1: % Template article for preprint document class `elsart'
2: % with harvard style bibliographic references
3: % SP 2001/01/05
4:
5: \documentclass{elsart}
6:
7: % Use the option doublespacing or reviewcopy to obtain double line spacing
8: %\documentclass[doublespacing]{elsart}
9:
10: % the natbib package allows both number and author-year (Harvard)
11: % style referencing;
12: %\usepackage{natbib}
13:
14: % if you use PostScript figures in your article
15: % use the graphics package for simple commands
16: %\usepackage{graphics}
17: % or use the graphicx package for more complicated commands
18: \usepackage{graphicx}
19: \usepackage{float}
20: % or use the epsfig package if you prefer to use the old commands
21: % \usepackage{epsfig}
22:
23: %\baselinestretch{2}
24: \usepackage{citesort}
25:
26: % The amssymb package provides various useful mathematical symbols
27: \usepackage{amssymb}
28:
29: \def \rb {\raisebox{.05ex}[.2cm][.2cm]}
30:
31: \begin{document}
32:
33: \begin{frontmatter}
34:
35: % Title, authors and addresses
36:
37: % use the thanksref command within \title, \author or \address for footnotes;
38: % use the corauthref command within \author for corresponding author footnotes;
39: % use the ead command for the email address,
40: % and the form \ead[url] for the home page:
41: % \title{Title\thanksref{label1}}
42: % \thanks[label1]{}
43: % \author{Name\corauthref{cor1}\thanksref{label2}}
44: % \ead{email address}
45: % \ead[url]{home page}
46: % \thanks[label2]{}
47: % \corauth[cor1]{}
48: % \address{Address\thanksref{label3}}
49: % \thanks[label3]{}
50:
51: \title{Estimation of time delay by coherence analysis}
52:
53: % use optional labels to link authors explicitly to addresses:
54: % \author[label1,label2]{}
55: % \address[label1]{}
56: % \address[label2]{}
57:
58: \author[adr1,adr2]{R. B. Govindan},
59: \author[adr1]{J. Raethjen},
60: \author[adr1]{F. Kopper},
61: \author[adr2]{J. C. Claussen},
62: \author[adr1]{G. Deuschl\corauthref{cor1}}
63: \corauth[cor1]{Prof. G. Deuschl, Department of Neurology, Christian-Albrechts
64: University of Kiel,
65: Schittenhelmstrasse 10, D-24105 Kiel, Germany}
66: \ead{g.deuschl@neurologie.uni-kiel.de}
67: \address[adr1] {Department of Neurology, Christian-Albrechts University of Kiel,
68: Schittenhelmstrasse 10, D-24105 Kiel, Germany.}
69: \address[adr2] {Institut f\"ur Theoretische Physik und Astrophysik, Christian-Albrechts University of Kiel,
70: Leibnizstrasse 15, D-24098 Kiel, Germany.}
71: \begin{abstract}
72: Using coherence analysis (which is an extensively used method to study
73: the correlations in frequency domain, between two simultaneously measured
74: signals) we estimate
75: the time delay between two signals. This method is suitable for time delay
76: estimation of narrow
77: band coherence signals for which the conventional methods cannot be reliably applied.
78: We show by analysing coupled R\"ossler attractors with a known delay, that the
79: method yields satisfactory results. Then, we apply
80: this method to human pathologic tremor. The delay between simultaneously measured traces of Electroencephalogram
81: (EEG) and Electromyogram (EMG) data of subjects with essential hand tremor is
82: calculated. We find that there is a delay of 11-27
83: milli-seconds ($ms$)
84: between the tremor correlated parts (cortex) of the brain (EEG) and the
85: trembling hand (EMG) which is in agreement with the experimentally observed delay value of
86: 15 $ms$ for the cortico-muscular conduction time. By surrogate analysis we
87: calculate error-bars of the estimated delay.
88: % Text of abstract
89:
90: \end{abstract}
91:
92: \begin{keyword}
93: Time series \sep Coherence \sep Spectral methods \sep Time delay
94: % keywords here, in the form: keyword \sep keyword
95:
96: % PACS codes here, in the form: \PACS code \sep code
97: \PACS 05.45.Tp \sep 42.25.Kb \sep 02.70.Hm
98: \end{keyword}
99:
100: \end{frontmatter}
101:
102: % main text
103: \section{Introduction}
104: %\label{}
105: The time delay between two dynamical systems can provide information on conduction velocity, and the nature and origin of
106: coupling, between the processes. So it is necessary to use
107: a well validated method for this purpose. Literature on time delay is vast and
108: well documented in \cite{ml1,ml2}. Often, in physiological time series
109: analysis, a single method cannot be made unique to be applicable for a wide class of
110: data stemming from the processes seemingly operated by similar
111: mechanisms. Methods used for
112: time delay estimation are no exception from this fact. Here, we use a
113: spectral based method, maximising coherence, for the time delay
114: estimation \cite{carter}. First, we apply this
115: method to uni- and bi-directionally coupled
116: R\"ossler attractors with a known delay between the systems.
117: In both cases, the results obtained are in good
118: agreement within a narrow range of the delay used in the simulation. Then, we
119: apply
120: this method to simultaneously recorded traces of Electroencephalogram
121: (EEG) and Electromyogram (EMG) of subjects with essential tremor a well known
122: pathological form of hand tremor. We obtain a delay
123: of 11-27 milli seconds ($ms$) between the tremor correlated cortical activity (EEG) and
124: EMG and the results are reasonably within the range of experimentally observed
125: value for cortico-muscular transmission \cite{roth}. This result, further confirms the involvement of cortex in
126: the generation of essential tremor \cite{hel1}.
127:
128: The paper is organised as follows: In section 2, we discuss in detail, the
129: methodology of coherence
130: analysis. Then, we extend the coherence analysis for time delay
131: estimation. Due to a time delay there will be a time
132: misalignment between the two time series thereby causing a reduction in the
133: coherence estimated between them. In order to compensate for the reduction in
134: coherence due to delay, we shift one of the time series keeping
135: the other constant and estimate the coherence as a function of the
136: shift. This method has been successfully applied to estimate the time
137: delay between the acoustic source and the receiving signals
138: \cite{carter}. On similar lines, phase synchronisation is used to
139: estimate the time delay between and among the atmospheric variables observed
140: at
141: different meteorological sites \cite{diego}. In section 3, we apply this method to afore mentioned
142: theoretical models. After validating the method by the results of standard
143: models, in section 4, we apply this method to estimate the time delay between
144: the simultaneously recorded EEG and EMG data of essential tremor subjects. We
145: discuss the results and conclude in section 5.
146:
147: \section{Methodology}
148: Let $x(t)$ and $y(t)$ be two simultaneously recorded data sets of length
149: $N$. The mean and standard deviation of the two data sets are, respectively,
150: set to zero and one. We divide the data sets into $M$ disjoint
151: segments of length $L$, such that $N=L \cdot M $. We calculate power
152: spectra, $\widehat{S_{xx}}$, $\widehat{S_{yy}}$ and cross spectrum
153: $\widehat{S_{xy}}$,
154: which is the
155: Fourier transform of the cross-correlation function of the signals $x(t)$ and
156: $y(t)$ \cite{ml1}, in each segment. Over cap in all the
157: quantities indicates that it is an estimate of that quantity. Finally, we
158: average
159: the power spectra
160: and the cross-spectrum across all the segments and calculate coherence as
161: follows \cite{hal1},
162: $$ \widehat{C(\omega)}~=~\frac{\widehat{|S_{xy}(\omega)}|^2}{\widehat{S_{xx}(\omega)} \cdot \widehat{S_{yy}(\omega)}}~.
163: \eqno 1 $$
164: The coherence spectrum provides the strength of correlation between the two
165: signals, $x(t)$ and $y(t)$. The confidence limit for coherence at the $100
166: \% ~\alpha $ is given by
167: $1-(1-\alpha)^{\frac1{(M-1)}}$. Thus, the coherence spectrum is always considered
168: with this
169: line. In all of our analysis, $\alpha$ is set to 0.99, and hence the confidence
170: limit is $1-0.01^{\frac1{(M-1)}}$. The estimated value of coherence, at a
171: frequency, above this
172: line indicates (a) significant coherence between the two time series at this
173: frequency;
174: (b) the magnitude (deviation of coherence from this line) determines the
175: degree of linear correlation between the two
176: time series at this frequency. The estimated value of coherence at a frequency below this line
177: is considered as the lack of correlation between the two time series at this
178: frequency.
179:
180: If the sampling frequency of the signals is $sf-$ $Hz$ (i.e. $sf$ number of
181: data points are sampled per second), then the frequency resolution of the quantities in
182: eq. 1, is $\frac{sf}{L}$. Thus, one should optimally choose the value of $L$
183: depending on the purpose of analysis, to compromise between the sensitivity
184: and reliability. Usage of the fixed segment length $L$ is questioned
185: and a variable segment length is suggested in \cite{jeti1}, to make
186: quantitative assessment of the signal. But, for all practical purposes, a fixed
187: segment length is easily implementable and hence, it is used in the
188: forthcoming analysis.
189:
190:
191: While the coherence spectrum provides the strength of correlation between the two
192: signals $x(t)$ and $y(t)$, time (delay) information between the two signals can
193: be obtained from the phase spectrum, which is the argument of the
194: cross-spectrum \cite{ml1,hal1},
195: $$ \widehat{\Phi(\omega)}~=~arg\{\widehat{S_{xy}(\omega)\}}. \eqno 2 $$
196: Following \cite{ml1}, eq. 2 can be further simplified to see the explicit
197: appearance of the time delay in it as follows,
198: $$\widehat{\Phi(\omega)}~=~\omega \delta, \eqno 3$$
199: where $\delta$ is the time delay.
200:
201: The phase estimate $\widehat{\Phi(\omega)}$ and its upper and lower 95\%
202: confidence interval are given by
203: \cite{hal1}
204: $$\widehat{\Phi(\omega)}~\pm 1.96~\left[\frac1{2M}\left(\frac1{\widehat{C(\omega)}}-1\right)
205: \right]^\frac12.
206: \eqno 4 $$
207: Thus, the confidence interval of the phase estimate is inversely related to coherence.
208: In the rest of the paper we use estimates of the coherence and phase without
209: the over hat for the sake of convenience.
210:
211: One of the conventional ways to estimate a time delay, in frequency domain, is to fit a straight
212: line to the phase spectrum (eq. 3) in the frequency band of significant
213: coherence, as the phase can be reliably estimated only in the frequency
214: band of significant coherence (see eq. 4). This method of estimating delay
215: is possible when we have a broad band coherence and limits its applicability
216: to the narrow band coherent signals. In some cases, coherence extends to first
217: harmonic and hence the phase values at the harmonics are used to increase the
218: reliability of the delay estimate \cite{pb1}.
219:
220: Further, if two time series show significant coherence over a wide
221: range of frequency band, but have a minimal phase relation, then the
222: estimation of time delay from the phase estimate is not straight
223: forward. Under such conditions, phase estimate, in addition to
224: the $\omega \delta $ term (see eq. 3), will also contain a
225: frequency dependent factor, namely, the argument of the transfer function,
226: ${\rm{arg}}~
227: A(\omega)$. In such cases, ${\rm{arg}}~A(\omega)$ is estimated by Hilbert
228: transform, and then delay is estimated from the phase estimate after
229: subtracting the estimated transfer function from the phase estimate \cite{ml2}.
230: From the above discussion it is clear that phase estimates cannot be used to
231: estimate the
232: delay from narrow band coherent signals which are often observed in human
233: physiological data \cite{hel1,jan1}. For this purpose, we
234: use the method of maximising coherence.
235:
236: As discussed in the last part of the introduction, a delay $\delta$
237: between two
238: time series, will cause a reduction in the estimated coherence between
239: them. In order to compensate for the reduction in coherence due to delay and
240: thereby to estimate the delay, we realign the time series by artificially
241: shifting them. For this, we shift one of the time series by a lag $(\tau)$
242: keeping the other constant and consider coherence at the selected
243: frequency band $\omega_0$ as a function of $\tau$, $C(\tau)_{\omega_0}$ and repeat the same for the other time series. If there is a
244: delay between the two time series, coherence in the selected frequency band
245: $C(\tau)_{\omega_0}$ will increase from its value at $\tau=0$ and reach a maximum at the $\tau$ corresponding to
246: $\delta$. After shifting one of the time series by a lag $\tau$, the length of
247: the shifted time series will be
248: less than the un-shifted
249: time series by $N-\tau$ ($\tau$ in sample units) data points. We discard the extra data points in
250: the un-shifted time series to have the same length for both time series. Further,
251: coherence is a relative measure and changes with the length of the data.
252: It is reflected in the 99\% confidence level (see above). For each shift,
253: we discard the points corresponding to the maximum lag from the entire length of the
254: time series, and consider only the length of the data which is integer
255: multiple of $L$. By doing so, coherence $C(\tau)_{\omega_0}$ calculated for
256: all the lags will have
257: the same confidence level, thereby confirming that the maximum
258: value of coherence is reached only because of the delay and not due to the
259: spurious effects caused in estimating coherence for different lengths
260: of data. As we have confidence level as a check for the
261: reliability of
262: $\delta$, we don't need any additional tests like, surrogate tests
263: \cite{kantz} to assess the reliability of the
264: results. However, in order to get the variability in the estimated delay, we
265: use surrogate analysis.
266:
267: Surrogate analysis is introduced in the context of nonlinear time series
268: analysis to check whether or not the time series under consideration has got
269: a nonlinear structure \cite{kantz,theiler}. Making the (null) hypothesis that
270: the time series has come from a linear process, several
271: linear realizations of
272: the time series namely, surrogates, are synthesised. Then, the original time
273: series and the surrogates are quantified by a suitable
274: discriminating statistic. Any deviation in the discriminating statistic
275: calculated for the
276: original time series and surrogates indicates the
277: presence of nonlinear structure in the original time series and thereby
278: rejecting the null
279: hypothesis \cite{theiler}. There are different ways of preparing
280: the surrogates \cite{kantz,schreiber}. The
281: misconception
282: of surrogate analysis is discussed in \cite{jeti2}. As mentioned above,
283: the objective of the methods to generate surrogates \cite{kantz,schreiber} is to
284: synthesise a new data set by
285: destroying the nonlinear structure present in the original data set. But our
286: aim is not to show the presence of nonlinear structure in our data. For
287: the purpose of calculating the error-bars of delay, surrogates are generated by exploiting one of the basic
288: assumptions of spectral analysis that distinct parts of the time series are
289: independent \cite{hal1}. Instead of shuffling the time series as a whole (which
290: is done in one of the methods to synthesise surrogates, amplitude adjusted surrogates) we
291: shuffle the disjoint data segments from which the original spectrum is
292: estimated (see above). We shuffle only the un-shifted time series (see above)
293: from which the information is assumed to flow to the other time series (which
294: is time shifted in advance in maximising coherence analysis) with a delay. By
295: doing so, the spectra of both the time series,
296: $\widehat{S_{xx}(\omega)}$ and $ \widehat{S_{yy}(\omega)}$ in
297: eq. 1 will remain the same but the cross spectrum $\widehat{S_{xy}(\omega)}$
298: in eq. 1 will be different. This type of surrogate is similar to the one
299: proposed in \cite{schreiber} where the whole time series is shuffled but by
300: preserving two point correlation (i.e. power spectrum) of the original time
301: series. For all the analyses reported in this paper we synthesise 19
302: different realizations of surrogates. We make a null hypothesis that
303: $C(\tau)_{\omega_0}$ obtained is due to spurious correlations between the two
304: time series. For each realization of surrogate we calculate time
305: delayed coherence $C(\tau)_{\omega_0}^{surr}$. We calculate the significance
306: of difference $S(\tau)$ between the $C(\tau)_{\omega_0}^{surr}$ calculated for
307: surrogates and $C(\tau)_{\omega_0}$ where $S(\tau)
308: =\frac{\left|C(\tau)_{\omega_0}-<C(\tau)_{\omega_0}^{surr}>
309: \right|}{\sigma[C(\tau)_{\omega_0}^{surr}]}$, where $<.>$ indicates the average
310: over different realizations of surrogates and $\sigma[.]$ indicates the standard deviation
311: between different realizations of surrogates. Any value of $S(\tau) > 2 $
312: indicates that the
313: $C(\tau)_{\omega_0}$ obtained is not due to spurious correlations
314: \cite{theiler} and hence the null hypothesis is rejected.
315: If the
316: null hypothesis is rejected we consider $C(\tau)_{\omega_0}$ for further analysis
317: otherwise we discard it as spurious correlation. For the
318: $C(\tau)_{\omega_0}$ qualified in the surrogate analysis (for which null
319: hypothesis is rejected) we calculate the
320: error in the delay in the following way: we subtract $C(\tau)_{\omega_0}^{surr}$
321: calculated for each realization from $C(\tau)_{\omega_0}$ and calculate delay
322: for each subtracted function (i.e.)
323: $C(\tau)_{\omega_0}-C(\tau)_{\omega_0}^{surr}$. We report here the delay of
324: the system as the mean value of the delays (calculated for the 19 surrogate
325: subtracted realizations) and their standard deviation as error-bar. As the
326: increase in the $C(\tau)_{\omega_0}$ when compensated for the delay is very
327: small, for the sake of clarity we plot $C'(\tau)_{\omega_0} = \left[
328: C(\tau)_{\omega_0}-<C(\tau)_{\omega_0}^{surr}>\right]
329: -\left[C(\omega_0)-<C(\omega_0)^{surr}>\right]$. Note that $C(\omega_0)$ is
330: the value of $C(\tau)_{\omega_0}$ at $\tau=0$ and $<C(\omega_0)^{surr}>$ is
331: the average of $C(\tau)_{\omega_0}^{surr}$ for different realizations of the
332: surrogates at $\tau=0$. By this definition $C'(\tau)_{\omega_0}$ will pass
333: through a zero value at $\tau=0$ and show a maximum value (above zero) at
334: $\tau=\delta$ (for $\delta \ne 0$) between the processes. Though we calculate $S(\tau)$ for different
335: $\tau$ values at which $C(\tau)_{\omega_0}$ is evaluated, we report here
336: $S(\tau=\delta)$ as it is the most relevant one in determining the
337: significance of the $C'(\tau=\delta)_{\omega_0}$ (and hence $C(\tau=\delta)_{\omega_0}$). Based on the above arguments
338: the delay between the two time series is given by:
339: $$\delta = {\max_\tau}~C'(\tau)_{\omega_0}.$$
340:
341: \section{Application to coupled R\"ossler attractors}
342: In this section, we apply the method of maximising coherence, to
343: coupled R\"ossler systems. The dynamics of the $i$-th attractor is given as
344: follows:
345:
346: \begin{eqnarray}
347: \dot{x}_i(t) &=& -[y_i(t)-z_i(t)]~+~\epsilon_{j,i}~x_j(t-\delta) \nonumber \\
348: \dot{y}_i(t)&=&x_i(t)~+~a~y_i(t) \nonumber \\
349: ~\dot{z}_i(t) &=&b~+~z_i(t)~[x_i(t)~+~c] \nonumber ,
350: \end{eqnarray} where $a~=~0.38,~b~=~0.3,~c~=~4.5$, are the
351: parameters. Coupling is established through the $x$ component between the
352: attractors. $\epsilon_{j,i}$, is the coupling constant and $x_j$ is the $x-$
353: component of the $j-$th attractor. In our study, we consider the dynamics of two
354: coupled R\"ossler attractors, and hence $i~=~1$ and $j~=~2$ and $\delta$ is
355: the delay. Further, we don't consider the self coupling of the oscillators.
356: \begin{figure}[H]
357: \begin{center}
358: \includegraphics[width=2in]{f1.eps}
359: \caption{Coupling scheme of the attractors 1 and 2. $\epsilon_{2,1}=0.16$ and
360: $\epsilon_{1,2}=0$. $\delta$ is set to 2 sec.}
361: \label{fig1}
362: \end{center}
363: \end{figure}
364: First, we consider the coupling scheme as shown in Fig. 1, where two
365: R\"ossler attractors are coupled and the information flow is from
366: attractor 2 to attractor 1 ($\epsilon_{2,1}=0.16$ and $\epsilon_{1,2}=0$) as
367: governed by the above set of equations. $\delta$ is set to 2
368: sec. The above equations are simulated by Euler scheme with a step size of
369: 0.01 sec and for the subsequent analysis the data are down sampled to 0.1
370: sec. As the coupling is established through the $x-$ component between the
371: attractors, we use the $x-$ time series to perform the delay estimation. We
372: have used 30000 data points for further analysis.
373:
374: Power spectra of the two attractors are shown in Fig 2a. Solid line
375: represents the power spectrum of attractor 1 and solid line with
376: dots represents the power spectrum of attractor 2. For the sake of
377: clarity the spectrum of attractor 1 is shifted vertically upwards by a factor
378: of 5 from
379: its original position. We have used a segment length
380: of $L=1000$ and hence the frequency resolution (of the quantities in eq.
381: 1) is 0.01 $Hz$. Spectra show dominant activities in the frequency range
382: of 0.1 to 0.2 $Hz$, which
383: can be related to the mean orbital period of $\sim5$ sec of the
384: attractor \cite{nar1,wolf}. Since the dynamics of the attractor 1 is
385: perturbed by coupling (see Fig. 1), its spectrum looks slightly
386: different from that of the attractor 2. The coherence spectrum of the
387: coupled attractors is shown in Fig. 2b. Horizontal line around the coherence of 0.15 indicates
388: the confidence level of $\alpha=0.99$.
389:
390:
391: In the frequency band between 0.1-0.2 $Hz$ (see Fig. 2a), where the
392: individual attractors show dominant activities, the coupled systems show
393: significant coherence. Estimated phase (eq. 2) between the coupled
394: attractors is shown in Fig. 2c. Error-bars represent the 95\% confidence
395: interval
396: of the phase estimate given by eq. 4. Thus the phase can be reliably
397: estimated in the frequency band (0.1-0.2 $Hz$), where the coupled
398: attractors show a significant coherence (see Fig. 2b). The result of
399: maximising coherence is shown in Fig. 2d where $C'(\tau)_{\omega_0}$ is
400: plotted as a function of lag $\tau$. We have chosen $\omega_0 = 0.13 Hz$
401: (and is shown in Fig. 2b) from the basic frequency of the oscillators. Solid line at
402: $C'(\tau)_{\omega_0}=0$ indicates $C(\omega_0)-<C(\omega_0)^{surr}> $(see the
403: last part of the methodology for details).
404: Negative
405: shift indicates that the time series of attractor 1 is in advance and positive
406: shift indicates that the time series of attractor 2 is in advance.
407: Since there is a delay of 2 sec, the coherence calculated at $\tau=0$ is
408: zero (see above). In order to compensate for the reduction in the coherence and
409: thereby to estimate the time delay, time series of the attractors are shifted
410: by a time lag $\tau$. As the delayed information flows from attractor 2 to
411: attractor 1, coherence $C'(\tau)_{\omega_0}$ increases from zero (see above) when the time series of attractor 1 is
412: shifted \textbf{back} in time and reaches a maximum at the $\tau=-2.1$ which is the
413: time delay between the two time series ($\delta$ used in the simulation is 2
414: sec). The significance of deviation from the surrogate data $S(-2.1)$ is
415: 2.24 and indicates that
416: $C'(\tau=-2.1)$ is not due to spurious correlations.
417: The error-bar for the delay (as
418: explained in the methodology section) is calculated as 0.4 sec. Thus, the
419: time delay $ \delta $ is $-2.1 \pm 0.4 $ sec and is well captured by the
420: maximising coherence analysis (expected delay value is 2 sec for the flow from
421: attractor 2 to attractor 1).
422: \begin{figure}[H]
423: \begin{center}
424: \includegraphics[width=5in,height=4.5in]{f2.eps}
425: \caption{(a). Solid line represents power spectrum of attractor 1 and
426: solid line with dots represents power spectrum of attractor 2. For the sake
427: of clarity the power spectrum of
428: attractor 1 is shifted vertically upwards by a factor of 5 from its original
429: position. (b). Estimated
430: coherence spectrum of the coupled R\"ossler attractors with
431: $\delta=2 $ sec. Horizonal line at the coherence of 0.15 indicates the confidence level
432: ($\alpha=0.99$). There is a significant coherence in the
433: frequency range of 0.1-0.2 $Hz$ between the two attractors. The value of
434: $\omega_0$ used for maximising coherence analysis is indicated along with
435: its
436: coherence value $C(\omega_0)$. (c). Phase estimate
437: $\Phi(\omega)$ of the coupled attractors. Error-bars represent
438: 95\% confidence interval defined by eq. 4. Errors are relatively small in
439: the region of significant coherence which is in accordance to eq.
440: 4. (d). Coherence $C'(\tau)_{\omega_0}$, as the function of lag
441: $\tau$. Horizontal line at $C'(\tau)_{\omega_0}=0$ indicates
442: $C(\omega_0)-<C(\omega_0)^{surr}>$. $<.>$ represents the average over
443: different surrogates. $C'(\tau)_{\omega_0}$ has reached the maximum value at the $\tau$ (delay) value of -2.1 sec
444: (from attractor 2 to 1), and the expected value of $\delta$ is 2 sec for
445: the flow from attractor 2 to attractor 1. The significance of deviation
446: from the surrogate $S(-2.1)$ is 2.24 and indicates that $C'(\tau)_{\omega_0}$ is not due to spurious
447: correlations.}
448: \label{fig2}
449: \end{center}
450: \end{figure}
451: Next, we consider the coupling scheme as shown in Fig. 3, where two R\"ossler
452: attractors are coupled as considered above, but with a bi-directional flow,
453: $\epsilon_{2,1}=0.15$ and $\epsilon_{1,2}=0.1$. $\delta$ used in the coupling
454: is same as in Fig. 1, which is 2 sec.
455: \begin{figure}[H]
456: \begin{center}
457: \includegraphics[width=2in]{f3.eps}
458: \vskip 0in
459: \caption{Coupling scheme of the attractors 1 and 2. $\epsilon_{2,1}=0.15$ and
460: $\epsilon_{1,2}=0.1$. $\delta$ is set to 2 sec.}
461: \label{fig3}
462: \end{center}
463: \end{figure}
464:
465: Power spectra of the attractors look
466: almost the same as the two attractors perturb each other almost to the same
467: extent (see Fig. 4a).
468: As obtained for uni-directional coupling, in this case
469: also, significant coherence is between 0.1-0.2 $Hz$ (Fig. 4b). As the number
470: of data points used is the same (30000) as in the uni-directional coupling,
471: the
472: confidence level $\alpha=0.99$ is also around the coherence of 0.15 (see the
473: horizontal line in Fig. 4 b which is same as in Fig. 2b). Phase estimate of
474: the bi-directionally coupled systems is shown
475: in Fig. 4c with the 95 \% confidence interval as the error-bars. Error-bars are
476: relatively narrow (see eq. 4) in the frequency band of 0.1-0.2 $Hz$ where
477: significant coherence is observed (as seen in Fig. 2c). Result of the
478: maximising coherence for bi-directionally coupled systems is shown in
479: Fig. 4d. The horizontal line at $C'(\tau)_{\omega_0}=0 $ has the meaning as
480: in Fig. 2d. In this case
481: 0.15 $Hz$ is used as $\omega_0$ (and is shown in Fig. 4b) which is the basic
482: frequency of the attractors. $C'(\tau)_{\omega_0}$ shows two maxima, one at $\tau=-2.5$ sec which
483: corresponds to the flow from attractor 2 to 1 and another at $\tau= 1.7$ sec
484: which corresponds to the flow from attractor 1 to 2, while the $\delta$ used
485: in the coupling is 2 sec in both directions (see Fig. 4d). Also, in this case,
486: the significance of deviation from surrogate calculated by $S(\tau)$ at $\tau$
487: values -2.5 sec and 1.7 sec are well above 2 (see Fig. 4d) indicating that the
488: $C'(\tau)_{\omega_0}$ obtained at these two $\tau$ values are not due to
489: spurious correlations.
490: Thus, the time delays of
491: the system are $-2.5 \pm 0.5 $ sec (for the flow from attractor 2 to
492: attractor 1)
493: and $1.7 \pm 0.4 $ sec (for the flow from attractor 1 to attractor 2) and are
494: well within the expected value of 2 sec in either direction (see Fig. 3). The
495: error-bars are obtained by surrogate analysis.
496: \begin{figure}[H]
497: \begin{center}
498: \includegraphics[width=5in,height=4.5in,angle=0]{f4.eps}
499: \caption{(a-c). Explanations are as in Fig. 2 but for bi-directionally coupled
500: R\"ossler attractors, $\epsilon_{2,1}=0.15$ and
501: $\epsilon_{1,2}=0.1$. (d). Result of maximising coherence
502: $C'(\tau)_{\omega_0}$.
503: $C'(\tau)_{\omega_0}$ has reached
504: maximum value at two places, one at -2.5 sec (from attractor 2 to
505: 1) and another at 1.7 sec (from attractor 1 to 2), where
506: $\delta$ is set to 2 sec in both directions. The significance of
507: deviation from surrogate at $S(-2.5)$ is 6.60 and at $S(1.7)$ is
508: 3.25 and indicates that $C'(\tau)$ at $\tau$ values -2.5 sec and 1.7 sec
509: are not due to
510: spurious correlations. The horizontal line at $C'(\tau)_{\omega_0}=0 $
511: has the same meaning as in Fig. 2d.}
512: \label{fig4}
513: \end{center}
514: \end{figure}
515: Results of the theoretically coupled systems are summarised in Fig. 5. For the
516: unidirectional flow, shown in Fig. 1, which is referred to as scenario 1 in
517: Fig. 5, the value of the delay obtained by maximising coherence, shown by '$
518: \ast$' with the error-bars (calculated by surrogate analysis)
519: shown as vertical line, is close to the expected value of -2 sec (shown as dot in
520: Fig. 5). Similarly for the bi-directional flow (see Fig. 3) which is referred
521: to as scenario 2 in Fig. 5, the values of the delay obtained by maximising
522: coherence with their error-bars (vertical lines) are close to the expected
523: delay
524: of 2 sec in either direction (see Fig. 3). Compared to the unidirectional
525: flow, the results of the bi-directionally coupled oscillators show slightly
526: more deviations from the expected value of 2 sec but when we include the
527: error-bar, they are close to the expected value. So when dealing with the
528: biological systems we should consider the delay always with the error-bar
529: before we draw any conclusion from it. For the uni-directionally coupled
530: oscillators there is no significant change in the result when $\omega_0$ is chosen as
531: 0.19 $Hz$ where there is a significant coherence between the two oscillators. Further for the above studied systems
532: similar results can be obtained by
533: cross-correlation analysis and by the time delayed phase synchronisation analysis
534: as well \cite{diego}.
535: \begin{figure}[H]
536: \begin{center}
537: \includegraphics[width=4in,height=3in,angle=0]{f5.eps}
538: \caption{Results of delay estimation by maximising coherence for
539: theoretically coupled systems. Negative delay indicates the flow from
540: attractor 2 to attractor 1 and the positive delay indicates the flow from
541: attractor 1 to attractor 2. Scenario 1 represents uni-directionally
542: coupled oscillators (flow from attractor 2 to attractor 1). Scenario 2
543: represents bi-directionally coupled oscillators. Expected delay values
544: are shown by dots. Delay values obtained by maximising coherence are shown
545: by '$\ast$'. Vertical lines are the error-bars of delay (see text for
546: details).}
547: \label{fig5}
548: \end{center}
549: \end{figure}
550:
551: \section{Application to EEG and EMG data}
552: Tremor is an involuntary, periodic movement of the parts of the body
553: \cite{deuschl}. It can be classified as normal and pathological tremor
554: depending on amplitude and the conditions under which tremor is activated
555: \cite{deuschl}. Essential tremor is a common movement disorder characterised
556: mostly by postural tremor of the arms. Other neurological abnormalities are
557: typically
558: absent in essential tremor \cite{findley}. It has been shown by experimental
559: studies in animals
560: \cite{llinas,lamarre} and by Positron emission tomography
561: or functional magnetic resonance imaging studies
562: \cite{hallet,jenkins,bucher} on human
563: beings, that different parts of the brain are involved in essential tremor.
564: The correlation between the thalamic activity and forearm electromyogram
565: provides a direct evidence for the involvement of the thalamus in the
566: tremor oscillations \cite{hua}. Further, control of essential tremor by
567: stereo-tactic lesions and high frequency stimulation of the ventrolateral
568: thalamus adds support for the involvement of thalamus in the
569: mechanism of essential tremor \cite{benabid,schuurman}. As the cerebellum has
570: its main outputs to thalamus which in turn has strong projections to
571: cortex, it is hypothesised that essential tremor arise from an oscillating
572: cerebello-thalamo-cortical loop the output station of which is the motor
573: cortex \cite{hel1}.
574: \begin{figure}[H]
575: \begin{center}
576: \includegraphics[width=5in,height=4.5in,angle=0]{f6.eps}
577: \caption{Results of spectral analysis and delay estimation for
578: subject 1. (a). Solid line represents power spectrum of C4 of EEG
579: (the central electrode overlaying the primary motor cortex, PMC) and solid line with dots
580: represents the power spectrum of extensor muscle
581: with a frequency resolution of 1 $Hz$. For the sake of clarity, the
582: power spectrum of EEG is shifted vertically upwards from its original
583: position by a factor of 10. (b). Coherence spectrum of C4
584: and extensor muscle. The horizontal line at the coherence of 0.024
585: indicates the confidence level of $\alpha=0.99$. There is a
586: significant coherence at the (tremor) frequency of 5
587: $Hz$ and at 10-15 $Hz$ as well. The value of $\omega_0$ used for the
588: maximising coherence is shown along with its coherence value $C(\omega_0)$. (c). Phase estimate of the C4 and
589: extensor muscle with 95\% confidence interval as the error-bars. Error-bars are narrow at the
590: significantly coherent frequency bands of 5 $Hz$ and 10-15 $Hz$ as well. (d). Coherence as a
591: function of $\tau$, $C'(\tau)_{\omega_0}$. It has reached maximum
592: at $\tau=16$ $ms$. The horizontal line at $C'(\tau)_{\omega_0}=0 $ has
593: the same meaning as in Fig. 2d. The significance of deviation
594: from the surrogate is given as $S(16) = 13.01$.}
595: \label{fig6}
596: \end{center}
597: \end{figure}
598: Cortico-muscular coherence for essential tremor has been studied earlier \cite{halliday} but has
599: not been ascertained at the tremor frequency in
600: \cite{halliday}. We show below that there is a significant
601: cortico-muscular coherence when the signal to noise ratio of the tremor
602: increases beyond a certain threshold value as seen in
603: \cite{hel1,hellwig2}.
604: Since the paper is aimed at time delay estimation, the detailed
605: results and discussion of the spectral analysis of the tremor study will be
606: reported elsewhere \cite{jan1}.
607:
608: For this purpose we consider five patients (with mean age 66 years and
609: standard deviation 5.1 years), which showed postural essential tremor of the
610: arms. In a
611: dimly lighted room, patients are asked to sit on a comfortable chair in a
612: slightly supine position with both hands held against gravity while the
613: forearms are supported. EEG is
614: recorded with a 64-channel EEG system (Neuroscan) with standard electrode
615: positions \cite{klem}. Surface EMG is recorded
616: from EMG electrodes attached to wrist flexors and extensors of both
617: arms. For all the subjects EEG and EMG are sampled
618: at a rate of 1000
619: $Hz$. EEG and EMG are bandpass filtered online, respectively, between
620: 0.01-200 $Hz$ and 30-200 $Hz$. Data are stored in a computer and are analysed offline. In all the cases EMG is recorded on both
621: hands. Each recording lasted for 1-4 minutes. Artifacts like base line shift,
622: eye blinks etc. are discarded by visual inspection.
623:
624: Before further analysis, EMG signals are full wave rectified (magnitude of
625: the deviations from mean)
626: and EEG is made reference free by constructing the second (spatial) derivative,
627: (Laplacian) \cite{hj1,hj2}. We discuss in detail for one of the subjects,
628: subject 1,
629: which we feel typical and for the remaining subjects we summarise the results in a
630: table.
631:
632: For subject 1, the power spectra of EEG (solid line) and EMG (solid line with dots), see
633: Fig. 6a, are
634: calculated with $L=1000$ and hence the frequency resolution is 1 $Hz$ (since
635: $sf=1000 $ $Hz$). In Fig. 6a, we show the power spectra of one of
636: the EEG channels (C4) in
637: the contralateral (right) side to tremor (left) hand and the rectified EMG
638: (extensor muscle) of the same hand. Power spectrum of EEG (shown as solid
639: line in Fig. 6a) is shifted vertically upwards by a factor of 10 from its initial
640: position for the sake of
641: clarity. Power spectrum of EEG shows dominant
642: activity in a broad band between 8-20 $Hz$ containing the bands of 10-13 $Hz$
643: and 15-20 $Hz$, which correspond, respectively, to the alpha and beta activity
644: of the brain \cite{guy}. Muscle spectrum (solid line with dots in Fig. 6a) shows tremor activity around 5
645: $Hz$. Correlation between the muscle and C4 (the central electrode
646: overlaying the primary motor cortex, PMC) of EEG cannot be guessed
647: from the EEG spectrum as there is no significant activity around 5 $Hz$ (tremor
648: frequency). Correlation becomes clear when we consider coherence. Figure 6b
649: shows coherence between C4 and extensor muscle. The horizontal line around the
650: coherence of 0.024 indicates the confidence level of $\alpha=0.99$. There is a
651: significant coherence (and hence correlation) between C4 and the tremor
652: activity reflected by the muscle \cite{hel1,hellwig2,mima1,mima2,mima3,hal3}
653: and also in the frequency band of 10-15 $Hz$.
654: Phase estimate of this system is given in Fig. 6c with $95 \%$ confidence interval as
655: error-bars. It is clear from Fig. 6c that phase can be reliably estimated at
656: the (tremor and the significant coherence) frequency of 5 $Hz$ (confidence
657: interval is narrow) and also in the frequency band of 10-15 $Hz$. In
658: an earlier study \emph{Hellwig et al.} \cite{hellwig2} investigated time
659: delays based on the value of the phase estimate $\widehat{\Phi(\omega)}$ at
660: the \textbf{tremor frequency}. But the results obtained \cite{hellwig2} are
661: not in the interpretable level. The reason postulated in \cite{hellwig2} that
662: the delay might be modulated by frequency dependent mechanism is hardly
663: conceivable in a narrow (single) frequency band at which phase estimate is
664: significant. The reason may be due to the weak assumption of the model
665: $\widehat{\Phi(\omega)} = \omega\delta + \theta$, where $\theta$ is the
666: indispensable constant \cite{mima1,mima2}, phase shift between the two
667: processes which is ignored in
668: \cite{hellwig2}. However, we show below that the delay estimated by maximising
669: coherence is reasonably in good agreement with the experimentally observed
670: conduction velocity and is in physiologically interpretable range.
671:
672: Application of the maximising coherence method to real life data like EEG is
673: not
674: straight forward. Since the maximising coherence method is aimed at to capture
675: the time delay (by making up for the reduction in the coherence) by
676: artificially shifting (local dynamics) the time series, it is very sensitive
677: to non-stationarities in the time series. However, in order to make coherence
678: analysis robust against non-stationarities within the time series, we use the
679: following way to discard the non-stationary parts of the data:
680: We calculate autocorrelation function
681: \cite{kaplan}, in each of the segments (of length 1000 samples). Then the
682: de-correlation time, time at which autocorrelation function falls to $exp(-1)$
683: (as the first value of autocorrelation function is 1 by definition)
684: \cite{kaplan} is calculated for all the segments. Non-stationarities due to
685: base line shift, eye blinks will introduce artificial trends (lifting up of the
686: mean value of the signal) \cite{chen} in the signal, which will have higher de-correlation
687: time than the stationary parts of the signal. In order to discard the
688: non-stationary parts, about 50 \% of the median of
689: the de-correlation time calculated in all the segments is taken as threshold
690: and parts of the EEG and EMG data with de-correlation time greater
691: than the above defined threshold are discarded in the delay estimation. After
692: discarding
693: the non-stationary parts of the data, the segments left out on either side of
694: the removed parts are stitched as if they are recorded continously. Of course,
695: care should be taken in applying the above method. The above method will
696: capture only the non-stationary parts if it is applied to the data of low
697: signal to noise ratio which can be avoided by visual inspection of the data.
698:
699: Result of the delay estimation between C4 and muscle, is shown in Fig.
700: 6d. The horizontal line at $C'(\tau)_{\omega_0}=0$ in Fig. 6d has the same
701: meaning as in Fig. 2d.
702: Here, we have used $\omega_0=5$ $Hz$ (see Fig. 6b), where
703: significant coherence is observed. Negative shift indicates that the time series
704: of EEG is in advance. In Fig. 6d, $C'(\tau)_{\omega_0}$ reaches a
705: maximum value at $\tau=16$ $ms$, indicating the cortex (time series of EMG
706: is in advance for positive values of $\tau$) is driving the
707: tremor. The significance of deviation from the surrogate is given by
708: $S(16)= 13.01$ and indicates that $C'(\tau)_{\omega_0}$ obtained is not
709: due to spurious correlations.
710: The error-bar obtained by surrogate analysis is 7 $ms$ and hence the delay
711: obtained
712: for C4 is 16$\pm$7 $ms$ which is in good agreement with the experimentally
713: observed value of around 15 $ms$ \cite{roth}. Comparison of Fig. 6d with
714: Fig. 2d,
715: shows that the information flow is in one direction (i.e.) from
716: cortex to muscle. But in other subjects, in (an electrode from)
717: this region, we have also observed a bi-directional flow, a situation
718: similar to Fig. 4d.
719: \begin{figure}[H]
720: \begin{center}
721: \includegraphics[width=5.5in]{f7.eps}
722: \caption{(a). A part of the 64-channel arrangement on the surface
723: of the scalp used to record EEG. Electrodes on the left hand side
724: are odd numbered and those on the right hand side are even
725: numbered. (b). Isocoherence map of the subject 1, constructed by
726: considering the coherence at the tremor frequency (5 $Hz$ in this
727: case). Magnitude of the coherence is indicated by the gray scale-bar
728: juxtaposed to the isocoherence map. Parts of the scalp which
729: showed significant coherence are coloured black and the
730: confidence level of $\alpha=0.99$ is marked with white colour (which is
731: set to zero, see text for details). (Electrodes O5-O6 are just
732: displayed and are not used for coherence analysis).}
733: \label{fig7}
734: \end{center}
735: \end{figure}
736: Of the 64-channels of EEG recorded, coherence and delay estimation are
737: explained in detail for C4, as it is overlaying the primary
738: sensorimotor cortex which is involved in motor control of the contralateral
739: hand affected by essential tremor. For the rest of the electrodes (see Fig.
740: 7a) we give the results of the coherence analysis as isocoherence
741: (region of scalp with same coherence) map. Electrodes on the boundary
742: of the scalp are not considered for the analysis as they had low signal
743: to noise ratio. Further, we have not used the electrodes
744: O5 - O6 for the study. To construct isocoherence map, we have considered
745: the coherence $C(\omega)$ at the (tremor) frequency of 5 $Hz$ for all the
746: thirty five channels (see above) shown in Fig. 7a, from F5 (top left of
747: scalp, Fig. 7a), to P6 (bottom right of scalp, Fig. 7a). We have
748: subtracted the confidence level ($\alpha=0.99$) from the coherence so
749: that the confidence level for coherence is zero. In Fig. 7b,
750: surface of the scalp exhibiting significant coherence is marked with black
751: colour and the coherence of the parts which do not cross the confidence
752: level of $\alpha=0.99$ (which are set to zero after the above
753: normalisation) are marked with white colour. In Fig. 7b, C4 shows maximum
754: coherence, (see
755: the gray scale-bar juxtaposed to Fig. 7b). Thus, there is a hot spot
756: (region of significant coherence) in the PMC region. Similar results have
757: been obtained in the spectral study of essential tremor
758: \cite{hel1,hellwig2}.
759:
760: There is another hot spot for subject 1 in the frontal region of the
761: scalp (FCZ). The detailed discussion of this will be dealt with in
762: \cite{jan1}.
763: \begin{figure}[H]
764: \begin{center}
765: \includegraphics[width=4in,height=5.5in,angle=270]{f8.eps}
766: \caption{Results of delay estimation by maximising coherence for (a). C2
767: (b). C6 (c). CP2 (d). CP4 (e). CP6 of the central hot spot and (f). FCZ of the
768: frontal hot spot. The horizontal line at $C'(\tau)_{\omega_0}=0$ in Fig. 8
769: (a-f) has the same meaning as in Fig. 2d. The
770: significance of deviation from the surrogate $S(\tau)$
771: at $\tau=\delta$ is given in all the plots.
772: %(Results of surrogate analysis are given in insets (see
773: % Fig. 1 for details of the inset).
774: Estimated values of delay by maximising
775: coherence are (c) -28 $ms$ (d) -2 $\sim$ 0 $ms$ (e) -30 $ms$ and 31 $ms$ (f)
776: 25 $ms$. As the maximum of $C'(\tau)_{\omega_0}=25~ms$ in Fig. 8f is not clearly
777: discernable (see text for details) it is magnified in the inset shown in
778: Fig. 8f. Negative shift indicates that the time series of EEG is in
779: advance and positive shift indicates that the time series of EMG is in
780: advance. A negative delay indicates the flow from muscle to EEG and a
781: positive delay indicates the flow from EEG to muscle. C2 and C6 do not
782: qualify in the surrogate analysis as their $S(\tau=\delta)$ values lie
783: below 2
784: indicating that the values of $C'(\tau)_{\omega_0}$ (see Fig. 8a and 8b) obtained for these
785: two electrodes are due to
786: spurious correlations and therefore are not considered for further analysis. }
787: \label{fig8}
788: \end{center}
789: \end{figure}
790: Spectral and delay estimations for C4 and extensor
791: muscle have been analysed in detail showing a uni-directional
792: flow from PMC to muscle, with a delay of 16 $ms$. For the rest of the
793: electrodes (C2, C6, CP2, CP4, CP6) of the
794: central hot spot and FCZ of the frontal hot spot, the results of delay
795: estimation are given in Fig. 8. The significance of deviation from surrogate
796: $S(\tau)$ at $\tau=\delta$ is given in all the plots, see Fig. 8(a-f). For C2
797: (Fig. 8a) and C6 (Fig. 8b) the value of $S(\tau=\delta)$ is below two
798: indicating that the values of $C'(\tau)_{\omega_0}$ obtained for these two
799: electrodes are
800: due to spurious correlations.
801: Therefore we do not consider them for further analysis. For CP2,
802: CP4, CP6 and FCZ, $S(\tau=\delta)$ values obtained are above 2 (see Fig. 8(c-f))
803: indicating that $C'(\tau)_{\omega_0}$ values obtained for these electrodes are not due to spurious correlations.
804: Delay values exhibited by
805: $C(\tau)_{\omega_0}$ shown in Fig. 8(c-f) are CP2 -28 $\pm$ 3.3 $ms$ (see
806: Fig. 8c), CP4, -2 $\sim$ 0 $ms$ (see Fig. 8d), CP6 -30 $\pm$ 7.4 and 31 $\pm$ 13.5 $ms$ (see
807: Fig. 8e) and FCZ 25 $\pm$ 10.4 $ms$ (see Fig. 8f). The maximum in
808: $C'(25)_{\omega_0}$ for FCZ shown in Fig. 8f is not clearly discernable and
809: hence this region is magnified in the inset shown in Fig. 8f where the maximum at
810: $\tau=25~ms$ can be seen without any ambiguity. The large
811: fluctuations in $C'(\tau)_{\omega_0}$ for FCZ (Fig. 8f) and C4 (Fig. 6d)
812: indicate larger uncertainty in the delay estimated by this method and is
813: one of the shortcomings
814: of this method. However, this uncertainty, to some extent is revealed by the
815: error-bars (see above) obtained by the surrogate analysis. A positive value of
816: $\tau$
817: indicates that the time series of muscle is in advance and a negative
818: value of $\tau$ indicates that the time series of EEG is in
819: advance. Error-bars are obtained by surrogate analysis as discussed in the
820: methodology section. A zero delay for CP4 (see Fig. 8d) may be due to the
821: nullification of two strong counter acting forces (drive from PMC to muscle
822: and opposing drive from muscle to PMC, a situation similar to phase
823: locking in coupled oscillators). The positive values of estimated delays including the error-bars are well in agreement with the
824: experimentally observed value of 15 $ms$ for forearm muscles \cite{roth}. For
825: the negative delay (flow from muscle to EEG), it is at least partly in keeping
826: with normal median Somatosensory Evoked Potential (SEP) latencies of around 20 $ms$.
827:
828: Further, in order to make the delay statistically significant we report the
829: delay of the hot spot as a whole as the delay of the system, by weighting the delay obtained for each
830: electrode (when it is qualified in surrogate analysis) by a fraction of the coherence
831: $C(\omega_0)$ value that each electrode contributes to the hot spot. By doing
832: so, electrodes showing higher tremor related correlations will
833: contribute more to the delay of the system. For subject 1, there is a hot
834: spot (central hot spot) in the PMC region of scalp and another in the
835: frontal region of the
836: scalp (i.e.) FCZ (see Fig. 7b). Of the electrodes contributing to the central
837: hot spot, we do not consider C2 and C6 (as they do not qualify in the surrogate analysis) for further analysis. We
838: weight the delay obtained for rest of the electrodes contributing to the
839: central hot spot by $NC^{(l)}(\omega_0)=\frac{C^{(l)}(\omega_0)}{\sum_{j=1}^KC^{(j)}(\omega_0)}$, where $K$
840: is the number of electrodes contributing to hot spot after qualifying the
841: surrogate analysis, and $[100 \cdot NC^{(l)}(\omega_0)]$ \% is the contribution of
842: the $l-$th electrode to the hot spot (it is
843: easy to check that $\sum_{j=1}^KNC^{(j)}(\omega_0)=1$). Similarly
844: $NC^{(l)}(\omega_0)$ is calculated for each hot spot. Further, in the calculation of $NC^{(l)}(\omega_0)$
845: for positive delay, only those electrodes which display positive delay are
846: considered and similarly for negative delay, only those electrodes which display
847: negative delay are considered. For subject 1, only one electrode, FCZ
848: contributes to the frontal hot spot, for which $NC^{(l)}(\omega_0)=1$. We
849: have normalised the error-bars of the delays in a similar way. The final
850: values of delay for both the hot spots are given in Table 1 along with the delay
851: values obtained by maximising coherence analysis for four other subjects.
852: \begin{table}[H]
853: \begin{center}
854: \begin{tabular}{|c|c|c|c|c|}\hline
855: & \multicolumn{4}{c|} {Delay $\delta$ ($ms$) obtained by maximising coherence}\\ \cline{2-5}
856: \rb{Subject} & \multicolumn{2}{c|} {Central hot spot} & \multicolumn{2}{c|
857: }
858: {Frontal hot spot}\\ \cline{2-5}
859: & EEG to Muscle & Muscle to EEG & EEG to Muscle
860: & Muscle to EEG\\ \hline
861: 1 & $12.0~\pm~6.6$ & $-16.2~\pm~2.6$ & $25.0~\pm~10.4$ & - \\ \hline
862: 2 & $14.0~\pm~4.4$ & $-9.0~\pm~4.4$ & $23.0~\pm~7.9$ &$-10.0~\pm~2.3$ \\ \hline
863: 3 & $11.0~\pm~3.7$ & $-9.0~\pm~3.8$ & $27.3~\pm~12.7$ & - \\ \hline
864: 4 & $16.0~\pm~7.7$ & $-24.0~\pm~9.0$ & - & - \\ \hline
865: 5 & $11.34~\pm~6.1$ & $-15.0~\pm~3.8$ & - & - \\ \hline
866: \end{tabular}
867: \caption{Summary of delay estimation by maximising coherence for five
868: essential tremor subjects. Positive and negative signs are introduced in
869: maximising coherence analysis just to
870: denote the directions of information flow and have got no real life
871: significance.}
872: \end{center}
873: \end{table}
874: Of the five subjects considered three show an additional hot spot in
875: the frontal region the delay values of which are also given in Table 1 (see third
876: column). As can be seen there is also a cortico-muscular delay between the frontal hot
877: spot and the periphery in all three cases. However, this delay is about double
878: the cortico-muscular delay from the central hot spot. These results indicate
879: that both the central area most likely corresponding to the primary
880: sensorimotor cortex and more frontal area are involved in the
881: generation of essential tremor. While the delays from the central area
882: (PMC) between 11 and 16 $ms$ are in keeping with a
883: direct transmission through fast conducting pyramidal pathways \cite{roth} the
884: cortico-muscular delay from the more frontal area rather indicates another way
885: of interaction with the periphery. Thus the
886: method of maximising coherence is capable of distinguishing different types of
887: cortico-muscular interactions which are relevant to the understanding of the
888: pathophysiology of essential tremor, for details see \cite{jan1} . The
889: musculo-cortical delays partly agree but are generally slightly lower than the
890: somatosensory conduction delays known from routine SEP studies and most likely reflect feed back of the peripheral tremor to the cortex.
891: \section{Conclusion}
892: We have used the method of maximising coherence to obtain the time delay between
893: two series. As a test case, we applied this method to uni- and
894: bi-directionally coupled R\"ossler attractors. In both cases, the delays
895: estimated by maximising coherence match well with the expected values,
896: within the error limits which are obtained by surrogate analysis. For EEG-EMG
897: time series
898: of essential tremor subjects, this method yields a delay in the range of
899: 11 to 16 $ms$ for the primary sensorimotor area and 23-27 $ms$ for the more
900: frontal area involved in the tremor oscillation whereas the experimentally observed value of the conduction time between primary cortex and muscle is 15 $\pm$ 2 $ms$
901: \cite{roth}. The larger delay observed for the frontal area may indicate a
902: different, possibly indirect interaction with the periphery. The musculo-cortical delays (9 to 24 $ms$) are partly in
903: keeping with the delay observed (around 20 $ms$) in SEP studies. One of the
904: reasons for the slight deviation of the delay from experimentally
905: observed value may be as follows: We have employed maximising
906: coherence method to capture the delay between two series, assuming that
907: there will be a continuous delayed flow of information from one time series to another. But, it may not be the case in biological data like EEG and EMG. Intermittently the flow may be absent. One way to check out this is to divide the time series into small
908: sub-portions and calculate the time delay in each portion. But coherence
909: analysis needs at least a minimum of 30000 data points to decide statistically
910: whether or not the two time series under study are linearly correlated. So we need to
911: analyse the nature of information flows using the nonlinear methods
912: like extended Granger causality \cite{govind}, transfer
913: entropy \cite{schreiber2}. This issue will be addressed in our future
914: work. Of course other methods like cross-correlation analysis and time delayed
915: phase synchronisation analysis \cite{diego} can be used to estimate the delay
916: between the EEG and EMG
917: of the essential tremor subjects. But the problems in using cross-correlation
918: are addressed in \cite{ml1} and references therein. There exists an analytical
919: expression for the confidence limit for coherence estimate but this is lacking for
920: phase synchronisation used in \cite{diego}. That
921: is the reason why phase synchronisation is not commonly used for
922: correlation analysis (subsequently for delay analysis in this work) though
923: this method is expected to be
924: superior to coherence analysis \cite{shay}. However comparison of the different methods of
925: delay estimation is beyond the scope of the current paper.
926:
927: RBG wishes to acknowledge Dr. Jens Timmer for discussions on surrogate
928: analysis. Financial support from Deutsche Forschungsgemeinschaft (German
929: Research Council) is gratefully acknowledged.
930:
931: % The Appendices part is started with the command \appendix;
932: % appendix sections are then done as normal sections
933: % \appendix
934:
935: % \section{}
936: % \label{}
937:
938: % Bibliographic references with the natbib package:
939: % Parenthetical: \citep{Bai92} produces (Bailyn 1992).
940: % Textual: \citet{Bai95} produces Bailyn et al. (1995).
941: % An affix and part of a reference:
942: % \citep[e.g.][Ch. 2]{Bar76}
943: % produces (e.g. Barnes et al. 1976, Ch. 2).
944:
945: \begin{thebibliography}{99}
946:
947: % \bibitem[Names(Year)]{label} or \bibitem[Names(Year)Long names]{label}.
948: % (\harvarditem{Name}{Year}{label} is also supported.)
949: % Text of bibliographic item
950:
951: \bibitem{ml1}
952: T.~M\"uller, M.~Lauk, M.~Reinhard, A.~Hetzel, C.H.~L\"ucking, J.~Timmer, Annals
953: of Biomedical Engineering, 31 (2003) 1423.
954: \bibitem{ml2}
955: M.~Lindemann, J.~Raethjen, J.~Timmer, G.~Deuschl, G.~Pfister,
956: J. Neurosci. Meth., 111 (2001) 127.
957: \bibitem{carter}
958: G.C.~Carter, Proc. IEEE, 75 (1987) 236.
959: \bibitem{roth}
960: J.C.~Rothwell, P.D.~Thompson, B.L.~Day, S.~Boyd, C.D.~ Marsden,
961: Exp. Physiol., 76 (1991) 159.
962: \bibitem{hel1}
963: B.~Hellwig, S. H\"au\ss ler, B.~Schelter, M.~Lauk, G.~Guschlbauer, J.~Timmer,
964: C.H.~ L\"ucking, Lancet, 357 (2001) 519.
965: \bibitem{diego}
966: D.~Rybski, S.~Havlin, A.~Bunde,
967: Physica A, 320 (2003) 601.
968: \bibitem{hal1}
969: D.M.~Halliday, J.R.~Rosenberg, A.M.~Amjad, P.~Breeze, B.A. Conway,
970: S.F. Farmer, Prog. Biophys. molec. Biol., 64 (1995) 237.
971: \bibitem{jeti1}
972: J.~Timmer, M.~Lauk, G.~Deuschl, Electroencephalogr. Clin. Neurophysiol., 101
973: (1996) 461.
974: \bibitem{pb1}
975: S.~Salenius, S.~Avikainen, S.~Kaakkola, R.~Hari, P.~Brown, Brain, 125 (2002)
976: 491.
977: \bibitem{jan1}
978: J. Raethjen, R.B. Govindan, F. Kopper, G. Deuschl, Brain (in preparation).
979: \bibitem{kantz}
980: Nonlinear Time Series Analysis, edited by H.~Kantz and T.~Schreiber (Cambridge
981: University Press, 1997).
982: \bibitem{theiler}
983: J.~Theiler, A.~Longtin, B.~Galdrikan, J.D.~Farmer, Physica D, 58 (1992) 77.
984: \bibitem{schreiber}
985: T.~Schreiber, A.~Schmitz, Phys. Rev. Lett., 77 (1996) 635.
986: \bibitem{jeti2}
987: J.~Timmer, Phys. Rev. Lett., 85 (2000) 2647.
988: \bibitem{nar1}
989: K.~Narayanan, R.B.~Govindan, M.S.~Gopinathan, Phys. Rev. E, 57 (1998) 4594.
990: \bibitem{wolf}
991: A.~Wolf, J.B.~Swift, L.~Swinney, A.~Vastano, Physica D, 16 (1985) 285.
992: \bibitem{deuschl}
993: G.~Deuschl, P.~Bain, M.~Brin, Scientific committee, Mov. Disord., 13 (1998) 2.
994: \bibitem{findley}
995: L.J.~Findley, W.C.~Koller, Neurology, 37 (1987) 1194.
996: \bibitem{llinas}
997: R.~Llin$\rm{\acute{a}}$s, R.A.~Volkind, Exp. Brain Res., 18 (1973) 69.
998: \bibitem{lamarre}
999: Y.~Lamarre, Animal models of physiological, essential and parkinsonian-like
1000: tremors. Editors L.J.~ Findley and R.~Capildeo, Movement
1001: disorders:tremor. (Macmillan Press, London, 1984) 183.
1002: \bibitem{hallet}
1003: M.~Hallett, R.M.~Dubinsky, J.~Neurol. Sci., 114 (1993) 45.
1004: \bibitem{jenkins}
1005: I.H.~Jenkins, P.G.~Bain, J.G.~Colebatch, P.D. Thompson, L.J. Findley,
1006: R.S.J. Frackowiak, C.D. Marsden, D.J. Brooks, Ann. Neurol., 34 (1993) 82.
1007: \bibitem{bucher}
1008: S.F.~Bucher, K.C.~Seelos, R.~Dodel, M.~Reiser, W.H. Oertel, Ann. Neurol, 41
1009: (1998) 273.
1010: \bibitem{hua}
1011: S.E.~Hua, F.A.~Lenz, T.A.~Zirh, P.M.~Dougherty, J. Neurosurg Psychiatry, 64
1012: (1998) 273.
1013: \bibitem{benabid}
1014: A.L.~Benabid, P. Pollak, C. Gervason, D. Hoffmann, D.M. Gao, M. Hommel,
1015: J.E. Perret, J. Derougemont, Lancet, 337 (1991) 403.
1016: \bibitem{schuurman}
1017: P.R.~Schuurman, D.A.~Bosch, P.M.M. Bossuyt, G.J. Bonsel, E.J.W. van Someren, R.M.S. de Bie, M.P. Merkus, J.D. Speelman, N. Engl. J. Med., 342 (2000)461.
1018: \bibitem{halliday}
1019: D.M.~ Halliday, B.A.~Conway, S.F.~Farmer, U.~Shahani, A.J.C.~Russell,
1020: J.R.~Rosenberg, Lancet, 355 (2000) 1149.
1021: \bibitem{hellwig2}
1022: B. Hellwig, S.~H\"au\ss ler, M.~Lauk, B.~Guschlbauer, B.~K\"oster,
1023: R.~Kristeva-Feige, J.~Timmer, C.H.~L\"ucking, Clin. Neurophysiol., 111 (2000)
1024: 806.
1025: \bibitem{klem}
1026: G.H. Klem, H.O. Jasper, H.H. Elgerl. The ten-twenty electrode system of the
1027: International Federation. Editors G. Deuschl and A. Eisen, In. Recommendations
1028: for the practice of clinical
1029: neurophysiology: Guidelines of the International Federation of Clinical
1030: Neurophysiology, Electroencephalogr. Clin. Neurophys. Suppl., 52 (1999) 3.
1031: \bibitem{hj1}
1032: B.~Hjorth, Electroencephalogr. Clin. Neurophysiol., 39 (1975) 526.
1033: \bibitem{hj2}
1034: B.~Hjorth, J. Clin. Neurophysiol., 8 (1991) 391.
1035: \bibitem{guy}
1036: A.C.~Guyton, Text Book of Medical Physiology, Editors M.J.~Wonsiewicz, (Prism
1037: Books (PVT) Ltd., Bangalore, India, 1991), 659.
1038: \bibitem{mima1}
1039: T.~Mima, M.~Hallett, J. Clin. Neurophysiol., 16 (1999) 501.
1040: \bibitem{mima2}
1041: T.~Mima, M.~Hallett, Clin. Neurophysiol., 110 (1999), 1892.
1042: \bibitem{mima3}
1043: T.~Mima, J.~Steger, A.E.~Schulman, C.~Gerloff, M.~Hallett, Clin. Neurophysiol.,
1044: 111 (2000) 326.
1045: \bibitem{hal3}
1046: D.M.~Halliday, B.A.~Conway, S.F.~Farmer, J.R.~Rosenberg, Neurosci. Lett., 241
1047: (1998) 5.
1048: %\bibitem{peter}
1049: %P.~Tass, M.G.~Rosenblum, J.~Weule, J.~Kurths, A.~Pivoksky, J.~Volkmann,
1050: %A.~Schnitzler, H.-J.~Freund, Phys. Rev. Lett., 81 (1998) 3291.
1051: \bibitem{kaplan}
1052: D.~Kaplan and L.~Glass, Understanding Nonlinear Dynamics, Editors
1053: J.E.~Marsden, L.~Sirovich, M.~Golubitsky and W. J\"ager (Springer-Verlag, New
1054: York, 1995).
1055: \bibitem{chen}
1056: Z.~Chen, P.Ch.~Ivanov, K.~Hu, H.E.~Stanley, Phys. Rev. E, 65 (2002) 041107.
1057: \bibitem{govind}
1058: Y.~Chen, G.~Rangarajan, J.~Feng, M.~Ding, Phys. Lett. A, 324 (2004) 26.
1059: \bibitem{schreiber2}
1060: T.~Schreiber, Phys. Rev. Lett., 85 (2000) 461.
1061: \bibitem{shay}
1062: A.~Gozolchiani, S.~Moshel, J.M. Hausdorff, E.~Simon, J.~Kurths, S.~Havlin,
1063: arXiv.org e-print cond-mat/0410617.
1064: \end{thebibliography}
1065:
1066: \end{document}
1067:
1068:
1069:
1070:
1071:
1072:
1073: