physics0110075/hrv.tex
1: %\documentclass[preprint,aps]{revtex4}
2: \documentclass[twocolumn,aps,showpacs]{revtex4}
3: %\documentclass[aps]{revtex4}
4: %\documentclass[singlecolumn,aps]{revtex4}
5: 
6: \usepackage[dvips]{graphicx}% Include figure files
7: \usepackage{dcolumn}% Align table columns on decimal point
8: %\documentstyle[aps,preprint]{revtex}
9: %\input{epsf.sty}
10: \begin{document}
11: %\draft  
12: \title{Zipf's law in human heartbeat dynamics}
13: \author{J.\ Kalda, M.\ S\"akki}  
14: \affiliation{
15: Institute of Cybernetics,
16: Tallinn Technical University,
17: Akadeemia tee 21,
18: 12618 Tallinn,
19: Estonia}
20: \author{M.\ Vainu}
21: \affiliation{Tallinn Diagnostical Center, P\"arnu mnt.\ 104, Estonia}
22: \author{M.\ Laan}
23: \affiliation{N\~omme Children Hospital, Laste 1, Tallinn, Estonia}
24: %\date{\today}
25: 
26: \begin {abstract} 
27: 
28: It is shown that the distribution of low variability periods in the activity of
29: human heart rate typically follows a multi-scaling Zipf's law. The presence or 
30: failure of a power law, as well as the values of the scaling exponents, are 
31: personal characteristics depending on the daily habits of the subjects.
32: Meanwhile, the distribution function of the low-variability periods as a whole 
33: discriminates efficiently between various heart pathologies. This new technique
34: is also applicable to other non-linear time-series and reflects these aspects
35: of the underlying intermittent dynamics, which are not covered by other 
36: methods of linear- and nonlinear analysis.
37: 
38: %can be used to efficiently discriminate 
39: %here four parameters are selected to reflect the  main part of 
40: %the diagnostically valuable information.
41: \end {abstract} 
42: \pacs{PACS numbers: 87.10.+e, 87.80.Tq, 05.45.-a, 05.45.Tp}
43: 
44: \maketitle
45: 
46: 
47: The nonlinear and scale-invariant aspects of the heart rate
48: variability (HRV) have been studied intensively during the last
49: decades. This continuous interest to the HRV can be attributed to the
50: controversial state of affairs: on the one hand, the nonlinear
51: and scale-invariant analysis of HRV has resulted in many methods of
52: very high prognostic performance (at least on test groups) \cite{Amaral,Ivanov,Thurner,Saermark};
53: on the other hand, practical medicine is still confident to the 
54: traditional ``linear'' methods.
55: The situation is quite different from what has been observed three decades ago,
56: when the ``linear'' measures of HRV became widely used as important
57: noninvasive diagnostic and prognostic tools,
58: soon after the pioneering paper \cite{Hon}.
59: Apparently, there is a need for further evidences for the superiority of
60: new methods and for the resolution of the existing ambiguities.
61: 
62: During recent years the main attention of studies has been
63: focused on the analysis of the scale-invariant methods. It has been
64: argued that measures related to a certain time-scale (e.g.\ 5 min) are
65: less reliable, because the characteristic time-scales of
66: physiological processes are patient-specific. The scale-invariant
67: measures are often believed to be more universal and sensitive to life-threatening 
68: pathologies \cite{Amaral,Ivanov}. However, carefully designed time-scale-related 
69: measures can be also highly successful, because certain
70: physiological processes are related to a specific time scale \cite{Thurner}. 
71: 
72: The scale invariance has been exclusively seen in the heart rhythm
73: following the (multi)fractional Brownian motion (fBm) \cite{Peng}. It
74: has been understood that the heart rhythm fluctuates in a very
75: complex manner and reflects the activities of the subject (sleeping,
76: watching TV, walking etc.) \cite{Poon,Struzik} and cannot be adequately described by
77: a single Hurst exponent of a simple fBm. In order to reflect
78: the complex behavior of the heart rhythm, the multi-affine
79: generalization of the fBm has been invoked \cite{Amaral,Ivanov}; it has been claimed that 
80: the multifractal scaling exponents are of a significant prognostic value.
81: 
82: The approach based on fBm
83: addresses long-time dynamics of the heart rhythm while completely neglecting 
84: the short-scale dynamics on time scales less than one minute (the respective 
85: frequencies are typically filtered out \cite{Peng}). The {\em short-time variability} 
86: has been described only by the so called linear measures, such as 
87: $p_{NN50}$ (the probability that two adjacent normal heart beat intervals 
88: differ more than 50 milliseconds).
89: Meanwhile, the level of the short-time variability of the human heart rate varies
90: in a very complex manner, the high- and low-variability periods are deeply intertwined \cite{Poon}. 
91: This %seems to be 
92: is a very important aspect, because the low-variability periods are the 
93: periods when  the heart is in a stressed state, with high level of signals arriving 
94: from the autonomous nervous system.
95: The conventional linear measures are not appropriate for describing
96: such a complex behavior. Thus, there is a clear need for suitable nonlinear methods.
97: 
98: In this paper we present a new scale-invariant description of the 
99: short-time variability of the heart rate.
100: It is shown that the distribution of low-variability
101: periods in the activity of a normal heart follows the Zipf's law. 
102: It is also shown that the distribution function of the low-variability
103: periods contains a considerable amount of diagnostically 
104: valuable information. This new 
105: technique is also applicable to other non-linear 
106: time-series, such as EEG signals and financial data \cite{KSK}.
107: 
108: 
109: Our analysis is based on ambulatory Holter-monitoring data 
110: (recorded at Tallinn Diagnostic Centre) of 218 patients with various diagnoses.
111: The groups of patients are shown in Table 1. The sampling rate of ECG was 
112: 180~Hz. %, which corresponds to the resolving power of 6 ms. 
113: The patients were monitored  during 24 hour under normal daily activities.
114: The preliminary analysis of the ECG recordings was performed using the 
115: commercial software; this resulted in the sequence of the {\em normal-to-normal} (NN)
116: intervals $t_{NN}$ (measured in milliseconds), %. The normal-to-normal intervals are 
117: which are defined as the intervals between two subsequent normal heartbeats (i.e.\ normal QRS complexes).
118: %TABLE I
119: \begin {table}[!bth]
120: \begin {tabular}{|l|c|c|c|c|c|c|c|}
121: \hline
122:          & Healthy & IHD & SND  & VES & PCI & RR & FSK\\
123:  \hline
124: No.\ of %&&&&&&&\\
125: patients & 103    & 8   & 11   & 16  & 7   & 11 & 6 \\
126: \hline
127: Mean age & $45.5$ & $65.4$ & $50.0$ & $55.9$ & $47.3$ & $55.5$ & $11.7$\\
128: \hline
129: Std.\ dev.\ %&&&&&&&\\
130: of age & $20.5$ &  $11.4$ & $19.3$ &  $14.3$ &  $11.6$ &  $14.4$ &  $4.6$ \\
131: \hline
132: \end{tabular}
133: \caption{
134: Test groups of patients. Abbreviations are as follows: IHD - Ischemic Heart Disease (Stenocardia); SND - Sinus Node Disease; 
135: VES - Ventricular Extrasystole; PCI - Post Cardiac Infarction; RR - Blood Pressure Disease; 
136: FSK - Functional Disease of Sinus Node.}
137: \end{table}
138: 
139: Originally, the Zipf's law addressed the distribution of words in a language \cite{Zipf}:
140: every word has assigned a rank, according to its  ``size'' $f$, 
141: defined as the relative number of occurrences in some long text 
142: (the most frequent word obtains rank $r=1$, the second frequent --- $r=2$, etc).
143: The empirical size-rank distribution law $f(r) \propto r^{-\alpha}$
144: is surprisingly universal:
145: in addition to all the tested natural languages, it applies
146: also to many other phenomena.  
147: The scaling exponent $\alpha$ is often close to one 
148: (e.g.\ for the distribution of words).
149: Typically, the Zipf's law is applicable to a
150: dynamical system at statistical equilibrium, when the following conditions are
151: satisfied: {\em (a)} the system consists of elements of different size; {\em (b)}
152: the element size has upper and lower bounds; {\em (c)} there is no intermediate 
153: intrinsic size for the elements.
154: As already mentioned, the human heart rhythm has a complex structure, where 
155: the duration $\tau$ of the low-variability periods varies in a wide 
156: range of scales, from few to several hundreds of heart beats.
157: Thus, one can expect that the 
158: distribution of the low-variability periods follows the Zipf's law
159: \begin {equation}
160: r \propto \tau ^{-\gamma}.
161: \end {equation}
162: However, the scaling behavior should not be expected to be perfect.
163: Indeed, the heart rate is a non-stationary signal affected by the non-reproducible daily 
164: activities of the subjects. The non-stationary pattern of these activities, together with their time-scales,
165: is directly reflected in the above mentioned distribution law.
166: This distribution law can also have a fingerprint of the characteristic time-scale (around ten to twenty seconds) 
167: of the blood pressure oscillations. 
168: Finally, there is a generic reason why the Zipf's law fails (or is non-perfect) at small rank numbers.
169: The Zipf's law is a statistical law; meanwhile, each rank-length curve is 
170: based on a single measurement. Particularly, there is only one longest low-variability period (and likewise, only one 
171: most-frequent word), the length of which is just as long as it happens to be, there is no averaging whatsoever.
172: For large rank values $r$, the relative statistical uncertainty can be estimated as $1/\sqrt{r}$.
173: %The effect of this phenomenon is reduced by the fact that the short-time variability of the heart rate variability level is less 
174: %While the mean heart rate is directly affected by the blood pressure ; 
175: %, which has a typical duration of ten to twenty seconds and can lead to the 
176: 
177: 
178: To begin with, we 
179: define the local variability for each ($i$-th) interbeat interval 
180: as the deviation of the heart rate from the local average, 
181: \begin {equation}
182: \delta(i) = \frac {|t_{NN}(i)-\left<t_{NN}(i) \right>|}{\left<t_{NN}(i) \right>}.
183: \end {equation}
184: % where $t_{NN}$ is the interval between two adjacent non-arrhythmic beats. 
185: The angular braces denote the local average, %which is 
186: calculated using a narrow (5 beats wide) Gaussian weight function. Further, we introduce a 
187: threshold value $\delta_0$; $i$-th interbeat interval is said to have a low variability,
188: if the condition
189: \begin {equation}
190: \delta(i) \le \delta_0
191: \end {equation}
192: is satisfied.
193: A {\em low-variability period}  is defined as a set of consecutive 
194: low-variability intervals; its length $\tau$ is measured in the number of
195: heartbeats. Finally, all the low-variability periods are arranged according to their lengths and
196: associated with ranks. % (the longest period obtains rank $r = 1$). 
197: The rank of a period is plotted versus
198: its length in a logarithmic graph, see Fig.~1;
199: Zipf's law would correspond to a straight descending line.
200: 
201: \begin {figure}[!bth]
202: \includegraphics{fig1.eps}
203: \caption{Multi-scaling distribution of the low-variability periods: the rank $r$ of a period is 
204: plotted versus its duration $\tau$ (measured in heartbeats) for different values of the threshold parameter 
205: $\delta_0$.}
206: \end{figure}
207: For a very low threshold parameter $\delta_0$, 
208: all the low-variability periods are very short, because it is difficult to satisfy the stringent condition (3).
209: In that case, the inertial range of scales is too short for a meaningful scaling law. On the other hand,
210: for a very high value of $\delta_0$, there is a single low-variability period occupying the entire
211: HRV-recording. Between these two %extreme 
212: cases, there is such a range of the values of $\delta_0$, which 
213: leads to a %Zipf's law-like behavior with $\gamma > 1$. 
214: non-trivial rank-length law.
215: For a typical healthy patient, the $r(\tau)$-curve 
216: is reasonably close to a straight line, and the scaling exponent $\gamma$ is  a function
217: of the threshold parameter $\delta_0$. Thus, unlike all the other well-known applications of the Zipf's law, 
218: we are dealing with a {\em multi-scaling law}. 
219: 
220: Recently, Ivanov et.\ al.\ \cite{Ivanov} have reported that anomalous multifractal spectra
221: of the HRV signal indicate an increased risk of sudden cardiac death. 
222: Therefore, it is natural to ask, does the presence or failure 
223: of the multiscaling behavior indicate the healthiness of the patient?
224: In what follows we discuss a somewhat more general question: what is the relationship between 
225: the properties of the distribution function of the low variability periods and the diagnosis 
226: of the patient. Testing the prognostic significance 
227: for predicting sudden cardiac death, which is also of a great importance, has been
228: postponed due to the nature of our test groups. 
229: 
230: 
231: First, let us analyze the correlation between the diagnosis of a patient and the 
232: scaling exponent $\gamma$.
233: To begin with, we have to determine the optimal value for the threshold parameter $\delta_0$. 
234: For a meaningful analysis, the 
235: scaling behavior should be as good as possible.
236: It turned out that for a typical patient, the best approximation of the %distribution 
237: function $r(\tau)$ with a power law is achieved for $\delta_0 \approx 0.05$ (see Fig.\ 2a); 
238: in what follows, all the values of the exponent $\gamma$ are calculated for $\delta_0 = 0.05$.
239: It should be noted that for some
240: patients, the length-rank distribution is still far from a power law (see Fig.\ 2b).
241: \begin {figure}[!b]
242: \includegraphics{fig2.eps}
243: \caption{Rank-length curves for a patient with a good power law {\em (a)} and for a patient 
244: with no power law {\em (b)}. In both cases, the threshold parameter $\delta_0=0.05$.}
245: \end{figure}
246: 
247: 
248: The slope of a curve on the logarithmic plot is calculated using root-mean-square (rms)
249: fit for such a range of lengths 
250: $[\tau_{\mbox{\scriptsize start}}, \tau_{\mbox{\scriptsize end}}]$, for which %where 
251: the $r(\tau)$-curve is nearly a power law, and the scaling range width
252: $\Delta=\ln \tau_{\mbox{\scriptsize end}} - \ln \tau_{\mbox{\scriptsize start}}$
253: is as large as possible.
254: %$\tau_{\mbox{\scriptsize start}}$ and $\tau_{\mbox{\scriptsize end}}$ 
255: %Here we need a quantitative definition, what is ``nearly a power law''.
256: Bearing in mind the statistical nature of the Zipf's law and non-stationarity of
257: the underlying signal, we have chosen a not 
258: very stringent definition of  what is ``nearly a power law'', see Fig.~3.
259: %We have chosen the way illustrated in Fig.~3. 
260: Around the rms-fit-line, two limit lines are drawn; 
261: $\tau_{\mbox{\scriptsize start}}$ and $\tau_{\mbox{\scriptsize end}}$ correspond 
262: to the points, where the $r(\tau)$-curve crosses the limit lines.
263: 
264: \begin {figure}[!b]
265: \includegraphics{fig3.eps}
266: \caption{Definition of the width of the scaling interval $\Delta$. The rank-length curve is fitted with 
267: a power law; the boundaries of the scaling interval are defined as the 
268: intersection points of limit lines and $r(\tau)$-curve.}
269: \end{figure}
270: Note that the precise placement and shape of the limit lines is arbitrary,
271: i.e.\ small variations do not lead to qualitative effects.
272: Here, the distance of the limit lines from the central line has been chosen to be 
273: $\ln 2$ at $\tau = \tau_{\max}$, and zero at $\tau = 1$,
274: where $\tau_{\max}$ is the length of the longest low-variability period. 
275: Admitting mismatch  $\ln 2$ at $\tau = \tau_{\max}$ is motivated by the observation that 
276: due to the lack of any statistics, 
277: the longest low-variability period could have been easily twice as long as we measured it to be.
278: %The statistical uncertainty of the rank $r$ is assessed as $\sqrt{r}$; 
279: For large rank values, the statistical uncertainty is assessed as $\sqrt{r}$;
280: in logarithmic graph, this would correspond to an exponentially decreasing (with increasing $r$) 
281: distance between a limit curve and the central fit-line. However, the above mentioned
282: effect of the non-stationary pattern of the subjects daily activities makes the situation 
283: more complicated. There is no easy way to quantify this effect and therefore, we opted for the simplest 
284: possible solution, simple straight limit lines.
285: 
286: The scaling exponent $\gamma$ has been calculated for all the patients and 
287: Student test was applied to every pair of groups. In most cases, the 
288: significance was quite low; two best distinguishable groups were RR and FSK, the
289: result of Student test being 5.7\%. Therefore, one can argue that 
290: the slopes of linear parts are highly personal characteristics depending also on the
291: daily habits of the subjects, which are weakly correlated with diagnosis.
292: 
293: Further we tested, how is the failure of the power law correlated with the diagnosis.
294: %Note that 
295: The width of the scaling range $\Delta$
296: was %can be 
297: used as a measure of how well the curve is corresponding to a power law. 
298: The Student test results for 
299: the parameter $\Delta$ turned out to be  similar
300: to what has been observed for the parameter $\gamma$: the correlation between the 
301: failure of the power law and diagnosis was weak. Thus, a rank-length curve
302: resembling the one depicted by a dashed line in Fig.~2, does not hint to
303: heart pathology. 
304: It should be also noted that the dashed curve in Fig.~2 can be considered as a {\em generalized 
305: form of scale-invariance} with scale-dependent  differential scaling exponent.
306: Such a %generalized scale-invariance % 
307: behavior 
308: seems to be universal; for instance, certain forest fire models \cite{Chen} 
309: lead to the differential fractal dimension depending on the local scale.
310: 
311: 
312: Finally, we analyzed the diagnostic significance of the parameters $\ln \tau_{\mbox{\scriptsize end}}$  
313: and $\ln \tau_{\mbox{\scriptsize start}}$. %Such an 
314: This analysis does make sense, because typically, 
315: the start- and end-points of the scaling range correspond to 
316: certain physiological time-scales. 
317: The parameter $\ln \tau_{\mbox{\scriptsize end}}$
318: %turned out to 
319: provided, indeed, a remarkable resolution between the groups of patients, see Table 2.
320: According to the Student test, the
321: healthy patients, were distinct from five heart pathology groups with  probability
322: $p<1.6$\%. The parameter $\ln \tau_{\mbox{\scriptsize start}}$
323: was diagnostically less significant.
324: %\begin{table}[!tbh]
325: %\includegraphics{table2a.eps}
326: %\caption{$p$-values of the Student test.
327: %Data in the left-bottom half are calculated using the parameter
328: %$r_{\max}$, data in the right-top half --- using $\tau_{\mbox{\scriptsize end}}$. 
329: %Gray background corresponds to values $p < 10\%$}
330: %\end{table}
331: %\begin{table}[!tbh]
332: %\includegraphics{table1a.eps}
333: %\caption{$p$-values of the Student test.
334: %Data in the left-bottom half  are calculated using the parameter
335: %$\tau_{40}$, data in the right-top half --- using $r_{100}$.}
336: %\end{table}
337: \begin{table}[!tb]
338: \includegraphics{table.eps}
339: \caption{$p$-values of the Student test.
340: Data in the topmost triangular region (with label {\em A}) are calculated using 
341: the parameter $\ln \tau_{\mbox{\scriptsize end}}$. Triangular region {\em B} corresponds to the parameter 
342: $\ln r_{\max}$, region {\em C} --- to $\ln r_{100}$, and region
343: {\em D} --- to  $\ln \tau_{40}$. Gray background highlights small $p$-values, $p < 10\%$.}
344: \end{table}
345: \begin{figure}[!!tb]
346: \includegraphics{fig4.eps} 
347: \caption{Definition of the parameters $r_{50}$, $r_{100}$, and $r_{200}$.}
348: \end{figure}
349: 
350: 
351: Unfortunately, the calculation of the parameter $\tau_{\mbox{\scriptsize end}}$ is technically quite 
352: a complicated task, not suited for clinical practice. Therefore, we aimed to 
353: find a simpler alternative to it.
354: Basically, the strategy was to find a simple parameter reflecting the behavior of
355: the rightmost (large-$\tau$) part of the $r(\tau)$-curve.
356: An easy option is $\ln \tau_{\max}$, which has been already analyzed \cite{JKMS}.
357: This parameter has indeed a considerable diagnostic value, but
358: its reliability is decreased by the above discussed statistical fluctuations.
359: Better alternatives are provided by {\em (a)} the overall number of low-variability periods  $r_{\max}$ 
360: (which is small, if there are lot of long low-variability periods); 
361: {\em (b)} the coordinates of specific points of the rank-length curve.
362: Here  we chose a set of critical ranks $R =10$, $20$ or $40$, and determined the respective
363: lengths $\tau_R$ so that $r(\tau_R)=R$. We also fixed  a set of critical 
364: length values, $T=50$, $100$, or $200$, and determined the respective
365: rank numbers $r_T = r(T)$, see Fig.\ 4.
366: %In full agreement with our expectations, 
367: Both techniques
368: turned out to be of high diagnostic performance;
369: %(in particular, better than the one based on $\ln \tau_{\max}$, discussed in \cite{JKMS}). 
370: illustrative $p$-values are given in Table 2. %and 3. 
371: Parameters $\tau_{10}$ and $\tau_{20}$ %are not presented, % in tabular data,  because they 
372: performed less well than  $\tau_{40}$ (for instance, 
373: the $p$-values for the healthy and VES-subject groups were 0.60\%, 0.58\% and 0.34\%, respectively),
374: and are not presented in tabular data.
375: Similarly, $r_{100}$ turned out to be more efficient than $r_{50}$ and $r_{200}$
376: (the respective healthy and VES-group $p$-values  being $0.02$\%, $0.01$\%, and $0.09$\%).
377: It also outperforms $\tau_{40}$, but is sometimes less efficient than $r_{\max}$ 
378: or $\tau_{\mbox{\scriptsize end}}$ (see Table 2).
379: Hence, various heart pathologies seem to affect the 
380: heart rate dynamics at the time scale around 100 heart beats (one to two minutes).
381: 
382: In conclusion, new aspect of non-linear time-series has been discovered,
383: the scale-invariance of low-variability periods.
384: We have shown that the distribution of low variability periods in the activity of
385: human heart rate typically follows a {\em multi-scaling Zipf's law}. 
386: The presence or failure of a power law, as well as 
387: the values of the scaling exponents, are personal characteristics depending 
388: on the daily habits of the subjects.
389: Meanwhile, the distribution function of the low-variability periods as a whole 
390: contains also a significant amount of diagnostically valuable information, the most part of 
391: which is reflected by the parameters $r_{100}$, $r_{\max}$, and $\tau_{\mbox{\scriptsize end}}$, see
392: Table 2. %and 3. 
393: These quantities characterize the complex structure of 
394: HRV signal, where the low- and high variability periods are deeply intertwined, aspect which 
395: is not covered by the other methods of heart rate variability analysis (such as 
396: fractional Brownian motion based multifractal analysis). As a future development, it would be of 
397: great importance to analyze the prognostic value of the above mentioned parameters for patients with sudden cardiac death.
398: % is an inherent feature of various time-series
399: 
400: 
401: The support of ESF %Etonian Science foundation 
402: grant No.\ 4151 is acknowledged.
403: 
404: \begin {thebibliography}{10}
405: 
406: \bibitem {Amaral} L.A.N.\ Amaral, A.L.\ Goldberger, P.Ch.\ Ivanov, and H.E.\ Stanley, 
407: %{\em Scale-Independent Measures and Pathologic Cardiac Dynamics.} 
408: Phys.\ Rev.\ Lett.\ {\bf 81}, 2388,%--2391 
409: (1998). 
410: 
411: 
412: \bibitem {Ivanov}
413: P.Ch.\ Ivanov  et al, %{\em Multifractality in Human Heartbeat Dynamics,} 
414: Nature, {\bf 399}, 461  (1999).%--465.
415: 
416: 
417: \bibitem {Thurner} S.\ Thurner, M.C.\ Feurstein, and M.C.\ Teich, 
418: %{\em Multiresolution wavelet analysis of heartbeat intervals discriminates
419: %healthy patients from those with cardiac pathology.}
420: Phys.\ Rev.\ Lett.\  {\bf 70}, 1544,%--1547 
421: (1998).
422: 
423: \bibitem {Saermark} K.\ Saermark et al, %Comparison of recent methods of analyzing heart rate variability
424: Fractals, {\bf 8}, 315,%-322
425: (2000).
426: 
427: 
428: \bibitem {Hon} E.H.\ Hon and S.T.\  Lee, 
429: %{\em Electronic evaluations of the fetal heart rate patterns preceding 
430: %fetal death, further observations.} 
431: Am.\  J.\ Obstet.\  Gynec.\  {\bf 87}, 814, %--826 
432: (1965).
433: 
434: 
435: \bibitem {Peng}
436: C.K.\ Peng et al., 
437: %{\em Long-range anticorrelations and non-Gaussian behavior of the heartbeat.}
438: Phys.\ Rev.\ Lett.\, {\bf 70}, 1343,%--1347.
439: (1993).
440: 
441: \bibitem {Poon}
442: C.S.\ Poon and C.K.\   Merrill,
443: %{\em Decrease of cardiac chaos in congestive heart failure.} 
444: Nature, {\bf 389}, 492,%--495 
445: (1997).
446: 
447: \bibitem {KSK} J.\ Kalda, M.\ S\"akki,  R.\ Kitt, unpublished.
448: 
449: \bibitem {Struzik}
450: Z.R.\ Struzik, 
451: %Revealing local variability properties of human heartbeat intervals with the local effective hölder exponent
452: Fractals, {\bf 9}, 77%-93 
453: (2001).
454: 
455: \bibitem {Chen}
456: K.\ Chen and P.\ Bak, Phys.\ Rev.\ E, {\bf 62}, 1613 (2000). %-1616 Scale-dependent dimension in the forest fire model
457: 
458: \bibitem {JKMS}
459: J.\ Kalda, M.\ Vainu, and M.\ S\"akki,  
460: %{\em  The methods of nonlinear dynamics in the analysis of heart rate variability for children,}
461: Med.\ Biol.\ Eng.\ Comp.\ {\bf 37}, 69%--72.
462: (1999).
463: %
464: \bibitem {Zipf} G.K.\ Zipf, {\em Human Behavior and the Principle of Least Effort} (Cambridge, 
465: Addison-Wesley, 1949).
466: \end{thebibliography}
467: 
468: 
469: 
470: 
471: 
472: \end{document}
473: