: Due to the varying event rate, we used a variable-$\tau$ approach. For the time
344: resolved event synchronization $Q'(n)$ we took a window $\Delta n = 1730$.
345:
346: \begin{figure}
347: \begin{center}
348: \epsfig{file=tl7-tls.eps, width=7.6cm,angle=0}
349: \end{center}
350: \caption{Time resolved event synchronization (upper plot) and delay asymmetries (lower plot)
351: between an
352: channel near the epileptic focus (TL7) and the remaining channels on the same side.
353: The red bar shows the duration of the epileptic seizure.}
354: \label{fig:tl7}
355: \end{figure}
356:
357: Figure \ref{fig:matrix} shows the time-averaged event synchronization values between all
358: channels. A detailed analysis of synchronization patterns for similar recordings has
359: already been described by Arnhold et al. \cite{arnhold} using a robust measure of
360: non-linear synchronization. Here, we just summarize the main results which are in
361: perfect agreement with those in Ref.~\cite{arnhold}. We first note that
362: synchronization between left and right electrodes is relatively low and that the right
363: contacts form two clusters: TR1-3 and TR4-10. This is just due to the fact that the
364: first 3 contacts were located in the entorhinal cortex and the remaining ones
365: in the hippocampus \cite{arnhold}. Moreover, for the right side we observe
366: a gradual decrease of synchronization with increasing distance between contacts.
367: The synchronization pattern for the left channels is different. There, the entorhinal
368: cortex/hippocampus separation is overshadowed by the epileptic activity leading to a
369: higher overall synchronization level.
370:
371: A visual analysis of the seizure onset revealed that contacts TL7 and TL8 showed the first
372: signs of seizure activity.
373: Figure \ref{fig:tl7} shows the time resolved synchronization $Q'(n)$ and delays $q(n)$
374: between TL7 and the remaining left side channels.
375: As expected, synchronization is largest between TL7 and its neighbors TL8 and TL6. It
376: is not homogeneous in time and we have several short drops before seizure starts.
377: Moreover, starting at seizure onset and during the whole seizure, synchronization of TL7
378: with TL8 and TL9 is high, while synchronization with TL6 and all others is
379: decreased. The lower panel shows that all left channels lag behind channel TL7.
380: There is just one exception: During the first part of the seizure,
381: channel TL7 falls back and channel TL8 leads for about half a minute (indeed, the
382: lead of TL7 is weakened already some 3 minutes before the seizure). After this,
383: TL7 takes up its lead even more vigorously than before. This might indicate
384: that the source of epileptic activity moves.
385: Whether these features are common to many epileptic seizures and whether they can have
386: clinical significance for e.g. seizure anticipation or focus localization requires
387: further study with a larger database.
388:
389: \begin{figure}
390: \begin{center}
391: \epsfig{file=tr7-tls.eps, width=7.6cm,angle=0}
392: \end{center}
393: \caption{Delay patterns between an contact in the non-focal side (TR7) against the other
394: contacts in the non-focal side (upper plot) and against the contacts in the focal
395: side (lower plot). No anomalous behavior is seen during the seizure (red bar). Notice
396: the different scales in the two plots.}
397: \label{fig:tr7}
398: \end{figure}
399:
400:
401: In Fig.~\ref{fig:tr7} we show the delays of the contralateral channel (TR7) with respect
402: to the other right channels (upper plot) and to the left channels (lower plot). Channels
403: TR4-6 strongly and steadily follow channel TR7, which itself follows channels TR8 and TR10.
404: This might reflect the source of `normal' synchronized activity.
405: A detailed analysis is outside the scope of this paper and will be further addressed
406: elsewhere. As seen from the lower panel, synchronization between both hemispheres is weak
407: and $q$ shows unbiased random walks. The complete absence of any deviant behavior during
408: the seizure reflects the fact that the seizure does not spread to the contralateral side.
409:
410: \section{Conclusion}
411:
412: In conclusion we presented a new approach to measure synchronization and time delays that is
413: based on the relative timings of events (in this study defined as local maxima). This also
414: gives an easy visualization of time-resolved synchronization and delay patterns. The method
415: is appealing due to its simplicity, straightforward implementation and speed. These features
416: make very easy its on-line implementation. In the particular case of EEGs, the proposed
417: approach is promising for the study of recordings of epileptic patients, where
418: synchronization is important and the analysis of time delay patterns could be useful for the
419: localization of the epileptic focus and the prediction of seizure onset.
420: Also, the method should be well suited for single-neuron recordings, where the fast
421: dynamics of spikes makes difficult the analysis with other measures.
422: In this paper we focussed on application to EEG signals, but the method can be easily
423: applied to other types of data just by adjusting the definition of events.
424:
425:
426: %\newpage
427: %\section*{Acknowledgments}
428:
429: We are very thankful to Ralph Andrzejak, Alexander Kraskov, Klaus Lehnertz, and Heinz
430: Schuster for stimulating discussions, to
431: Giles van Luijtelaar and Joyce Welting from NICI, University of Nijmegen, for the rats data
432: used in this paper and to K. Lehnertz and C. Elger from the Department of Epileptology,
433: University of Bonn, for the intracranial EEG data. T.K. is supported by the
434: Deutsche Forschungsgemeinschaft, SFB TR3.
435:
436: %\newpage
437:
438: \begin{thebibliography}{10}
439:
440: \bibitem{schiff}
441: S.J. Schiff, P.~So, T.~Chang, R.E. Burke and T. Sauer, Phys. Rev. E {\bf 54}, 6708 (1996).
442:
443: \bibitem{quyen}
444: M.~Le~Van Quyen, J.~Martinerie, C.~Adam, and F.J. Varela, Physica D {\bf 127}, 250 (1999).
445:
446: \bibitem{arnhold}
447: J.~Arnhold, P.~Grassberger, K.~Lehnertz, and C.E. Elger, Physica D {\bf 134}, 419 (1999).
448:
449: \bibitem{quian}
450: R. Quian Quiroga, J. Arnhold and P. Grassberger, Phys. Rev. E, {\bf 61}, 5142 (2000).
451:
452: \bibitem{quian1}
453: R. Quian Quiroga, A. Kraskov, T. Kreuz and P. Grassberger, Phys. Rev. E, in press.
454:
455: \bibitem{rosemblum}
456: M. Rosenblum, A. Pikovsky and J. Kurths, Phys. Rev. Lett, {\bf 76}, 1804 (1996).
457:
458: \bibitem{tass}
459: P. Tass, M. Rosenblum, J. Weule, J. Kurths, A. Pikovsky, J. Volkmann,
460: A. Schitzler and H. Freund, Phys. Rev. Lett, {\bf 81}, 3291 (1998).
461:
462: \bibitem{florian}
463: F. Mormann, K. Lehnertz, P. David and C.E. Elger, Physica D, {\bf 144}, 358
464: (2000).
465:
466: \bibitem{lachaux}
467: J. Lachaux, E. Rodriguez, J. Martinerie and F. Varela, Human Brain Mapping, {\bf 8}, 194 (1999).
468:
469: \bibitem{rodriguez}
470: E. Rodriguez, N. George, J. Lachaux, J. Martinerie, B. Renault and F. Varela, Nature, {\bf 397}, 430 (1999).
471:
472: \bibitem{sleepwake}
473: G. van Luijtelaar, J. Welting and R. Quian Quiroga,
474: in: van Bemmel {\it et al.} (eds.) {\sl Sleep-wake research in the Netherlands},
475: vol 11, pp:86-95 (Dutch Society for Sleep-Wake Research, 2000).
476:
477: \bibitem{pijn} J.P.M. Pijn, {\it Quantitative Evaluation of EEG Signals in Epilepsy},
478: PhD. Thesis, Amsterdam University (1990)
479:
480: \bibitem{giles}
481: G. van Luijtelaar and A. Coenen (eds.), {\sl The WAG/Rij rat model of absence epilepsy:
482: Ten years of research} (Nijmegen University Press, 1997).
483:
484: \bibitem{www}
485: The EEG signals can be downloaded from \\ www.vis.caltech.edu/\~{}rodrigo.
486:
487: \end{thebibliography}
488:
489: \end{document}
490: