nlin0312027/HRV.tex
1: \documentclass[aps,pre,showpacs,floatfix]{revtex4} 
2: \usepackage{epsfig} 
3:   
4: \begin{document}  
5:   
6: \def\r{{\rm r}}  
7: \def\dr{{\Delta{\rm r}}}   
8: \def\C{{\bf C}}  
9:  
10: \title{Hierarchical Structure in Healthy and  
11: Diseased Heart Rate Variability in Humans}  
12: \author{Emily S.C. Ching$^1$, D.C. Lin$^2$ and C. Zhang$^1$}  
13: \affiliation{ 
14: $^1$Department of Physics, The Chinese University of Hong Kong,  
15: Shatin, Hong Kong.\\  
16: $^2$Department of Mechanical and Industrial Engineering, 
17: Ryerson University, Toronto, Canada.}  
18:   
19: \begin{abstract}  
20: It is shown that the heart rate variability (HRV)   
21: in healthy and diseased humans possesses a hierarchical structure  
22: of the She-Leveque (SL) form. This structure, first found in  
23: measurements in turbulent fluid flows, implies further details  
24: in the HRV multifractal scaling. The potential of diagnosis is also  
25: discussed based on the characteristics derived from the SL hierarchy.  
26: \end{abstract}  
27:   
28: \pacs{87.19.Hh, 87.10.+e}  
29:   
30: \maketitle  
31:   
32: The heart beat interval in humans is known to exhibit 
33: fluctuation which is referred to as heart rate
34: variability (HRV). Power spectrum analysis of the fluctuation 
35: revealed a $1/f$-like scaling \cite{ref3}. Recent studies indicated that
36: healthy human HRV exhibits even higher complexity which 
37: can be characterized by multifractal scaling \cite{ref1,ref2}. 
38: In contrast, HRV in the 
39: pathological state such as congestive heart failure
40: exhibits more monofractal-like scaling \cite{ref1}.
41: The change of the HRV 1/f law in congestive heart
42: failure is consistent with this result \cite{ref4}.
43: Such a multifractal-monofractal transition was also
44: reported in parasympathetic nervous system (PNS)
45: blockade experiment \cite{ref2}. Hence the
46: manifestation of multifractal HRV is indicative of the
47: proper autonomic regulation of the heart rate. Further
48: studies revealed that the multifractal HRV have
49: properties analogous to those found in fluid turbulence
50: \cite{ref5}. However, there is little understanding
51: beyond the phenomenological description of multifractal
52: HRV.
53:  
54: In this paper, we exploit further the 
55: analogy of HRV to fluid turbulence
56: and show the existence of a hierarchical 
57: structure in healthy and diseased HRV. This structure
58: allows us to model the multifractality of HRV and make
59: conjecture to the heart beat dynamics responsible for
60: the multifractal scaling. The hierarchy, first proposed
61: by She and Leveque (SL) to understand the statistical
62: properties of turbulent flows, provides a successful
63: framework to discuss and characterize the deviation
64: from Kolmogorov monofractal scaling in fluid turbulence
65: \cite{ref6}. When applied to study HRV, the SL
66: hierarchy provides a model structure which possesses
67: two advantages: (a) it simplifies the functional
68: description of the multiscaling by using a maximum of
69: only three parameters, and (b) it contains predictive
70: power for HRV scaling in physiological states related
71: to PNS withdrawal. One immediate implication is the
72: potential use of this notion in applications such as
73: diagnosis.
74:  
75: Let the beat-to-beat RR interval (RRi) be $\r(t)$, where 
76: $t$ is the beat number, and its increment be $\dr(\tau)= 
77: \r(t+\tau)-\r(t)$. The SL hierarchy implies, for a range 
78: of $\tau$, 
79: \begin{equation} 
80: \left[{S_{p+2}(\tau)\over S_{p+1}(\tau)}\right] =  
81: A_{p}\left[{S_{p+1}(\tau)\over S_p(\tau)}\right]^\beta  
82: \left[S^\infty(\tau)\right]^{1-\beta}. 
83: \label{1}  
84: \end{equation} 
85: Here $0 < \beta < 1$ is a parameter 
86: of the hierarchy, $A_p$, a function of $p$, $S_p(\tau)= 
87: \langle|\dr(\tau)|^p\rangle$, the $p$-th order moment of 
88: $|\dr(\tau)|$ denoted as the $p$-th order RRi structure 
89: function, $S^\infty(\tau)\equiv\lim_{p\to\infty}S_{p+1}( 
90: \tau)/S_p(\tau)$ and $\langle\cdot\rangle$ denotes 
91: statistical average. Since $S^\infty(\tau)$ is dominated 
92: by the statistics of large $\dr(\tau)$, it characterizes
93: the most intense fluctuation in HRV. Moreover, given the
94: empirical law $S_p(\tau)\sim\tau^{\zeta(p)}$ in HRV
95: \cite{ref5}, the hierarchy (\ref{1}) implies \cite{ref6} the scaling
96: model 
97: \begin{equation} 
98: \zeta(p) = h_0 p + C ( 1 - \beta^p)  
99: \label{2}  
100: \end{equation}  
101: where $h_0$ and $C$ are two other parameters of the hierarchy. 
102: It follows from (\ref{1}) and (\ref{2}) that $S^\infty(\tau)
103: \sim\tau^{h_0}$. A nonlinear functional dependence of $\zeta(p)$
104: on $p$ indicates multifractal scaling. Thus, the parameter
105: $\beta$ measures the degree of multifractality. In particular,
106: $\beta\to 1$ leads to monofractal scaling. In the multifractal
107: description of fluid turbulence, the parameter $C$ can be shown
108: to be the codimension of the most intense structure of the flow
109: \cite{ref6}. Since it is not yet possible to write down the
110: equation of motion for long-term cardiovascular dynamical system,
111: we assume a working definition for $C$ as the ``codimension
112: parameter" of the signal.
113:  
114: We follow the procedure developed in Ref. \cite{ref7} to check
115: whether the RRi data possess a SL hierarchical form. This
116: approach, based on the scaling property implied by the
117: hierarchy, the so-called generalized extended self-similarity
118: (GESS) in fluid turbulence \cite{ref8,ref9}, describes a
119: power-law relationship between the normalized structure
120: functions: 
121: \begin{equation}  
122: {S_p(\tau) \over [S_n(\tau)]^{p/n}} \sim   
123: \left\{{S_q(\tau) \over [S_n(\tau)]^{q/n}}\right\}^{\rho_n(p,q)} 
124: \label{5}  
125: \end{equation}  
126: In the case of SL hierarchy, the exponents $\rho_n(p,q)$ depends 
127: only on the model parameter $\beta$: 
128: \begin{equation}  
129: \rho_n(p,q) = {n(1-\beta^p)-p(1-\beta^n) \over  
130: n(1-\beta^q)-q(1-\beta^n)}. 
131: \label{6}  
132: \end{equation} 
133: It follows that 
134: \begin{equation}  
135: \Delta \rho_n(p+\delta p,q) = \beta^{\delta p} \Delta \rho_n(p,q)  
136:  - {\delta p(1-\beta^n)(1-\beta^{\delta p}) \over n(1-\beta^q)-q(1-\beta^n)}  
137: \label{7}  
138: \end{equation}  
139: where $\Delta \rho_n(p,q) \equiv \rho_n(p+\delta p,q)-\rho_n(p,q)$. 
140: One can then plot $\Delta \rho_n(p+\delta p,q)$ vs $\Delta
141: \rho_n(p,q)$ to check (\ref{7}) and hence the validity of the SL
142: hierarchy. We use several databases to perform the calculations (3)
143: and (5). The first database (DB1) contains 10 sets of daytime
144: ambulatory RRi recordings taken from healthy young adults \cite{ref5}.
145: The second database (DB2) contains 18 sets of daytime normal sinus
146: rhythm RRi data downloaded from public domain \cite{ref10}. We also
147: analyze RRi data from congestive heart failure patients (DBCHF) from
148: the same public domain \cite{ref10} to study the intrincity of the
149: hierarchy.
150:  
151: Except for a few cases where excessive ectopic beats in the RRi 
152: data complicates the calculation of $S_p(\tau)$, and were 
153: therefore discarded from the analysis, GESS is found in both 
154: healthy and CHF HRV. The exponent $\rho_n(p,q)$ is then 
155: estimated from (\ref{5}) and used in (\ref{7}) to calculate 
156: $\Delta\rho_n(p,q)$. Typical $\Delta \rho_n(p+\delta p,q)$ vs 
157: $\Delta \rho_n(p,q)$ plots are shown in Fig.~1. The observed 
158: linear trend implies (\ref{7}). Hence, SL hierarchy is 
159: compatible with the multifractal scaling in HRV. From such
160: plots, we estimate the value of $\beta$ simultaneously from
161: the slope and the intercept of the fitted straight lines. The
162: results are given in Fig.~2a. The $\beta$'s from healthy HRV
163: (DB1,DB2) cluster in the range [0.65,0.85] with those $\zeta 
164: (p)$ showing less curvature being characterized by larger 
165: $\beta$ values (Fig.~1). The $\beta$'s from DBCHF are generally
166: larger in values due to the monofractal-like scaling.
167:   
168: To gain insight of the hierarchy, She and Waymire (SW) arrived 
169: at the hierarchy (\ref{1}) using multiplicative random cascade 
170: \cite{ref11}. Their cascade consists of two dynamic components. 
171: One is the basic component that generates the singular 
172: dynamics over a continuum of scales. It can be shown that this 
173: dynamical component gives rise to the scaling term $h_0p$ in 
174: (\ref{2}). SW's cascade contains an extra component, which 
175: they called the ``defect dynamics" (DD), that modulates the 
176: singular structure through the multiplication of $\beta$ in 
177: discrete steps \cite{ref11}. It can be shown that this dynamical 
178: component contributes to the nonlinear term $C(1-\beta^p)$ in 
179: (\ref{2}). To apply SW's interpretation requires some 
180: clarification since continuous scale invariance does not exist 
181: in HRV, due to the fact that RRi fluctuation between heart 
182: beats cannot be defined. This suggests a dominant DD in the
183: generation of the multifractal scaling of HRV and a scaling
184: model with $h_0\sim 0$. Equivalently, this implies a hierarchy
185: with a $\tau$-independent $S^\infty$. Since $S^\infty$ cannot
186: be directly calculated, to verify such a model we first re-write
187: (\ref{1}) as  
188: \begin{equation}  
189: S_p(\tau) \sim [S^\infty(\tau)]^p\left\{S_q(\tau)  
190: \over S^\infty(\tau)^q\right\}^{\mu(p,q)}  
191: \label{8}  
192: \end{equation}  
193: where $\mu(p,q) \equiv (1-\beta^p)/(1-\beta^q)$ \cite{ref7}. We then 
194: form the quotient of (\ref{8}) at distinct values of $\tau$ and $\tau_0$, 
195: which, after some algebra, yields 
196: \begin{equation} 
197: \log_2\left[{S^\infty(\tau)  
198: \over S^\infty(\tau_0)}\right]  
199: ={\log_2[S_p(\tau)/S_p(\tau_0)]  
200: -\mu(p,q)\log_2[S_q(\tau)/S_q(\tau_0)]\over  
201: p-q\mu(p,q)}  
202: \equiv F_{p,q}(\tau,\tau_0).
203: \label{9}  
204: \end{equation}  
205: Hence, $F_{p,q}(\tau,\tau_0)$ is independent of $p$ and $q$ and  
206: a $\tau$-independent $S^\infty(\tau)$ implies a ``constant"
207: $F_{p,q}(\tau,\tau_0)$ over a range of $\tau$ and $\tau_0$
208: values. Figure~3 shows $F_{p,q}(\tau,\tau_0)$ for healthy and
209: CHF HRV. It is seen that the condition $h_0\sim 0$ can be
210: statistically ascertained and that $\zeta(p) \sim C(1-\beta^p)$.
211: Given this, $C$ is obtained by averaging $\zeta(p)/(1-\beta^p)$
212: over a range of $p$ and its result has been shown in Fig.~2b. 
213: Moreover, $C$, as a function $\beta$, shows an increasing trend
214: as $\beta\to 1$ (Fig.~4). This functional relationship is
215: consistent with the observations $h_0\sim 0$ and that $\zeta(p)$
216: becomes almost proportional to $p$ as $\beta \to 1$. It can
217: also be inferred from the earlier experimental studies, as we now explain.
218: 
219: Recall that the fractal dimension $D(h)$ of the set with a local
220: scaling exponent $h$ is related to $\zeta(p)$ through a Legendre
221: transform:
222: \begin{equation}  
223: D(h) = {\rm min}_p [ph + d - \zeta(p)]   
224: \label{3}  
225: \end{equation}  
226: where $d$ is the dimension of the embedding space. From (\ref{2})
227: and with $h_0\approx 0$, $D(h)$ can be explicitly obtained as: 
228: \begin{equation}  
229: D(h) = d - C + \left[ 1+ \ln C + \ln (\ln 1/\beta) \over \ln(1/\beta)  
230: \right]h    
231:  - {h \ln h \over \ln(1/\beta)}  
232: \label{4}  
233: \end{equation}  
234: Let $h^*$ be the scaling exponent of the singular
235: set with largest dimension, i.e., $D(h^*)$ is the maximum. 
236: For HRV, $h^*$ was
237: found to increase its value from the multifractal-like
238: scaling in healthy state to the monofractal-like scaling
239: in the diseased and pathological states \cite{ref1,ref2}.
240: Using (\ref{4}), $h^*$ is derived explicitly as
241: \begin{equation}  
242: h^*=C\ln(1/\beta).  
243: \label{10}  
244: \end{equation}  
245: If $C$ is constant, $h^*$ decreases as monofractal scaling 
246: is approached ($\beta\to 1$), which contradicts what was 
247: observed \cite{ref1,ref2}. In order for $h^*$ to increase
248: as $\beta \to 1$, $C(\beta)$ must diverge as $1/\ln (1/
249: \beta)$. Indeed, we find that the dependence of $C$ on
250: $\beta$ can be well described by $0.2\beta/\ln(1/\beta)$
251: (Fig.~4). Thus we have the result $h^*\sim 0.2\beta$
252: with $h^*$ increasing with $\beta$ in accord with the
253: experimental observations \cite{ref1,ref2}.
254: Finally, the asymptotic behaviour of $C$ (as $\beta\to 1$)
255: presents a more favorable condition for diagnosis: i.e., the
256: more monofractal-like scaling falls into the range with
257: larger ``separation" of the $C$ values (see also Fig.~2b).  
258:   
259: In summary, we show that a hierarchical structure of the SL 
260: form exists in the healthy and diseased HRV. This property 
261: allows us to model the multifractal HRV in terms of only 
262: two parameters $C$ and $\beta$. Interestingly, $C$ and
263: $\beta$ are related by an empirical law captured in Fig.~4.
264: This finding is important for two reasons. First, the
265: empirical law appears universal and is capable of describing
266: both healthy and CHF data. Second, the divergence of $C$ as
267: $\beta\to 1$ implies potential in diagnosis using the
268: hierarchical structure. To find a model that is compatiable
269: with the current finding, we adopted SW's cascade which
270: leads to further implications beyond the phenomenological
271: description of multifractal HRV scaling. This model is
272: appealing in that it contains a modulating component (DD)
273: acting on the singular dynamics to ``tame" the
274: fluctuation (since $\beta<1$.) Its effect reminds us of
275: the function of feedback regulation in biological systems.
276: Hence, the SW's model provides a concrete example of how
277: (additive) feedback mechanism may be integrated
278: multiplicatively in a cascading structure to produce the
279: observed HRV phenomenology. Further experiments will be
280: needed to test and quantify these possibilities in more
281: detailed physiological terms.
282:   
283: \begin{center}  
284: {\bf Acknowledgment}  
285: \end{center}  
286: ESCC acknowledges the Hong Kong Research Grants Council (CUHK 
287: 4286/00P) for support. D.C. Lin would like to thank the hospitality
288: of the Physics Department at the Chinese University of Hong Kong,
289: the C.N. Yang Visiting Fellowship and Natural Science and
290: Engineering Research Council of Canada for supports.  
291:   
292:   
293: \begin{thebibliography}{99} 
294: \bibitem{ref3} M. Kobayashi and T. Musha, IEEE Tans. Biomed. Eng. {\bf 
295: 29}, 456 (1982). 
296: \bibitem{ref1}  
297: P.CH. Ivanov, L.A.N. Amaral, A.L. Goldberger, S. Havlin, M.G. 
298: Rosenblum, Z.R. Struzik and H.E. Stanley, Nature {\bf 399}, 
299: 461 (1999). 
300: \bibitem{ref2}  
301: L.A.N. Amaral, P.Ch. Ivanov, N. Aoyagi, I. Hidaka, S. 
302: Tomono, A.L. Goldberger, H.E. Stanley, Y. Yamamoto, Phys. Rev. 
303: Lett. {\bf 86}, 6026 (2001). 
304: \bibitem{ref4} 
305: G.C. Butler, J. Floras, Clin. Sci. {\bf 92}, 545 (1997). 
306: \bibitem{ref5}  
307: D.C. Lin and R.L. Hughson, Phys. Rev. Lett. {\bf 86}, 
308: 1650 (2001); D.C. Lin, Fractals {\bf 11}, 63 (2003). 
309: \bibitem{ref6}  
310: Z-S. She and E. Leveque, Phys. Rev. Lett. {\bf 72}, 336 (1994). 
311: \bibitem{ref7}  
312: E.S.C. Ching, Z-S. She, W. Su, Z. Zou, Phys. Rev. E 
313: {\bf 65}, 066303 (2002). 
314: \bibitem{ref8} 
315: R. Benzi, L. Biferale, S. Ciliberto, M. V. Struglia, 
316: and R. Tripiccione, Europhys. Lett. {\bf 32}, 709 (1995). 
317: \bibitem{ref9} R. Benzi, L. Biferale, S. Ciliberto, M. V. Struglia, 
318: and R. Tripiccione, Physica D {\bf 96}, 162 (1996). 
319: \bibitem{ref10}  
320: http://physionet.org. See also
321: A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, 
322: P.Ch. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.-K. Peng, and H.E. 
323: Stanley, {\it Circulation}, {\bf 101}, e215 (2000). 
324: \bibitem{ref11}  
325: Z-S. She and E.C. Waymire, Phys. Rev. Lett. {\bf 74}, 262 (1995). 
326: \end{thebibliography} 
327:  
328: \newpage 
329:  
330: \begin{figure}
331: \centering
332: \includegraphics[width=.45\textwidth,angle=90]{fig1.eps}
333: \caption{
334: Typical $\Delta\rho_n(p+\delta p,q)$ vs.
335: $\Delta\rho_n(p,q)$ plots for (a) healthy HRV and (b) CHF
336: HRV with $\delta p=0.2$. We used $q=1$ and $n=2$ for
337: healthy HRV and $q=0.8$ and $n=1.2$ for CHF HRV. The
338: arrow direction indicates the increasing $p$ direction.
339: The estimated slopes for the shown cases are $\beta^{
340: \delta p}$ = 0.9422 and 0.9881, respectively. Hence,
341: $\beta\sim(0.9422)^5=0.7425$ and $(0.9881)^5\sim 0.9491$,
342: respectively. The insets show the corresponding
343: $\zeta(p)$. It is seen that a large $\beta$ in CHF HRV
344: implies a $\zeta(p)$ with less curvature, i.e.,
345: monofractal-like scaling.}
346: \end{figure}
347: 
348: \begin{figure}
349: \centering
350: \includegraphics[width=.95\textwidth]{fig2.eps}
351: \caption{Estimated values for model parameters (a) $\beta$ and
352: (b) $C$. Estimates from DB1, DB2 for 24 healthy subjects
353: are given in circles and those from DBCHF for
354: 30 congestive heart failure subjects are given in squares.
355: Means and standard deviations in both cases are also shown by the
356: bar-chart.}
357: \end{figure}
358:   
359: \newpage
360: 
361: \begin{figure}
362: \centering
363: \includegraphics[width=.6\textwidth,angle=90]{fig3.eps}
364: \caption{Evidence of $\tau$-independent $S^\infty
365: (\tau)$. $F_{p,2}(\tau,64)$ vs. $\log_2(\tau)$ from (a) a 
366: healthy subject ($p=0.2\sim 5$) and (b) a CHF patient
367: ($p=0.4\sim 2.6$). The $\beta$ needed in the calculation
368: of $\mu(p,q)$ (see \ref{9}) is obtained from that
369: estimated by (\ref{7}).}
370: \end{figure}
371: 
372:   
373: \begin{figure}
374: \centering
375: \includegraphics[width=.45\textwidth]{fig4.eps}
376: \caption{$C$ vs. $\beta$ empirical law for healthy
377: subjects (open circles) and congestive heart failure
378: patients (solid circles). The solid line is the
379: fit $C = 0.2\beta/\ln(1/\beta)$.}
380: \end{figure}
381: 
382: \end{document}
383: