1: % todo: Rever biografia Raya:2003

2: 

3: \documentclass[aps,preprint,amsmath,amssymb]{revtex4}

4: %\documentclass[aps,twocolumn,amsmath,amssymb]{revtex4}

5: %\documentclass[aps,pre,preprint,groupedaddress,floatfix]{revtex4}

6: 

7: %\documentclass[aps,twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}

8: %\documentclass[preprint,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}

9: %\documentclass[twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}

10: \usepackage{graphicx}% Include figure files

11: \usepackage{dcolumn}% Align table columns on decimal point

12: \usepackage{bm}% bold math

13: \usepackage{amssymb}

14: \usepackage{amsmath}

15: %\nofiles

16: 

17: \bibliographystyle{apsrev}

18: 

19: \begin{document}

20: 

21: %\preprint{APS/123-QED}

22: 

23: \title{Spectral Detrended Fluctuation Analysis and

24: Its Application to Heart Rate Variability Assessment}

25: 

26: 

27: \author{Luciano da Fontoura Costa$^1$} 

28: \email{luciano@if.sc.usp.br}

29: 

30: \author{Ruth Caldeira de Melo$^2$}

31: \author{\'Ester da Silva$^3$}

32: \author{Audrey Borghi-Silva$^2$}

33: \author{Aparecida Maria Catai$^2$}

34: 

35:  \affiliation{// 1 -- Instituto de F\'{\i}sica de S\~{a}o Carlos, Universidade

36:  de S\~{a}o Paulo, Av. Trabalhador S\~{a}o Carlense 400, Caixa Postal

37:  369, CEP 13560-970, S\~{a}o Carlos, S\~ao Paulo, Brazil \\

38: 2 -- N\'ucleo de Pesquisas em Exerc\'{\i}cio F\'{\i}sico,

39: Laborat\'orio de Fisioterapia Cardiovascular, Universidade Federal

40: de S\~ao Carlos, SP, Brazil \\

41: 3 -- Faculdade de Ci\^encias da Sa\'ude, Universidade Metodista de

42: Piracicaba, SP, Brazil}

43: 

44: 

45: \date{5th June 2005}

46: 

47: 

48: \begin{abstract}

49: 

50: Detrend fluctuation analysis (DFA) has become a choice method for

51: effective analysis of a broad variety of nonstationary signals. We

52: show in the present article that, provided the nonstationary

53: fluctuations occur at a large enough time scale, an alternative

54: approach can be obtained by using the Fourier series of the signal.

55: More specifically, signal reconstructions considering Fourier series

56: with increasing number of higher spectral components are subtracted

57: from the signal, while the dispersion of such a difference is

58: calculated.  The slope of the loglog representation of the dispersions

59: in terms of the time scale (reciprocal of the frequency) is calculated

60: and used for the characterization of the signal.  The detrend action

61: in this methodology is performed by the early incorporation of the low

62: frequency spectral components in the signal representation.  The

63: application of the spectral DFA to the analysis of heart rate

64: variability data has yielded results which are similar to those

65: obtained by traditional DFA.  Because of the direct relationship with

66: the spectral content of the analyzed signal, the spectral DFA may be

67: used as a complementary resource for characterization and analysis of

68: some types of nonstationary signals.

69: 

70: \end{abstract}

71: 

72: %\pacs{Valid PACS appear here}

73: 

74: \maketitle

75: 

76: \section{Introduction}

77: 

78: Several natural signals are characterized by nonstationarity, i.e. the

79: variation of its statistical description along time.  Introduced by

80: Peng and collaborators~\cite{Peng:1994}, detrend fluctuation analysis

81: --- DFA --- has proven to be particularly effective for coping with

82: nonstationary signals, providing results which tend to be more robust

83: than previous methods such as those based on the Hurst coefficient.

84: One of the interesting features of DFA is its ability to reveal the

85: fractal nature (e.g.~\cite{Bunde_Havlin1,Bunde_Havlin2}) of

86: one-dimensional signals.  A particularly impacting application of DFA

87: has been the characterization of heart rate variability (HRV)

88: (e.g.~\cite{Iyengar:1996}), as estimated from sequences of interbeat

89: intervals (RRi).  The traditional approach to such a problem involves

90: normalizing the interbeat sequence so as to achieve null mean and unit

91: variance, integrating the resulting sequence, dividing it into boxes

92: of fixed size, obtaining linear fit (or higher polynomial

93: approximations) inside each box, and estimating the standard deviation

94: of the difference between the sequence and linear fit.  The absolute

95: value of the inclination of the loglog curve yields the $\alpha$

96: parameter, which has been shown to be useful for applications to

97: several analyses.  Note that all logarithms in this article are

98: natural logarithms.

99: 

100: Many interesting results have been obtained by DFA analysis of RRi

101: sequences, including the fact that the coefficient $\alpha$ provides

102: good sensitivity to the age of the analyzed individuals, with younger

103: subjects tending to present smaller values of $\alpha$

104: (e.g.~\cite{Iyengar:1996, Acharya:2004,

105: Lipsitz_circulation,Peng_Mietus:2002}).  This effect is to a great

106: extent a consequence of the fact that the heart rate of youngers tend

107: to exhibit higher small scale variability, which populates the

108: lefthand side of the loglog curve, therefore decreasing the value of

109: $\alpha$.  At the same time, complementary studies have shown that the

110: effect of exercise in humans tend to lead to a reduction of the value

111: of $\alpha_1$, namely the slope of the loglog curve considering 4 to

112: 11 heatbeats\cite{Tulppo:2001}.  As with the aging effect, such a

113: reduced slope is also a consequence of the enhancement of small scale

114: information in the interbeat interval signals.  Additional works

115: dealing with the effect of exercise and physical fitness on heart

116: dynamics include but are not limited to~\cite{Mikko:2002,

117: Hautala:2003, Arja:2004, Arto:2004}.

118: 

119: While traditional DFA analysis has consolidated itself as an important

120: method for analysing heart dynamics, it is also interesting to

121: consider modifications of the primary approach which can lead to

122: eventual complementation of interpretation or improvements.  For

123: instance, Willson and collaborators ~\cite{Willson_1, Willson_2} have

124: shown that the $\alpha$ coefficient in DFA can be related to the

125: frequency-weighted Fourier spectrum of the original signal. The

126: present work reports on a modified DFA method where the piecewise

127: partition and fitting of the original signal is replaced by its

128: respective Fourier series with increasing number of coefficients.  In

129: addition to its inherent simplicity -- it only involves calculating

130: the fast Fourier transform and standard deviations, the proposed

131: methodology bears a direct relationship with the progressive

132: approximation of the signal by its Fourier series.  The potential of

133: the proposed DFA variant is illustrated with respect to human

134: interbeat interval analysis involving young and older subjects

135: submitted or not to regular exercises.  It is shown that the spectral

136: DFA leads to results which are similar to those obtained by using the

137: traditional DFA regarding the effects of age and physical

138: conditioning.

139: 

140: This article starts by briefly revising the traditional DFA,

141: describing the spectral DFA, and illustrating its application in the

142: characterization of interbeat interval analysis by using the

143: traditional and spectral DFA methods.

144: 

145: 

146: \section{Traditional DFA}

147: 

148: Typically, ECG is recorded during a fixed total period of time, and

149: the interbeat intervals are extracted afterwards (e.g. by considering

150: the signal peaks).  The interbeat intervals values are represented as

151: a discrete sequence $r$, such that $r(k)$ corresponds to the k-th

152: interbeat interval, with $k = 1, 2, \ldots, N$.  Let $\left< r

153: \right>$ and $\sigma_r$ stand for the average and standard deviation

154: of the sequence $r(k)$.  A normalized sequence $s(k)$ is obtained

155: through the transformation

156: \begin{equation}

157:   s(k) = \frac{r(k)- \left< r \right>}{\sigma_r}   \label{eq:normal}

158: \end{equation}

159: 

160: It can be shown that $s(k)$ has null average and unit variance.  The

161: sequence $s(k)$ is then divided into boxes of size $L = L_{min},

162: \ldots, L_{max}$ and linear interpolation is performed within each box

163: considering several box sizes.  The difference $d(k)$ between $s(k)$

164: and the obtained regression is calculated for each box.  The standard

165: deviation $sd(L)$ of the difference sequence $d(k)$ is estimated, and

166: a loglog curve is obtained for $sd(L)$ in terms of $L$.  In case the

167: loglog curve corresponds to a straight line, the original sequence

168: $r(k)$ is understood to present fractal strucutre, with the absolute

169: value of the inclination of the loglog curve, called $\alpha$,

170: providing a parameter which has shown to be particularly useful for

171: characterization of the heart rate variability, especially regarding

172: the age of the individuals~\cite{Iyengar:1996}, as well as the effect

173: of aerobic training~\cite{Tulppo:2001}. 

174: 

175: Figure~\ref{fig:DFA} illustrates two RRi signals typical of a

176: sedentary (a) and active older subjects (b) as well as the respective

177: loglog curves (c-d), considering box sizes ranging from 4 to 11

178: heartbeats.

179: 

180: 

181: \begin{figure}

182: \begin{center}

183: \includegraphics[scale=0.5]{dfa_sed.eps}  \hspace{0.5cm}

184: \includegraphics[scale=0.5]{dfa_act.eps}  \\

185: 

186: \caption{The RRi time series for a sedentary (a) and active (b) older

187: subjects. The respective loglog curves are shown in (c) and (d). The

188: values of $\alpha$ (i.e. the curve slope) obtained for these curves

189: were 0.98 and 0.94, respectively.}

190: \label{fig:DFA}

191: 

192: \end{center}

193: \end{figure}

194: 

195: 

196: 

197: 

198: \subsection{Spectral DFA}

199: 

200: The Fourier series of the normalized interbeat interval signal $s(k)$

201: can be calculated in matrix form as

202: 

203: \begin{eqnarray}

204:   S = W s \\

205:   where \nonumber \\

206:   W= [w_{i,j}], w_{i,j} = exp \{ -  2 \pi \sqrt{-1} \: i j / N   \} \\

207:   i, j = 0, 1, \ldots, N-1 \nonumber

208: \end{eqnarray}

209: 

210: The reconstruction $s_m(k)$ of $s(k)$ considering the $m$ first

211: Fourier coefficients (i.e. $0, 1, \ldots, m-1$) can be obtained as

212: 

213: \begin{eqnarray}

214:   s_m = Q S \\

215:   where \nonumber \\

216:   Q = [q_{i,j}], \\

217:   q_{i,j} = 

218:   \begin{cases}

219:      exp \{  2 \pi \sqrt{-1} \: i j /N \} &  i=0, 1, \ldots, N-1; \: j= 0, 1, \ldots, m-1 \\

220:     0  &  i=0, 1, \ldots, N-1; \: j= m, \ldots, N-1 \nonumber

221:   \end{cases}  

222: \end{eqnarray}

223: 

224: Note that this equation corresponds to the Fourier series of the

225: signal considering m spectral components, where the quantity $u=ij$ is

226: proportional to the frequency.  The difference between the original

227: signal and its reconstruction is therefore given as $d(k) = s(k) -

228: s_m(k)$, whose standard deviation is henceforth expressed as $sd(m)$.

229: 

230: The spectral DFA proposed in this work involves calculating $sd(m)$

231: for several values of $m$, ploting the curve of $log(sd(m))$ in terms

232: of $log(1/m) = -log(m)$, and obtaining the slope $\gamma$, which has a

233: similar role as the parameter $\alpha$ in the traditional DFA.  Note

234: that as the values of $m$ are proportional to the Fourier series

235: frequencies, the quantity $1/m$ is proportional to the period of the

236: sinusoidal functions in the Fourier kernel.  Thus, $1/m$ can be

237: understood as a time scale parameter (analogous to the box size in the

238: traditional DFA) underlying the modified DFA analysis.

239: 

240: The detrending effect of the above described approach is illustrated

241: in Figure~\ref{fig:trend}.  The original signal $s0(k)$ in (a) is

242: contaminated with a triangular trend function $tr(k)$, yielding the

243: signal $s(k)$ to be analyzed (b).  Because of the slow variation of

244: such a trend function, the Fourier power spectrum of $s(k)$ can be

245: separated into two groups, one with lower frequencies containing the

246: trend signal, at the left hand side in (c), and the other corresponding

247: to the original signal $s0$, at the righthand side in (c).  Provided

248: the original signal and trend contamination are reasonably

249: well-separated along the spectral decomposition, the trend effect can

250: be removed by the above difference procedure before the signal of

251: interest start influencing the slope $m$.  In other words, the

252: dispersion of the difference signal becomes independent of the trend

253: function.

254: 

255: \begin{figure}

256: \begin{center}

257: \includegraphics[scale=0.7]{ex_trend.eps} 

258: 

259: \caption{\label{fig:trend} An original signal of interest $s0(k)$ (a),

260: its contamination by a triangular function (b), and the respective

261: Fourier power spectrum (c).}

262: \end{center}

263: \end{figure}

264: 

265: 

266: 

267: Figures~\ref{fig:FDFA}(a) and (b) illustrate the Fourier power

268: spectrum for the same sedentary and active older individuals in

269: Figure~\ref{fig:DFA}. Note that the interbeat dynamics of the active

270: older, whose power spectrum is shown in (b), involves a more intense

271: higher frequency composition.  The respective loglog curves obtained

272: by the spectral DFA are illustrated in (c) and (d), respectively.  It

273: is clear from these curves that physical exercise has as effect the

274: reduction of the inclination of the loglog curve, indicating that

275: regular physical activity tends to add small scale detail (i.e. high

276: frequency components) to the interbeat dynamics.  Note that the second

277: curve in Figure~\ref{fig:FDFA} is not well fitted by the linear

278: regression line.  This is a consequence of the fact that the Fourier

279: series approximations considered in the modified DFA methodology

280: contain \emph{all} coefficients ranging from $0$ to $m$, implying

281: faster convergence to the original interbeat signal than that which

282: could be obtained by using polynomial piecewise fitting.  Note the

283: substantially higher inclination (slope) of the curve obtained for the

284: sedentary subject, reflecting its lack of high frequency composition.

285: 

286: \begin{figure}

287: \begin{center}

288: \includegraphics[scale=0.5]{fdfa_sed.eps} \hspace{0.5cm}

289: \includegraphics[scale=0.5]{fdfa_act.eps}  \\

290: 

291: \caption{\label{fig:FDFA} The Fourier spectrum for the same sedentary

292: (a) and active (b) older subjects considered in Figure~\ref{fig:DFA}.

293: The respective loglog curves, shown in (c) and (d), have slopes 

294: of 0.87 and 0.34, respectively.}

295: \end{center}

296: \end{figure}

297: 

298: 

299: 

300: \section{Interbeat Interval Analysis}

301: 

302: In order to illustrate the performance and features of the Fourier

303: DFA, this method has been applied to a database of 40 interbeat

304: interval sequences.  

305: 

306: The RRi database includes measuremeents obtained at the same time of a

307: day from 40 male individuals divided into four classes according to

308: age and fitness (i.e. sedentary or active), all with good health as

309: confirmed by extensive clinical, physical and laboratory examinations

310: detailed in previous studies~\cite{Catai1}.  The individual included

311: in the active classes had been through regular physical activities

312: over a long period of time (at least 15 years in the case of the older

313: group).  On the day of experiments, after further medical

314: examinations, the subjects remained at rest in supine position during

315: 20 minutes and then ECG was recorded for 15 minutes.  The ECG and

316: heart rate were acquired through a one-channel heart monitor (ECAFIX

317: TC500, Sp, Brazil) and processed by using an analog-digital converter

318: Lab.PC+ (National Instruments, Co., Austin, TX, USA), interfaced to a

319: personal computer.  After analog to digital conversion, the R-R

320: interval (ms) was calculated on a beat-to-beat basis by using a

321: customised software (SP, Brazil)~\cite{Catai2}.  At least 5 minutes of

322: ECG recording, characterized by the highest stability, were considered

323: for the DFA investigations in the current work.

324: 

325: After normal transformation by using Equation~\ref{eq:normal}, each

326: RRi signal was processed by both traditional and spectral DFA.  The

327: intervals considered for calculating the loglog slopes were from 4 to

328: 11 beats in the traditional DFA and from -4 to -2 (in log scale) in

329: the spectral method.  The scatterplot obtained by considering these

330: two measurements for each anlyzed human individual is shown in

331: Figure~\ref{fig:scatts} (a).  The normal density functions fitted for

332: each of the four classes considering each of the slopes is given in

333: (b) and (c).  It is clear from (a) that the results obtained by the

334: traditional and spectral DFA methods are highly correlated, while the

335: inspection of (b) and (c) indicates that similar separations between

336: the several classes are obtained by the two methodologies.  In both

337: cases, the sedentary individuals yielded loglog curve slope which is

338: higher than those for the active subjects.

339: 

340: 

341: \begin{figure}

342: \begin{center}

343: \includegraphics[scale=0.7]{scatts.eps} 

344: 

345: \caption{\label{fig:scatts} The scatterplot (a) obtained by

346: considering the slope of the loglog curves produced by the traditional

347: and spectral DFA methodologies. Density probability functions for the

348: four types of interbeat data obtained by normal fit of the

349: measurements obtained by using the traditional (b) and spectral (c)

350: DFA.}

351: 

352: \end{center}

353: \end{figure}

354: 

355: 

356: \section{Concluding Remarks}

357: 

358: This article has reported on a spectral variation of the detrend

359: fluctuation analysis, DFA.  Provided the trend effects are slow

360: varying in time, the suggested methodology allows the detrend of

361: higher frequency components of the signal of interest.  The slope of

362: the loglog curve of the dispersion against time scale has been shown

363: to have an interpretation analogous to the $\alpha$ coefficient in the

364: traditional DFA.  The behavior of the spectral DFA has been

365: illustrated with respect to interbeat analysis considering four types

366: of individuals, yielding discrimination results which are closely

367: correlated with those obtained by traditional DFA.  Because of its

368: spectral nature, the new type of DFA may be of interest for some

369: applications in which one want to have a more direct comparison of

370: relationship with the spectral content of the signals.

371: 

372: 

373: \begin{acknowledgments}

374: 

375: Luciano da F. Costa thanks FAPESP (proc. 99/12765-2), CNPq

376: (proc. 3082231/03-1) and Human Frontier for financial support.  Ester

377: da Silva thanks FAPESP (proc. 01/07427-2) and Aparecida Maria Catai

378: thanks CNPq (proc. 478799/2003-9).

379: 

380: \end{acknowledgments}

381: 

382: \bibliographystyle{unsrt}

383: 

384: %\newpage %Just because of unusual number of tables stacked at end

385: \bibliography{dfa}% Produces the bibliography via BibTeX.

386: 

387: \end{document}

388: 

389: 

390: 

391: 

392: 

393: