1: %\documentstyle[preprint,aps]{revtex}
2: \documentstyle[aps,twocolumn,epsfig]{revtex}
3: %\def\btt#1{{\tt$\backslash$#1}}
4: %\def\BibTeX{\rm B{\sc ib}\TeX}
5:
6: \begin{document}
7: \draft
8:
9: \title {Scaling Properties of Fluctuations in\\
10: Human Electroencephalogram}
11:
12: \author{Rudolph C. Hwa}
13: \address{Institute of Theoretical Science and Department of
14: Physics\\ University of Oregon, Eugene, OR 97403-5203, USA}
15:
16: \author{Thomas C. Ferree\cite{byline}}
17:
18: \address{Electrical Geodesics, Inc., Riverfront Research Park,
19: Eugene, OR 97403}
20:
21: \date{\today}
22: \maketitle
23:
24:
25: \begin{abstract} The fluctuation properties of the human
26: electroencephalogram (EEG) time series are studied using
27: detrended fluctuation analysis. For all 128 channels in each of
28: 18 subjects studied, it is found that the standard deviation of
29: the fluctuations exhibits scaling behaviors in two regions.
30: Topographical plots of the scaling exponents reveal the spatial
31: structure of the nonlinear electrical activities recorded on the
32: scalp. Moment analyses are performed to extract the gross features
33: of all the scaling exponents. The correlation between the two
34: scaling exponents in each channel is also examined. It is found that
35: two indices can characterize the overall properties of the
36: fluctuation behaviors of the brain dynamics for every subject and
37: that they vary widely across the subjects.
38:
39: \end{abstract}
40:
41: \pacs{PACS numbers: 05.45.Tp, 87.19.La, 87.90.+y}
42:
43: \section{Introduction}
44:
45:
46: The scalp electroencephalogram (EEG) provides a wealth of
47: information about human brain dynamics. The complex nature
48: of brain dynamics results in a high degree of fluctuations in
49: both the spatial and temporal aspects of the EEG signals. To
50: extract the salient properties from the data is the primary
51: objective of any method of analysis. We present in this paper
52: a novel method that explores the scaling behavior of the
53: fluctuations and uses moment analysis to reduce the complexity
54: of the results obtained.
55:
56: The most common methods of EEG time series analyses are
57: event-related time ensemble averaging and Fourier decomposition,
58: both of which are based implicitly on assumptions of
59: linearity \cite{eeg,nunez}. Since the physiological
60: mechanisms underlying the scalp EEG are generally nonlinear,
61: they can generate fluctuations that are not best described
62: by linear decomposition. Moreover, the resting EEG always
63: displays a broad-banded power spectrum, so in Fourier
64: analysis one must arbitrarily define frequency bands
65: ($\delta, \theta, \alpha, \cdots$) which may not actually
66: delineate different dynamical mechanisms. Wavelet analyses
67: have also been applied to examine EEG time series \cite{bla},
68: but at a sacrifice of the ability to describe long-range
69: temporal correlations. Chaos analyses have been applied
70: to quantify the nonlinear behavior of the brain
71: \cite{chaos,jan,leh}, but typically require a long period
72: of time to compute attractor properties for a single time series.
73: Moreover, chaos-based approaches assume the existence of
74: low-dimensional attractors, and this is probably not generally a
75: valid assumption for the brain. In this paper, we discuss a
76: method that analyzes the fluctuations in EEG over a short
77: period of time (around 10s), and makes use of the information
78: conveyed by all 128 channels on the scalp. We show the
79: existence of scaling behaviors of certain measures of
80: fluctuations in all channels and in all subjects. We
81: propose two global measures of the spatio-temporal signals
82: that characterize the distinctive nature of EEG for each subject.
83:
84:
85: The study of scaling behavior emphasizes the relationship
86: across time scales. We aim to find what is universal among
87: all channels as well as what varies among them. The former
88: is obviously important by virtue of its universality for a
89: given subject; how that universal quantity varies from
90: subject to subject is clearly interesting. What varies
91: from channel to channel is perhaps even more interesting,
92: since it has implications for describing focal features
93: which could have functional or clinical relevance.
94:
95: Our procedure is to focus initially on one channel at a time.
96: Thus it is a study of the local temporal behavior and the
97: determination of a few parameters (scaling exponents) that
98: effectively summarize the fluctuation properties of the time
99: series. The second phase of our procedure is to describe the
100: global behavior of all channels and to arrive at two numbers that
101: summarize the variability of these temporal measures across the
102: entire scalp surface. This dramatic data reduction necessarily
103: trades detail for succinctness, but such reduction is exactly
104: what is needed to allow easy discrimination between brain
105: states.
106:
107: The emphasis in this paper is on the method of analysis. It is not
108: our aim here to perform clinical and cognitive analyses. Due to
109: the fact that the EEG data available to us are short in time
110: duration and few in the number of subjects studied, it is not
111: feasible for us to make reliable inference on the physiological
112: implications of our findings. Nevertheless, the data are
113: sufficient for the extraction of interesting behaviors that are
114: channel dependent as well as subject dependent.
115:
116: \section{Detrended Fluctuation Analysis}
117:
118: The specific method we use in the first phase is detrended
119: fluctuation analysis (DFA). This analysis is not new. It was
120: proposed for the investigation of correlation properties in
121: non-stationary time series and applied to the studies of
122: heartbeat \cite{peng} and DNA nucleotides \cite{pbh}. It has
123: also been applied to EEG \cite{wat}, but with somewhat different
124: emphases than those presented here. Since the analysis
125: considers only the fluctuations from the local linear trends,
126: it is insensitive to spurious correlations introduced by
127: non-stationary external trends. By examining the scaling
128: behavior one can learn about the nature of short-range and
129: long-range correlations.
130:
131: Let an EEG time series be denoted by $y(t)$, where $t$ is
132: discrete time ranging from 1 to $T$. Divide the entire range of
133: $t$ to be investigated into $B$ equal bins, discarding any
134: remainder, so that each bin has $k={\rm floor}(T/B)$ time points.
135: Within each bin, labeled $b\ (b=1,\cdots, B)$, perform a least-square
136: fit of $y(t)$ by a straight line, $\overline y_b(t)$, i.e.,
137: $\overline y_b(t)=$ Linear-fit$[y(t)]$ for $(b-1)k<t\leq bk$.
138: That is the semi-local trend for the $\it b$th bin. Define
139: $F_b^2(k)$ to be the variance of the fluctuation $y(t)$ from
140: $\overline y_b(t)$ in the $b$th bin, i.e.,
141: \begin{equation}
142: F_b^2(k) = {1\over k} \sum_{t=(b-1)k+1}^{bk} [y(t)-\overline
143: y_b(t)]^2
144: \label{(1)}
145: \end{equation}
146: It is a measure of the semi-locally detrended fluctuation in bin $b$.
147: The average of $F_b^2(k)$ over all bins is
148: \begin{equation}
149: F^2(k)={1\over B}\sum_{b=1}^B\,F_b^2(k).
150: \label{(2)}
151: \end{equation}
152: $F(k)$ is then the RMS fluctuation from the semi-local trends in
153: $B$ bins each having $k$ time points.
154:
155: The study of the dependence of $F(k)$ on the bin size $k$ is the
156: essence of DFA \cite{peng,pbh}. If it is a power-law behavior
157: \begin{equation}
158: F(k) \propto k^{\alpha} , \label{(3)}
159: \end{equation}
160: then the scaling exponent $\alpha$ is an
161: indicator of the correlations of the fluctuations in EEG, which
162: depends on the relationship of these fluctuations across time
163: scales. Since DFA considers only the fluctuations from
164: the semi-local linear trends, it is insensitive to spurious correlations
165: introduced by non-stationary external trends. This is a practical
166: advantage since EEG acquisition systems often suffer from slow
167: drifts associated with gradual changes in the quality of electrode
168: contact to the skin. The analysis also liberates our result
169: from the dependence on the overall magnitude of the voltage
170: $y(t)$ recorded by each probe, which is an advantage since
171: overall signal amplitude can vary across subjects,
172: presumably due to differences in skull conductivity and other factors.
173:
174: Resting EEG data were collected for 18 subjects using a
175: 128-channel commercial EEG system, with scalp-electrode
176: impedences ranging from 10 to 40 k$\Omega$. The data were
177: hardware filtered between 0.1 and 100 Hz, then digitized at
178: 250 points/sec. After acquisition, $T\approx 10 $s lengths of
179: simultaneous time series in all channels are chosen, free of
180: artifacts such as eye blink and head movements. At each time
181: point, the average across all electrodes was subtracted, to
182: remove approximately the effect of the reference electrode
183: \cite{nunez}. We investigate the range of $k$
184: from 3 to 500 in approximately equal steps of $\ln k$.
185:
186: In Fig.\,1 we show three typical time series $y(t)$ in three
187: widely separated channels for subject A, labeled 1-3, for
188: brevity. While it is clear that both channels 2 and 3
189: have substantial 10 Hz oscillations after 0.2s, it is much
190: less apparent that there exist any scaling behaviors in all three
191: channels. The corresponding values of $F(k)$ are shown in the
192: log-log plot in Fig.\,2. Evidently, the striking feature is
193: that there are two scaling regions with a discernible bend when
194: the two slopes in the two regions are distinctly different.
195: With rare exceptions this feature is found in all channels for
196: all subjects. Admittedly, the extents of the scaling regions
197: are not wide, so the behavior does not meet the qualification
198: for scaling in large critical systems or in fractal geometrical
199: objects. However, since the behavior is so universal across
200: channels and subjects, and since the temporal scales involved are
201: physiologically relevant, the scaling behavior is a feature of
202: EEG that conveys an important property of the brain activity and
203: should not be overlooked.
204:
205:
206: \section {Scaling and Nonscaling Properties}
207:
208: The fact that there exist two scaling regions suggests that the
209: lack of scaling in the region between the two implies the
210: existence of some significant time scale. From Fig.\ 1 one indeed
211: sees roughly periodic oscillations in Ch.\ 2 and 3. One may
212: therefore be tempted to think that if, instead of considering the
213: fluctuations from the linear semi-local trends $\bar y_b(t)$, one
214: studies the fluctuations from periodic oscillations, then the
215: bend would disappear and the two scaling regions might be joined
216: to become one. However, even if that were true, such a procedure
217: should not be used for two reasons. First, not all channels
218: exhibit obvious oscillatory behaviors with definite frequencies.
219: Channel 1 in Fig.\ 1 is one such example. Whatever detrending one
220: chooses should be universally applied to all channels in order to
221: avoid introducing discrepencies across the channels due to
222: external intervention. Second, to determine the frequency of the
223: oscillatory trend requires a Fourier analysis, which is precisely
224: what our approach attempts to circumvent. To decide on a sinusoidal
225: wave of a particular frequency as reference for detrending
226: involves arbitrariness and is unlikely to lead to any simplification
227: in the global picture. The simplest and unbiased approach is to use
228: the semi-local linear trends, as we have done.
229:
230: To quantify the scaling behavior, we perform a linear fit in
231: Region I for $1<{\ln}k < 2.5$ and denote the slope by
232: $\alpha_1$, and similarly in Region II for $3.5 < {\ln}k < 5.75$ with
233: slope denoted by $\alpha_2$. Visual inspection for each of the
234: 18 subjects verifies that fitting this way does a remarkably
235: good job of characterizing the slopes in the two regions.
236: Knowing the two straight lines in each channel allows us to
237: determine the location of their intercept, $\ln\kappa$, which
238: gives a good approximation for the position of the bend in
239: ${\ln}k$. We find that, whereas $\alpha_1$ and $\alpha_2$ can
240: fluctuate widely from channel to channel, $\kappa$ is limited to
241: a narrow range in most subjects. The average value of ln$\kappa$
242: for each subject ranges from 2.6 to 3.6, with a grand average
243: across subjects to be approximately 3.1. It should, however, be
244: noted that when $\alpha_1$ and $\alpha_2$ are nearly the same, as
245: is the case for Ch.\ 1 in Fig.\ 2, the determination of $\kappa$
246: by the intersection of the two straight lines is not reliable.
247: Nevertheless, it is visually clear that the bend occurs in the
248: vicinity of ln$\kappa = 3.1$.
249:
250: Since scaling behavior means that the system examined has no
251: intrinsic scale, scale noninvariance at $\kappa$ implies that
252: $\kappa$ is related to a dominant frequency of oscillation in the
253: time series. It is at this point that a contact can be made with
254: the usual Fourier analysis. Although our analysis focuses on scale
255: invariant quantities, i.e., the dimensionless scaling exponents, it
256: is worth digressing momentarily to establish this contact. To do
257: this, we loosely associate the time scale $\kappa$ with the period
258: of a sine wave with frequency $f$. If the data acquisition rate is
259: denoted by $r$, then the frequency $f$ corresponding to $\kappa$
260: is
261: \begin{equation}
262: f=r/\kappa. \label{(4)}
263: \end{equation}
264: For our data acquisition we have $r=250$ points/sec.
265: For the across-subject average of ln$\kappa = 3.1$, we get from
266: Eq.\ (4) $f=11.3$ Hz. That is in the middle of the traditional
267: $\alpha$ (8-13 Hz) EEG frequency band. Thus the dominant periodic
268: oscillation apparent in Fig.\ 1 does reveal itself in the study of
269: the scaling behavior. If one's interest is in the frequency
270: content of the EEG time series, then Fourier analysis is more
271: direct. However, if the interest is in the fluctuation properties
272: and their relationships across time scales, then DFA is
273: more effective. Hereafter, frequency will play no essential role
274: in the remainder of this paper.
275:
276: For each subject we have 128 pairs of values of $(\alpha_1,
277: \alpha_2)$, which summarize the temporal fluctuations in terms
278: of scaling exponents. In Fig.\ 3, we exhibit by scatter plot
279: the values of $\alpha_1$ and $\alpha_2$ of all channels for
280: subject A. The three points marked with circles correspond
281: to the channels shown in Fig.\ 2. The error bars in Fig.\ 3
282: indicate the goodness of fit of the two regions by straight
283: lines. Since the variability of the scaling exponents across
284: channels is large compared to the error in fit, the different
285: values convey numerically meaningful information. For this subject,
286: Region I exhibits better scaling behavior than Region II, although
287: the error bars on $\alpha_2$ are not so large as to call into
288: question the power-law description in Region II.
289:
290: Overall, for subject A, the scaling exponents are in the ranges:
291: $0.19<\alpha_1<1.44$ and $0.018<\alpha_2<0.489$. Whereas
292: $\alpha_1$ is widely distributed, $\alpha_2$ is sharply peaked
293: at 0.1 and has a long tail. The value of $\alpha=0.5$
294: corresponds to random walk with no correlation among the various
295: time points. For $\alpha\neq 0.5$ there are correlations: Region
296: I corresponds to short-range correlation, Region II long-range,
297: with $\kappa$ giving a quantitative demarkation between the two.
298: In most channels we find $\alpha_1 > \alpha_2$, although there
299: are a few where $\alpha_1 \approx \alpha_2$. The scatter plots
300: of all other subjects are similar in general features to the one
301: shown in Fig.\ 3, but vary in detail from subject to subject.
302: It is impractical to show them all in this paper. Evidently, it
303: is desirable to find a way to quantify succinctly these 128 pairs
304: of numbers so that one can effectively compare the results
305: across subjects.
306:
307: A scatter plot such as Fig.\ 3 reveals very well how the
308: $\alpha_i$ exponents of all the channels are related to one
309: another. However, it shows nothing about the locations of the
310: channels on the scalp. To show that, we use the topographical
311: plots of $\alpha_1$ and $\alpha_2$ separately, as in Fig.\ 4, to
312: exhibit the spatial structure of the signals extracted. The
313: dissociation of $\alpha_1$ from $\alpha_2$ is a price paid to
314: gain the spatial perspective on the scalp. Topographic plots
315: such as this may be useful for specifying the location
316: of focal features, e.g., associated with particular brain
317: functions and/or pathologies. Thus topographical and scatter plots
318: present different aspects of the fluctuation properties of the
319: brain electrical activity of any given subject. Both are
320: inefficient for comparison across subjects. What we need is a
321: global measure that describes the general, overall feature of all
322: $\alpha_i$ pairs in the form of a single parameter.
323:
324:
325: \section {Moments of the Scaling Exponents}
326:
327: We propose to consider the moments of the scaling exponents. In
328: general, if we have $N$ numbers, $z_j, j=1,\cdots, N,$ we can
329: calculate the moments
330: \begin{equation}
331: G_q = {1\over N} \sum_{j=1}^N z_j^q , \label{5}
332: \end{equation}
333: where $q$ is a positive integer \cite{gar}. The information
334: contained in the first $N$ moments (i.e., $q=1,\cdots,N$) is
335: enough to reproduce all the $z_j$ by inversion. However, we may be
336: interested in only a few of the $G_q$ with lower order $q$, each of which
337: contains some information of all the $z_j$. In our
338: present problem we have
339: $N=128$, and we shall consider the first ten orders, $1\le q\le
340: 10$. That is a significant reduction of numbers, a process worth
341: investigating.
342:
343: Before calculating the moments of $\alpha_i$, let us see how
344: those values are distributed. Let $x$ be either $\alpha_1$ or
345: $\alpha_2$. Since no value of $\alpha_i$ has been found to exceed
346: 1.5 in the subjects we have examined, we consider the interval
347: $0\leq x\leq 1.5$. Divide that interval into $M$ equal cells,
348: which for definiteness we take to be $M=150$ here. Let the cells
349: be labeled by $m=1,\cdots, M$, each having the size $\delta
350: x=1.5/M$. Denote the number of channels whose
351: $x$ values are in the $m$th cell by $n_m$. Define
352: \begin{equation}
353: P_m = n_m\,/\,N. \label{6}
354: \end{equation}
355: It is the fraction of channels whose $x$ values
356: are in the range $(m-1)\,\delta x \leq x < m\,\delta x$. By
357: definition, we have $\sum_{m=1}^M\,P_m=1$. In Fig.\,5 we show as
358: an illustration the two graphs of $P_m$ for subject A. The two
359: graphs correspond to $\alpha_1$ and $\alpha_2$, and are, in
360: essence, the projections of the scatter plot in Fig.\,3 onto the
361: $\alpha_1$ and $\alpha_2$ axes. From Fig.\,5 we see that
362: $\alpha_1$ is widely distributed, while $\alpha_2$ is not. $G_1$
363: gives the average, and $G_2$ is related to the width.
364:
365: Since the fluctuation of $m$ in $P_m$ should be measured relative to
366: its mean, let us consider the normalized moments \cite{rch}
367: \begin{equation}
368: M_q^{(i)}=G_q^{(i)}\left/ \left(G_1^{(i)}\right)^q\right.=
369: \sum_{m=1}^M m^q P_m^{(i)} \left/
370: \left(\sum_{m=1}^M m P_m^{(i)}\right)^q\right. ,
371: \label{7}
372: \end{equation}
373: where $i=1$ or 2.
374: Since these moments are averages of $(m/\overline m)^q$, where
375: $\overline m$ is the average-$m$, they are not very sensitive to
376: $\overline m$ itself. They contain the essence of the
377: fluctuation properties of $\alpha_{1,2}$ in all channels.
378: In terms of the scaling exponents explicitly, let us
379: use $\alpha_i(j)$ to denote the value of $\alpha_i$
380: for channel $j$ so that Eq.(\ref{7}) may be rewritten as
381: \begin{equation}
382: M_q^{(i)}={1\over N}\sum_{j=1}^N
383: \alpha_i(j)^q\, \left/ \left({1\over N}\sum_{j=1}^N
384: \alpha_i(j)\right)^q . \right. \label{7.5}
385: \end{equation}
386:
387: In principle, it is possible to examine also the moments for
388: $q<0$, which would reveal the properties of $P_m$ at low values
389: of $m$. However, the accuracy of our data is not too reliable
390: for low-$k$ analysis, since the 60 Hz noise due to ambient
391: electric and magnetic fields has not been cleanly filtered out.
392: In this paper, therefore, we restrict our study to only the
393: positive $q$ values. For high $q$, the large $m/\overline m$
394: parts of $P_m^{(1,2)}$ dominate $M_q^{(1,2)}$.
395:
396: In Fig.\,6 the $q$-dependences of $\ln M_q^{(1,2)}$ are shown
397: for the distributions exhibited in Fig.\,5 for
398: $2\leq q\leq 10$. They are approximately linear except for the
399: low values of $q$. The same type of dependencies on $q$ are found
400: for all subjects. In Fig.\,6 we show two straight lines that can
401: fit very well the nearly linear behaviors of ln
402: $M_q^{(i)}$ vs $q$ for $q \ge 5$. Thus for large $q$ we have
403: \begin{equation}
404: M_q^{(i)} \propto {\rm exp}\ (\mu_i\,q), \qquad q\ge 5. \label{8}
405: \end{equation}
406: The linear extrapolations of the lines to lower values of $q$ show
407: the degree of deviation of the the calculated values of
408: $\ln M_q^{(1,2)}$ from linearity. Since ln$M_q^{(1)}$ and
409: ln$M_q^{(2)}$ behave so similarly in their departures from the
410: linear dependencies on $q$, we plot ln$M_q^{(2)}$ vs
411: ln$M_q^{(1)}$ in order to exhibit their direct relationship without
412: explicit dependence on $q$. We find that they are linearly related
413: over a wider range of values. This linearity is found to be true for
414: all subjects. The plots for three of them are illustrated in
415: Fig.\,7, where the straight lines are the linear fits. Thus the
416: implication is that there exists a universal power-law behavior
417: \begin{equation} M_q^{(2)} \propto
418: \left(\,M_q^{(1)}\,\right)^{\eta}
419: \label{9}
420: \end{equation}
421: valid for all subjects examined. From Eqs.(\ref{8}) and (\ref{9}) we
422: obtain
423: \begin{equation}
424: \eta = \mu_2 / \mu_1 , \label{10}
425: \end{equation}
426: but now $\eta$ is meaningful for all $q$ (except for the lowest
427: points) and in that sense independent of $q$. Thus we have
428: discovered a global measure $\eta$ that characterizes all
429: $\alpha_i$ values of a subject, and varies from subject to
430: subject. We postpone the display of all the
431: $\eta$ values for all subjects until later.
432:
433: To understand the exponential behavior in Eq.(\ref{8}), we note
434: that $G_q$ is dominated by large $z_j$ when $q$ is large, as is
435: self-evident in Eq.(\ref{5}). For asymptotically large $q$, we
436: have $G_q\propto {\rm exp}\ (q \,{\rm ln}\,z_{\rm max})$, where
437: $z_{\rm max}={\rm max}\,\{z_j\}$. For intermediate $q$, all large
438: values of $z_j$ can make important contributions, and the
439: exponential dependence on $q$ can still prevail. The denominator
440: in Eq.(\ref{7}) is $G_1^q=\bar z^q={\rm exp}\ (\,q\,{\rm ln} \bar
441: z\,)$, where $\bar z$ is the average of $z_j$, so it is also
442: exponential for any $q$. It is therefore clear that Eq.(\ref{8})
443: follows, and that $\mu_i$ depends on all $z_j$ with more weight on
444: the large $z_j$ values. The power-law behavior of Eq.(\ref{9})
445: implies that the exponent $\eta$ is independent of $q$ and that
446: all $\alpha_i$ values are relevant contributors to the universal
447: behavior. This is an important point worth emphasizing: the
448: independence of $\eta$ on $q$ implies that the whole spectra of
449: $\alpha_1$ and $\alpha_2$ are summarized by the one index $\eta$.
450: The fact that $\eta$ varies from subject to subject is a
451: consequence of the variability of all 128 pairs of
452: $(\alpha_1,\alpha_2)$ across the subjects, and offers the possibility
453: that $\eta$ can be used as a discriminating representation of the brain
454: state.
455:
456: \section{Correlations of the Scaling Exponents}
457:
458: The analysis in the preceding section treats the moments of
459: $\alpha_1$ and $\alpha_2$ separately. Only in the last step are the
460: global properties embodied in $M_q^{(1)}$ and $M_q^{(2)}$ related
461: through the exponent $\eta$ in Eq.(\ref{9}). In that approach the
462: pairing of
463: $\alpha_1$ with
464: $\alpha_2$ in each channel is not taken into account. However, we
465: know that there are channels, such as Ch.\ 1 in Figs.\ 1 and 2, where
466: the absence of a dominant mode of oscillation results in $\alpha_1
467: \approx \alpha_2$. Thus the correlation between the two scaling
468: exponents is an important feature that should be explored and
469: quantified. To that end we define
470: \begin{equation}
471: \beta=\alpha_2 / \alpha_1 \label{11}
472: \end{equation}
473: for each channel. In most cases we have $\beta<1$, but $\beta>1$
474: is possible and, by its rarity, noteworthy.
475:
476: From a scatter plot, such as Fig.\,3, it is possible to visualize
477: the $\beta$ distribution, since $\beta$ is just the slope of a line
478: from the origin to each point. We show in Fig.\,8 the $\beta$
479: distributions for the same three subjects as those in Fig.\,7.
480: Subject B is chosen for display because it has the largest $\eta$,
481: while subject C is chosen because it has several
482: $\beta$ values that exceed 1.
483:
484: To summarize the 128 values of $\beta_j$ for each subject, we
485: apply to them the moment analysis that is developed in Sec.\,IV.
486: Let us therefore define
487: \begin{equation}
488: N_q={1\over N} \sum_{j=1}^N \beta_j^q \left/ \left({1\over N}
489: \sum_{j=1}^N \beta_j\right)^q \right.. \label{12}
490: \end{equation}
491: The $q$ dependence of ln\,$N_q$ for the same three subjects are
492: shown in Fig.\,9. Again, linear fits are very good. Thus we have
493: \begin{equation}
494: N_q\propto {\rm exp} (\,\nu q\,) \label{13}
495: \end{equation}
496: with a distinct $\nu$ for each subject. Clearly, the ones with
497: wide $\beta$ distributions relative to their means have higher
498: values of $\nu$.
499:
500: We now have found two indices, $\eta$ and $\nu$, for each subject.
501: They describe different aspects of the scaling exponents. To
502: display those values, it is illuminating to show the scatter plot
503: of $(\eta,\nu)$, as in Fig.\,10, which has 18 points for the 18
504: subjects studied. The subjects A, B and C are denoted by distinct
505: symbols, same as in Fig.\,9. We see that the points in Fig.\,10
506: appear to form a band, roughly correlated in $\eta$ and
507: $\nu$. Only subject C has a value of $\nu$ that lies above the
508: band, and it is C who has several
509: $\beta$ values above 1. Whether that is an anomaly carrying some
510: physiological significance is an issue outside the scope of this
511: paper, since we assume no knowledge of the physical conditions of
512: the subjects. We are similarly ignorant at this point about the
513: meaning of the spread of the $(\eta,\nu)$ values. However, it is
514: encouraging that the scatter plot in $(\eta,\nu)$ is widely
515: distributed for the 18 subjects, since it offers the possibility
516: of being a discriminating tool, quite different from the
517: alternative scenario, if the parameters determined in an
518: analysis had turned out to be nearly the same for all subjects.
519:
520: \section{Conclusion}
521:
522: Recognizing that the brain is a highly nonlinear system, we have
523: explored a possible way of analyzing the EEG time series that
524: avoids the assumptions made in linear analyses and in chaos
525: studies. By studying the fluctuations from linear trends defined
526: over varying time scale, we have found two scaling regions in
527: which the RMS fluctuations can be characterized by two
528: dimensionless scaling exponents,
529: $\alpha_1$ and $\alpha_2$, for each channel. We then performed
530: moment analyses to reduce the large number of pairs of
531: $(\alpha_1,\alpha_2)$ to simple and manageable forms. The two
532: types of independent moments,
533: $M_q^{(i)}$ and $N_q$, yield two indices, $\eta$ and $\nu$, which provide
534: concise signatures of the nonlinear
535: behavior of all channels of the EEG signals.
536:
537: Our emphasis in this paper has been on the method of analysis
538: rather than on the physiological interpretation of the results. For
539: the latter task it is necessary to have not only more data, but
540: also detailed information on the physical conditions of the
541: subjects so as to have a reference frame to calibrate the indices
542: obtained. However, working with 18 subjects is sufficient to
543: demonstrate the effectiveness of the method, to show the
544: universality of the scaling behaviors, to reveal the range of
545: variability of the indices derived, and to offer the possibility of
546: a new way of understanding human brain activity.
547:
548:
549: \vspace{1cm}
550:
551: \begin{center}
552: {\bf Acknowledgment}
553: \end{center}
554:
555: We are grateful to Prof.~Don Tucker and Dr.~Phan Luu for
556: supplying the EEG data for our analysis. We have also benefited from
557: the computational assistance of Wei He. This work was supported, in
558: part, by the U.\,S.\,Department of Energy under Grant No.\,
559: DE-FG03-96ER40972, and the National Institutes of Health under
560: Grant No.\,R44-NS-38829.
561:
562: \newpage
563: %\begin{bibliography}
564: \begin{references}
565: \bibitem[*]{byline}Present address: Dynamic Neuroimaging Laboratory,
566: Department of Radiology, University of California at San Francisco,
567: San Francisco, CA 94143.
568:
569:
570: %1
571: \bibitem{eeg}{\it Electroencephalography: Basic Principles,
572: Clinical Applications, and Related Fields}, edited by E.\
573: Niedermeyer and F.\ H.\ Lopes da Silva (Urban and Schwarzenberg,
574: Baltimore, 1987); ibid (Williams and Wilkins, Baltimore, 1998).
575:
576: %2
577: \bibitem{nunez} P.\, L.\, Nunez, {\it Neocortical Dynamics and
578: Human EEG Rythms} (Oxford University Press, 1995).
579:
580: %3
581: \bibitem{bla} S.\ Blanco, C.\ D'Attellis, S.\ Isaacson, O.\ A.\
582: Rosso, and R.\ Sirne, Phys.\ Rev.\ E {\bf 54}, 6661 (1996); S.\
583: Blanco, A.\ Figliola, R.\ Quian Quiroga, O.\ A.\ Rosso, and E.\
584: Serrano, Phys.\ Rev.\ E {\bf 57}, 932 (1998).
585:
586: %4
587: \bibitem{chaos} {\it Chaos in Brain?}, edited by K.\ Lehnertz,
588: J.\ Arnhold, P.\ Grassberger and C.\ E.\ Elger (World
589: Scientific, Singapore, 2000).
590:
591: %5
592: \bibitem{jan} B.\ H.\ Jansen and M.\ E.\ Brandt, {\it Nonlinear
593: Dynamical Analysis of the EEG} (World Scientific, Singapore,
594: 1993).
595:
596: %6
597: \bibitem{leh} K.\ Lehnertz and C.\ E.\ Elger, Phys.\ Rev.\
598: Lett.\ {\bf 80}, 5019 (1998).
599:
600: %7
601: \bibitem{peng} C.-K.\ Peng, S.\ Havlin, H.\ E.\ Stanley, and A.\
602: L.\ Goldberger, Chaos {\bf 5}, 82 (1995).
603:
604: %8
605: \bibitem{pbh}C.-K.\ Peng, S.\ V.\ Buldyrev, S.\ Havlin, M.\
606: Simons, H.\ E.\ Stanley, and A.\ L.\ Goldberger, Phys.\ Rev.\ E
607: {\bf 49}, 1685 (1994).
608:
609: %9
610: \bibitem{wat} Watters, P.\ A.\ Complexity International {\bf 5}, 1
611: (1998).
612:
613: %10
614: \bibitem{gar} C.\ W.\ Gardiner, {\it Handbook of Stochastic Methods}
615: (Springer-Verlag, Berlin, 1983).
616:
617: %11
618: \bibitem{rch} R.\ C.\ Hwa, Phys.\ Rev.\ D {\bf 41}, 1456 (1990).
619:
620: \end{references}
621: %\end{bibliography}
622:
623:
624: % figure 1
625: \begin{figure}[ht]
626: \center\epsfig{figure=Fig1_ill.eps,width=3.25in}
627: \caption{A sample of EEG time series in three channels. The
628: vertical scales of Ch.\ 1 and Ch.\ 2 are shifted upward by 60 and
629: 30 $\mu$V, respectively.}
630: \end{figure}
631:
632:
633: % figure 2
634: \begin{figure}[ht]
635: \center \epsfig{figure=Fig2_ill.eps,width=3.25in}
636: \caption{$F(k)$ vs $k$ for the three channels in Fig.\,1. The
637: vertical scales of Ch.\ 1 and Ch.\ 2 are shifted upwards by 1.0 and
638: 0.5 units, respectively.}
639: \end{figure}
640:
641: % figure 3
642: \begin{figure}[ht]
643: \center\epsfig{figure=Fig3_ill.eps,width=3.25in}
644: \caption{Scatter plot of $\alpha_2$ vs $\alpha_1$ for
645: subject A. The three channels exhibited in Figs.\,1 and 2 are
646: shown as circles.}
647: \end{figure}
648:
649: % figure 4
650: \begin{figure}[ht]
651: \center\epsfig{figure=Fig4_ill.eps,width=3.25in}
652: \caption{Topographical plots of $\alpha_1$ (top)
653: and $\alpha_2$ (bottom). In each figure, ten contour lines
654: are drawn within the data range: solid lines above the
655: mean, dotted lines below.}
656: \end{figure}
657:
658: % figure 5
659: \begin{figure}[ht]
660: \center\epsfig{figure=Fig5_ill.eps,width=3.25in}
661: \caption{The distributions $P_m$ for $\alpha_1$ and
662: $\alpha_2$. The bin size in $\alpha$ for this plot is 0.02.}
663: \end{figure}
664:
665: % figure 6
666: \begin{figure}[ht]
667: \center\epsfig{figure=Fig6_ill.eps,width=3.25in}
668: \caption{The $q$ dependence of $\ln M_q^{(i)}$ for
669: subject A. The straight lines are linear fits of the points for
670: $q \ge 5$.}
671: \end{figure}
672:
673: % figure 7
674: \begin{figure}[ht]
675: \center\epsfig{figure=Fig7_ill.eps,width=3.25in}
676: \caption{A log-log plot of $M_q^{(2)}$ vs $M_q^{(1)}$
677: for three subjects A, B, and C. The solid lines have the slopes
678: given by Eq.\,(\ref{10}).}
679: \end{figure}
680:
681: % figure 8
682: \begin{figure}[ht]
683: \center\epsfig{figure=Fig8_ill.eps,width=3.25in}
684: \caption{The distributions of the $\beta$ values of
685: the subjects A, B, and C.}
686: \end{figure}
687:
688: % figure 9
689: \begin{figure}[ht]
690: \center\epsfig{figure=Fig9_ill.eps,width=3.25in}
691: \caption{The $q$ dependence of $\ln N_q$ for
692: subjects A, B, and C.}
693: \end{figure}
694:
695: % figure 10
696: \begin{figure}[ht]
697: \center\epsfig{figure=Fig10_ill.eps,width=3.25in}
698: \caption{Scatter plot of $\nu$ vs $\eta$ for all 18
699: subjects, three of which have individual symbols: triangle (A),
700: square (B), and star (C).}
701: \end{figure}
702:
703:
704: \end{document}
705:
706:
707:
708:
709:
710:
711: