1: % ****** Start of file apssamp.tex ******
2: % This file is part of the APS files in the REVTeX 4 distribution.
3: % Version 4.0 of REVTeX, August 2001
4: % Copyright (c) 2001 The American Physical Society.
5: % See the REVTeX 4 README file for restrictions and more information.
6: % TeX'ing this file requires that you have AMS-LaTeX 2.0 installed
7: % as well as the rest of the prerequisites for REVTeX 4.0
8: % See the REVTeX 4 README file
9: % It also requires running BibTeX. The commands are as follows:
10: % 1) latex apssamp.tex
11: % 2) bibtex apssamp
12: % 3) latex apssamp.tex
13: % 4) latex apssamp.tex
14: %\documentclass[preprint,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}
15: % Some other (several out of many) possibilities
16: %\documentclass[preprint,aps]{revtex4}
17: %\documentclass[preprint,aps,draft]{revtex4}
18: %\documentclass[prb]{revtex4}% Physical Review B
19: % Include figure files
20: % Align table columns on decimal point
21: % bold math
22: %\nofiles
23:
24:
25: \documentclass[showpacs,preprintnumbers,prl,a4paper,twocolumn]{revtex4}
26: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
27: \usepackage{amssymb}
28: \usepackage{amsfonts}
29: \usepackage{amsmath}
30: \usepackage{graphicx}
31: \usepackage{dcolumn}
32: \usepackage{bm}
33:
34: \setcounter{MaxMatrixCols}{10}
35: %TCIDATA{OutputFilter=LATEX.DLL}
36: %TCIDATA{Version=4.00.0.2312}
37: %TCIDATA{LastRevised=Monday, September 05, 2005 16:50:44}
38: %TCIDATA{<META NAME="GraphicsSave" CONTENT="32">}
39: %TCIDATA{Language=American English}
40:
41: %\input{tcilatex}
42:
43: \begin{document}
44:
45: \preprint{}
46: \title{\bf Dynamics and Synchrony from Oscillatory Data
47: via Dimension Reduction}
48: \author{Jie Zhang$^1$, Kai Zhang$^2$, Jianfeng Feng$^3$, Junfeng Sun$^1$, Xiaoke Xu$^1$, Michael Small$^1$\\
49: \small{\emph{$^1$Department of Electronic and Information
50: Engineering, Hong Kong
51: Polytechnic University, Hong Kong\\
52: $^2$Department of Computer Science and Engineering, Hong Kong
53: University of Science and Technology, Hong Kong\\ $^3$Department of
54: Computer Science and Mathematics, Warwick University, Coventry CV4
55: 7AL, UK}} }
56:
57: \email[$^1$]{enzhangjie@eie.polyu.edu.hk}
58: \date{\today }
59:
60: \begin{abstract}
61: Complex, oscillatory data arises from a large variety of biological,
62: physical, and social systems. However, the inherent oscillation and
63: ubiquitous noise pose great challenges to current methodology such
64: as linear and nonlinear time series analysis. We exploit the state
65: of the art technology in pattern recognition and specifically,
66: dimensionality reduction techniques, and propose to rebuild the
67: dynamics accurately on the ``cycle'' scale. This is achieved by
68: deriving a compact representation of the cycles through global
69: optimization, which effectively preserves the topology of the cycles
70: that are embedded in a high dimensional Euclidian space. Our
71: approach demonstrates a clear success in capturing the intrinsic
72: dynamics and the subtle synchrony pattern from uni/bivariate
73: oscillatory data over traditional methods. Application to the human
74: locomotion data reveals important dynamical information which allows
75: for a clinically promising discrimination between healthy subjects
76: and those with neural pathology. Our results also provide
77: fundamental implications for understanding the neuromuscular control
78: of human walking.
79:
80: \end{abstract}
81: \pacs{05.45.Tp, 05.45.Xt}
82: \maketitle
83:
84:
85: Time series exhibiting complex oscillatory dynamics are widely
86: observed in the real world: from finance to biomedical systems
87: \cite{leo}. In this Letter we focus on an important class of data
88: with a stable oscillation frequency but irregular waveform
89: fluctuations, also called pseudoperiodic time series. Such data
90: arise from broad application domains and have gained particular
91: interest in recent years, with examples ranging from human ECG and
92: gait data \cite{ecg, Dingwell}, white blood-cell count and tremor
93: data in patients \cite{blood,tremor}, epidemic dynamics \cite{epi},
94: light intensity of laser \cite{laser}, sun spot numbers and global
95: temperature variation \cite{sun, temp}. Accurate extraction and
96: characterization of the dynamics of these time series is a general
97: problem, which holds the key to understanding the inner workings of
98: many important physical and biological systems. To this end,
99: traditional methods rely primarily on linear spectral analysis or
100: computing nonlinear characteristics, and recent attempts include
101: producing pseudoperiodic surrogate data \cite{pps}, or to establish
102: a nonlinear transform from cycles in time domain to nodes in a dual
103: complex network domain \cite{zj}.
104:
105:
106: However, there are still no generic, systematic and robust
107: approaches to handle such oscillatory time series. The Fourier
108: transform, though widely applied to oscillatory data, is inherently
109: a linear method and cannot account for the nonlinear nature of the
110: data, if any. For nonlinear approaches, the cyclic trend overlying
111: the signal, together with noise from unknown sources, can often mask
112: the intrinsic dynamics \cite{periodicity, noise} and pose great
113: challenges to the nonlinear time series analysis/modeling
114: techniques. For example, the most popular Poincare section method,
115: which reduces the flow data to intersections of the trajectories
116: with a lower-dimensional subspace, may produce highly corrupted
117: results under noisy environment, in that the intersection points can
118: no longer preserve the original dynamics in the presence of noise.
119: Things get even worse when the cycles possess complex noncoherent
120: waveforms. In this paper we propose a novel, generic approach that
121: can effectively capture the dynamics of oscillatory time series. By
122: mapping the cycles in the time series to a multi-dimensional
123: Euclidian space, we seek a low-dimensional representation which
124: topologically preserves the important proximity relation among
125: cycles, through efficient global optimization. It is the first time
126: that advances in dimension reduction are explored in reconstructing
127: the dynamics of complex oscillatory data. Our approach utilizes the
128: richer information of pairwise cycle correlation, therefore it not
129: only excavates the inherent dynamics obscured by the cyclic trend,
130: but also offers an extra robustness to noise due to the global
131: nature of the method. Exploration of our approach to probing the
132: synchrony between bivariate oscillatory time series is shown to be
133: able to reveal the subtle synchrony for which traditional approaches
134: will fail. We applied the proposed method to the human locomotion
135: data, and are capable of discriminating between the healthy subjects
136: and those with neuropathology reliably. Based on the results we are
137: able to make the important biological inference that the human
138: walking is not critically dependent on the peripheral neural
139: feedback.
140:
141: \section{Theory}
142: \subsection{Reconstructing Dynamics Underlying Cyclic Trend by Spectral Clustering}
143: We demonstrate how intrinsic dynamics of psuedoperiodic data can be
144: extracted using \emph{Laplacian Eigenmaps} \cite{laplacian}, a
145: nonlinear dimension reduction method based on \emph{spectral graph
146: theory}. We illustrate with the $X$-component of the chaotic
147: R\"{o}ssler system. Motivated by the fact that such data usually
148: exhibit a highly redundant patterns in the form of repeated cycles,
149: we first partition the time series into individual cycles $C_{i}$'s
150: ($i=1,...,k$) by local minimums as shown in Fig. \ref{ts} (a). Each
151: cycle is then mapped to a high dimensional vector $x_i$ whose
152: dimension equals the number of points in that cycle. Our goal is to
153: compute a set of new, low-dimensional (in the simplest case, $1D$)
154: representation, or embedding $y_i$'s, such that the proximity
155: relations among $x_i$'s are maximally preserved in their
156: low-dimensional counterparts $y_i$'s and that the redundancy in the
157: cycles are removed.
158:
159: \begin{figure}
160: \begin{center}
161: \includegraphics[height=0.3\textwidth]{1.eps}
162: \caption{ (a) Time series form $X$-component of the chaotic R\"{o}ssler system,
163: which is divided into cycles by local minima (circles). (b)
164: Laplacian embedding $c(i)$ extracted on the cycle scale for the time
165: series shown in (a), and (c) the return plots of $c(i)$ derived by
166: two dimension reduction methods, i.e., the Laplacian Eigenmaps and
167: Multidimentional scaling.} \label{ts}
168: \end{center}
169: \end{figure}
170:
171: To achieve this, a weighted graph $\xi$ is constructed, where each
172: node corresponds to a cycle $x_i$ and edges are assigned between all
173: pairs of nodes. The graph can either be fully connected or only bind
174: those vertices within the $k$-nearest-neighbors and we aims at
175: extracting the dynamics from the topology of the graph. We use
176: $W_{ij}$ to denote the similarity between cycle $x_i$ and $x_j$,
177: which can be chosen conveniently as the correlation coefficient
178: $\rho_{ij}=Cov(x_i,x_j)/(\sigma_{x_i} \sigma_{x_j})$ (in case $x_i$
179: and $x_j$ differ in length, shift the shorter vector on the longer
180: one until $\rho_{ij}$ maximizes). Then, the low-dimensional
181: representation $y_i$'s can be cast as the solution of the following
182: optimization problem, $\min \sum_{i,j\in\xi}W_{ij}||y_i-y_j||^2$,
183: which penalizes those mappings where nearby points $x_i$'s are
184: relocated far apart in the space of $y_i$'s. In case of univariate
185: $y_i$'s, the objective can be written as $\textbf{y}^TL\textbf{y}$,
186: where $\textbf{y} = [y_1, y_2,...,y_k]^{\top}$, $L=D-W$ is the
187: \emph{graph Laplacian}, and $D$ is the diagonal degree matrix such
188: that $D_{ii}=\sum_j W_{ji}$. Hence the name \emph{Laplacian
189: Eiganmap}. To prevent arbitrary scaling in the solution, one can
190: enforce the constraint $\textbf{y}^TD\textbf{y}=1$.
191:
192:
193: \begin{figure}[htbp]
194: \begin{center}
195: \includegraphics[height=0.3\textwidth]{2.eps}
196: \caption{ Return plots for the time series from (a) Poincare section points $P(i)$,
197: which is obtained by collecting the local minima $X_{min}$, and (b)
198: the extracted $c(i)$. The R\"{o}ssler system here is corrupted by
199: $5\%$ dynamical noise and $30\%$ measurement noise. It is obvious
200: that the return plots from $c(i)$ takes on a clearer form of
201: quadratic function, indicating that the original nonlinear dynamics
202: is sufficiently kept.}\label{noise1}
203: \end{center}
204: \end{figure}
205:
206:
207: The above constrained minimization is solved by the generalized
208: eigenvalue problem $L \textbf{y}_i=\lambda_i D \textbf{y}_i$, where
209: $\lambda_i$'s (${i = 1,2,...,k}$) are eigenvalues sorted in an
210: ascending order, and $\textbf{y}_i$'s are the corresponding
211: eigenvectors. It can be shown that the minimum eigenvalue
212: $\lambda_1$ is zero, corresponding to an eigenvector
213: ($\textbf{y}_1$) of all 1's. Therefore it is degenerate and the
214: optimal solution is actually provided by $\textbf{y}_2$, the
215: eigenvector of the second smallest eigenvalue \cite{laplacian}. As
216: is shown in Fig. \ref{ts}, the eigenvector $\textbf{y}_2(i)$
217: provides a unique, reduced representation of the original time
218: series by capturing the dynamics of the oscillatory data on the
219: cycle scale. We use a general notation $c(i)$ ($c$ for cycle) for
220: such reduced series representation (i.e., other dimension reduction
221: schemes can also be applied, and we denote the reduced series all as
222: c(i)). More generally, to obtain an $m$-dimensional solution for
223: $y_i$'s, one can simply extract the $m$ trailing eigenvectors
224: $\textbf{y}_{i}$'s ($i=2,3,...,m+1$).
225:
226: \subsection{The Significance of Extracting Dynamics on Cycle Scale }
227: The comparison between Laplacian embedding $c(i)$ and the Poincare
228: section points $P(i)$ (obtained by collecting the local minima)
229: indicates that they are dynamically identical by sharing the same
230: chaotic invariants, i.e., Lyapunov exponent and correlation
231: dimension. However, in case of significant noise (see Fig.
232: \ref{noise1}), our approach can capture more of the deterministic
233: structure than does $P(i)$ and is therefore more robust. This is not
234: surprising, since acquisition of $c(i)$ is based on an optimization
235: process that preserves the proximity relation between all pairs of
236: cycles simultaneously, while $P(i)$ is obtained by treating each
237: cycle independent of each other. It is reasonable to expect the
238: former to excavate more useful, global patterns than the latter.
239: %Later we shall see that our approach is also more reliable than
240: %simply extracting the cycle interval.
241:
242: Many other forms of dimension reduction can also be applied here,
243: which produce similar results. For example, the
244: \emph{Multidimensional Scaling} method that preserves the pairwise
245: distance yields almost the same outputs as the \emph{Laplacian
246: Eigenmaps}, see Fig. \ref{ts}(c). Actually, the applicability of
247: dimension reduction techniques in extracting the dynamics shall be
248: generally justifiable, considering the intrinsically low correlation
249: dimensions of most real world pseudoperiodic data. The corresponding
250: trajectories in phase space tend to have similar orientations for
251: nearby cycles (see, e.g., Fig. \ref{gait}). Such redundancy can be
252: effectively removed through dimension reduction, leaving only the
253: useful degrees of freedom for manifestation of the dynamics of
254: interest.
255:
256: \begin{figure}[htbp]
257: \begin{center}
258: \includegraphics[height=0.3\textwidth]{3.eps}
259: \caption{Correlation between $X$ and $Y$ component revealed by (a)
260: Poincare section points $P(i)$ and (b) the extracted $c(i)$. The
261: R\"{o}ssler system is corrupted by $5\%$ dynamical noise and $30\%$
262: measurement noise. }\label{noise2}
263: \end{center}
264: \end{figure}
265:
266: Reconstructing the dynamics on the cycle scale not only enhances the
267: robustness to noise, but bring new vitality to a number of nonlinear
268: methods which are otherwise not suitable for oscillatory data due to
269: the inherent periodicity, such as detrended fluctuation analysis,
270: surrogate data method, recurrence plot, causality, entropy and so
271: on. For example, the detrended fluctuation analysis, which computes
272: the variance of the detrended data at different scales, will produce
273: similar variance curves (i.e., uprising at small scale and saturate
274: at the periodicity of the signal) for all pseudoperiodic data
275: despite their distinct dynamics. This is because the periodic trend
276: dominating in the data will conceal the underlying, subtle dynamics.
277: Thus extracting the dynamics on the cycle scale proves to be crucial
278: for subsequent analysis.
279:
280: \subsection{Detecting Synchrony from Oscillatory Data} A topic
281: which is of great interest to oscillatory data analysis in recent
282: years is to detect the degree of synchronization between
283: self-sustained oscillators. For example, the peakness of the phase
284: difference distribution and the consistency of mutual nearest
285: neighbors are used to characterize the phase synchrony \cite{Tass,
286: phase} and the dynamical interdependence \cite{a}, respectively.
287: Here we are especially interested in the case where the phase
288: relation between two oscillator is evident, but is hard to estimate
289: due to non-phase-coherence and noise, or is not sensitive to degree
290: of changes of synchrony. For example, the blood pressure interacted
291: with heartbeat, where each cycle of blood pressure variation
292: correspond to one heartbeat. The phase index in this case may not be
293: quite informative of the synchrony strength. On the other hand, the
294: presence of noise in most real world data will inevitably degrade
295: the dependency measures.
296:
297:
298:
299: Unlike traditional methods, we
300: propose to quantify the synchrony the noisy, noncoherent
301: pseudoperiodic data \emph{on their cycle scale} through their
302: Laplacian embedding $c(i)$. We first test with the $X$ and $Y$
303: components of the noisy R\"{o}ssler system. Although these two
304: components are non-separable and cannot be strictly defined as being
305: synchronized, they are actually ``in phase'' and therefore suitable
306: in our context. The two time series are first segmented into cycles
307: by the local maximums of one of the data (i.e., the \emph{i}th cycle
308: in both time series share the same time span). We find that the
309: Poincare section points picked up from the two time series can
310: hardly capture the interdependency due to noise, in that the scatter
311: plot between the Poincare section points of $X$ and $Y$ leads to a
312: group of random points (Fig. \ref{noise2} (a)). Comparatively, the
313: extracted Laplacian embedding $c(i)$'s extracted from $X$ and $Y$ by
314: our approach successfully reveal the synchrony pattern by
315: demonstrating an uprising trend in the corresponding scatter plot
316: (Fig. \ref{noise2} (b)).
317:
318:
319:
320: \begin{figure}[htbp]
321: \begin{center}
322: \includegraphics[height=0.3\textwidth]{4.eps}
323: \caption{ Time series (left panel) and the corresponding phase space
324: reconstructions (right panel) for knee ((a)-(b)) and ankle
325: kinematics data ((c)-(d)). The two time series are typically
326: non-phase-coherent by demonstrating multi-oscillation within each
327: cycle. This is also evident from the multi-center rotations of the
328: attractor in phase space (see (b) and (d)). The two signals are
329: divided by their respective local maxima into consecutive
330: cycles.}\label{gait}
331: \end{center}
332: \end{figure}
333:
334: \section{Application to Human Locomotion Data}
335: \subsection{Characterizing the Locomotion Dynamics}
336: Now we apply our approaches to human gait data collected from
337: electrogoniometers at the knee and ankle joints that measure their
338: sagittal plane kinematics during overground walking. Human
339: locomotion is a highly complex, rhythmic process that involves
340: multilevel control of central nerve system and feedback from various
341: peripheral sensors. Here we consider two groups of subjects from
342: healthy controls (CO) and diabetic patients (NP, with significant
343: peripheral neuropathy), each with 10 subjects \cite{Dingwell}.
344:
345: \begin{figure}
346: \begin{center}
347: \includegraphics[height=0.3\textwidth]{5.eps}
348: \caption{ Power spectrum density (PSD) for typical healthy subject
349: (the left panel) and the diabetic patient (the right panel). The
350: top, middle and bottom rows are PSDs for the extracted $c(i)$,
351: stride interval series, and the original ankle data, respectively.
352: The frequency here is normalized by the sampling rate. The PSDs for
353: the stride interval series and the original data show no significant
354: difference between the healthy subject and the diabetic patient.
355: %The sampling rate of $c(i)$ in (b) can be taken as the mean cycle interval
356: % %These log-log PSDs are then fitted with a linear function using
357: %least squire regression.
358: }\label{psd}%c03 and np16
359: \end{center}
360: \end{figure}
361:
362: Typically, the human gait time series (Fig. \ref{gait} (a) and (c))
363: exhibit a relatively fixed frequency with the stride interval
364: demonstrating little variability, which is suited for dimension
365: reduction. We first extract the dynamics of the ankle on its cycle
366: scale by Laplacian embedding. Figure \ref{psd} (a)-(b) display the
367: power spectrum density (PSD) of the extracted $c(i)$ typically
368: observed for the two groups. We find that most CO subjects
369: demonstrate broad band spectrums that scale as $1/f^\beta$,
370: $\beta=0.76 (mean) \pm 0.23 (std)$, indicating the presence of long
371: range correlation (i.e., the strides separated by a large time span
372: are still statistically correlated). In comparison, the spectrum of
373: the diabetic patients are mostly flat resembling white noise
374: processes ($\beta=0.37\pm 0.16$), which means that the strides at
375: different times are mostly uncorrelated. This difference, however,
376: has not been found with either the stride interval (SI) series (Fig.
377: \ref{psd} (c)-(d)) or the raw data (Fig. \ref{psd} (e)-(f)). This is
378: because, the periodicity in the raw data can obscure the underlying
379: fluctuations and that a linear Fourier transform is not capable of
380: characterizing nor distinguishing the nonlinear dynamics in the
381: data; on the other hand, although SI removes the cyclic trend and
382: has been widely used to quantify pathological locomotion
383: \cite{Haus}, it loses much of the dynamical information by only
384: recording the duration of cycles. In comparison, the transform
385: $c(i)$ successfully removes the periodic trend while preserving the
386: dynamical fluctuation within each cycle, in particular, the specific
387: patterns of the four phases within a pace. Therefore it keeps more
388: valuable information about the neuromuscular control of human
389: walking.
390:
391:
392: \subsection{ Synchrony Detection between Knee and Ankle Movement}
393: The human walking involves the coordination of two major joints,
394: i.e., the knee and the ankle. So we are also interested in examining
395: the synchrony between the knee and ankle data, which may provide
396: further insights in understanding the neuralcontrol of walking. The
397: knee and ankle movements during continuous walking are obviously in
398: phase due to the physical connection of the two joints. The strength
399: of coupling, however, can hardly be recognized from the phase due to
400: the noncoherence \cite{explain} (see Fig. \ref{gait}). Also, noise
401: tends to destroy the local structure in phase space and thus hampers
402: the dynamical dependency measures \cite{qian}. To circumvent these
403: difficulties, we compare the dynamics of the two time series by
404: using their Laplacian Embedding $c(i)$'s. Note that each time series
405: can be segmented by either its own local maximums, or those of its
406: partner series (markers in Fig. \ref{gait}). Therefore we will
407: segment each time series twice and compute the averaged correlation
408: coefficients $\rho_{ij}$'s between $c_{ankle}$ and $c_{knee}$.
409:
410: \begin{figure}
411: \begin{center}
412: \includegraphics[height=0.3\textwidth]{6.eps}
413: \caption{ Synchrony between knee and ankle locomotion
414: revealed by Laplacian embedding $c(i)$ ((a) and (b)) and stride
415: interval series ((c) and (d)), with the left and right columns for
416: an healthy subject and a diabetic patient, respectively. As can be
417: seen in (b), the stride interval series cannot distinguish the
418: healthy from the diabetics, as the scatter plots of the stride
419: interval series show similar uprising trends for both of them.
420: }\label{y2xianguan}%c01 and np24
421: \end{center}
422: \end{figure}
423:
424: Then we examine the interrelation between the Laplacian embeddings
425: from the ankle and the knee data. For healthy subjects, the scatter
426: plots demonstrate a significant uprising trend (Fig.
427: \ref{y2xianguan} (a), $\rho=0.68\pm0.19$), indicating that the knee
428: and ankle movements are highly synchronized; while for diabetics
429: patients, the synchrony almost vanishes (Fig. \ref{y2xianguan} (b),
430: $\rho=0.26\pm0.18$). Again, the discrimination between CO and NP
431: groups is unavailable by stride interval series, which always
432: exhibits a strong correlation between the joints (Fig.
433: \ref{y2xianguan} (c)-(d)), corresponding to high degree of phase
434: synchronization. Note that our approach avoids the difficulty of
435: defining the phase. On the other hand, it is superior to directly
436: computing the correlation between pairs of equal-time cycles picked
437: respectively from the two time series, which is sensitive to noise
438: and nonstationarity. Finally, we point out that a more comprehensive
439: description of synchrony can be achieved by examining more Laplacian
440: eigenvectors. In the current case the single eigenvector
441: $\textbf{y}_2$ already encodes the primary variability and is thus
442: sufficient for the discriminative task.
443:
444:
445: Note that we have observed a lack of long range correlation in the
446: ankle kinematics of the NP group (Fig. \ref{psd} (b)). This suggests
447: the alteration of the locomotor pattern in the lower limbs, due to
448: the nerve deterioration in feet and toes typical of the diabetics.
449: Despite the impaired peripheral feedback caused by the dying nerves,
450: the \emph{knee} kinematics of the NP group are \emph{still} found to
451: demonstrate a stable long range correlation indistinguishable from
452: the CO group. This ``mismatch" between the ankle/knee dynamics for
453: diabetics can be reasonably explained by the weak synchrony between
454: these two joints, as is shown in Fig. \ref{y2xianguan} (b). From
455: these findings, we can see that the impaired neuralfeedback from the
456: feet influences only the lower limb locomotion while not that of the
457: knees. We therefore conclude that the human walking is not
458: critically dependent on the feedback from peripheral nervous system.
459: It should be noted that although Gates \emph{et.al.} \cite{gates}
460: checked the same data, they do not consider the interaction among
461: the knee and angle locomotion which may be important in
462: characterizing understanding neuralcontrol of human walking.
463: Furthermore the authors relied solely on the extracted stride
464: interval series, which have been shown to contain limited dynamical
465: information and therefore may be insufficient for the analysis and
466: evaluation of the human walking.
467:
468: \section{ Conclusion}
469: To conclude, we have for the first time circumvented the problem of
470: rebuilding the complex dynamics in oscillatory data from the
471: viewpoint of dimension reduction. We use a global optimization
472: procedure to compute the inherent, low-dimensional representation
473: that maximally preserves the topology of the cycles when embedded in
474: the multi-dimensional Euclidian space. Our approach has been shown
475: to be very robust to noise, and is capable of extracting the
476: underlying dynamics and identifying the subtle synchrony pattern
477: which usually degrade traditional linear or nonlinear methods.
478: Application to human gait data provides clinically promising
479: information in discriminating the healthy from the neuropathological
480: patients, and further enables us to make fundamental inference on
481: the neuromuscular control mechanism of human walking.
482:
483: Our approach may be of great relevance, and is expected to provide
484: more accurate and powerful diagnostics, to the general biological
485: and biomedical fields where complex oscillatory data abound. For
486: example, we have applied the proposed method to the phonation data
487: from the Parkinson patients, and find a significant lack of long
488: range correlation among the cycles in the signal compared to the
489: healthy people. Our approach can therefore serve as an alternative
490: tool in early Parkinson disease (PD) detection, where there are no
491: blood or laboratory tests currently that can help in diagnosing PD.
492: Other potential applications which will be involved in the future
493: works include the discrimination between Parkinson and essential
494: tremor, and the evaluation of the interaction between the blood
495: pressure variation with the heart beat.
496:
497:
498:
499:
500:
501: \begin{thebibliography}{50}
502: \bibitem{leo} Glass L (2001) Synchronization
503: and rhythmic processes in physiology. Nature 410:277-284.
504: \bibitem{ecg} Bezerianos A \emph{et al}. (1995) Time
505: Series Analysis of Electrocardiograms. \emph{Chaos} 5:95-101.
506: \bibitem{Dingwell} Dingwell JB, Cusumano JP (2000) Nonlinear
507: time series analysis of normal and pathological human walking.
508: \emph{Chaos} 10:848-863.
509: \bibitem{blood} Haurie C \emph{et al.} (1998) Cyclical neutropenia and other periodic
510: hematological disorders: a review of mechanisms and mathematical
511: models. \emph{Blood} 92:2629-2640.
512: \bibitem{tremor} Edwards R, Beuter AJ (2000) Using time domain
513: characteristics to discriminate physiologic and parkinsonian
514: tremors. \emph{J. Clin. Neurophysiol.}:17, 87-100.
515: %\bibitem{4} K. T. Alligood, T. D. Sauer, and J. A. Yorke, Chaos: An Introduction
516: %to Dynamical Systems (Springer, New York, 1996).
517:
518: \bibitem{epi}Olsen LF, Schaffer WM (1990) Chaos vs. noisy periodicity: Alternative
519: hypotheses for childhood epidemics. \emph{Science}:249, 499-504.
520: \bibitem{laser} R. Roy \emph{et al.} (1992) Dynamical control of a chaotic laser:
521: Experimental stabilization of a globally coupled system. \emph{Phys.
522: Rev. Lett.}:68, 1259-1262.
523: \bibitem{sun} Palus M, Novotna D(1999) Sunspot cycle: A driven nonlinear
524: oscillator? \emph{Phys. Rev. Lett.}:83, 3406-3409.
525: \bibitem{temp} Ghil M, Vautard R (1991) Interdecadal oscillations and the warming
526: trend in global temperature time series. \emph{Nature}:350, 324-327.
527: \bibitem{pps} Small M, Yu DJ, Harrison RG (2001) A surrogate test
528: for pseudoperiodic time series data. \emph{Phys. Rev. Lett.} 87,
529: 188101.
530: \bibitem{zj} Zhang J, Small M (2006) Complex network from pseudoperiodic
531: time series: Topology vs dynamics. \emph{Phys. Rev. Lett.}:96,
532: 238701; Zhang J \emph{et al}. (2006) Detecting temporal and spatial
533: correlations in pseudo-periodic time series. \emph{Phys. Rev. E}:73,
534: 016218.
535: \bibitem{periodicity} Cazelles B, Ferriere RH (1992) How predictable is chaos.
536: \emph{Nature} 355:25-26.
537: \bibitem{noise} Theiler J (1991) Some comments on the correlation dimension
538: of 1/f noise. \emph{Phys. Lett. A}:155, 480-493.
539: \bibitem{laplacian} Belkin M, Niyogi P (2003) Laplacian eigenmaps for
540: dimensionality reduction and data representation. \emph{Neural.
541: Comput.}:15(6),1373-1396.
542:
543: \bibitem{Haus} Hausdorff JM \emph{et al.}, Altered fractal dynamics of gait: reduced
544: stride-interval correlations with aging and Huntington's disease.
545: \emph{J. Appl. Physiol.}:82, 262-269 (1997).
546: \bibitem{Tass} Tass P \emph{et al}.(1998) Detection of n:m Phase Locking from Noisy Data:
547: Application to Magnetoencephalography \emph{Phys. Rev. Lett.}:81,
548: 3291-3294.
549: \bibitem{phase} Rosenblum M, Pikovsky A, Kurths J (1996) Phase
550: synchronization of chaotic oscillators \emph{Phys. Rev. Lett.}:76,
551: 1804-1807.
552: \bibitem{a} Schiff SJ \emph{et al.}(1996) Detecting dynamical interdependence and
553: generalized synchrony through mutual prediction in a neural
554: ensemble. \emph{Phys. Rev. E}:54, 6708-6724.
555:
556: \bibitem{explain} Though phase can be defined by using the marker
557: events (denoted by circles in Fig. \ref{gait}), it actually contains
558: the same information as the stride interval data by performing a
559: linear interpolation from \emph{0} to $2\pi$ over a cycle. We have
560: shown that the SI data contains insufficient dynamical information
561: for a discriminative purpose.
562:
563: \bibitem{qian} Quiroga RQ \emph{et al.}(2000) Learning driver-response
564: relationships from synchronization patterns. \emph{Phys. Rev. E}:61, 5142-5148.
565: \bibitem{gates} Gates DH, Dingwell JB (2007) Peripheral neuropathy does not
566: alter the fractal dynamics of stride intervals of gait. \emph{J.
567: appl. physiol.}:102 965-971.
568:
569: \end{thebibliography}
570:
571: \end{document}
572: