Você está na página 1de 7

Gzrdiovascular Research

Cardiovascular Research 3 1 (19%) 434-440

Characterization of complex heart rate dynamics and their pharmacological disorders by non-linear prediction and special data transformations
Dirk Hoyer * , Karin Schmidt, Ulrich Zwiener, Reinhard Bauer
Imritutfir Pathologische Physiologic, Medizinische Fakultiit der Friedrich Schiller Uniuersitiit Jena, D-07740 Jew, Germany

Received 9 February 1995; accepted 5 May 1995

Abstract Objective: The aim of the presented method was the characterization of different complex heart rate dynamics in conscious rabbits as well as during general anaesthesia and vagal blockade. This wasdoneby non-linear predictionof originalmeasured special and surrogate data in the phase space. Method: The development the predictionerror in dependence the predictiontime interval wasinvestigated of on

in the phase space. Two kindsof surrogate wereproduced tested data and with regardto non-linearities orientationin the phase and space.
Typical characteristics of prediction error development were shown for simulated uncorrelated stochastic, correlated stochastic, regular

deterministic,anddeterministic chaoticsignals. These characteristics were used to evaluate the measured heart rate data in connection with tests surrogate data. Results: It couldbe shown heartratefluctuationcannotbe described oneof these of that by idealmodels alone.
Common consideration of all investigated prediction characteristics indicated chaos in the heart rate of conscious rabbits as well as during

anaesthesia vagal blockade,where non-linearcorrelatedstochastic and propertiescould not he excluded.The different amountof
non-linearites and orientation was described quantitatively. Conclusions: Detailed analysis of prediction error development in the phase space, connected with testsfor non-linearities orientation,enableda specific quantitativecharacterization complex heart rate and of

dynamics their pharmacological and disorders.


Keywords:

Heart rate variability; Non-linear phenomena; Autonomic

nervous

system;

Rabbit,

anesthetized

1. Introduction

In healthy subjects the time coursesof the heart rate show irregular behaviour which changesto become more regular according to functional limitations because of anaesthesiaand particular blockades of the autonomous nervous system [l-S]. It could be shown that the seemingly stochastic behaviour of the heart rate fluctuations is causedby dynamic non-linear processes [6,71. Positive Lyapunov exponents and non-integer correlation dimensions, which are estimated from measured time

seriessuch as heart rate fluctuations (HRF9, can be based on different deterministic (chaotic) or stochastic processes

[8-101. The characteristicsof the underlying processwhich are to be distinguishedas far as possible are linearity and non-linearity, and more specifically regular determinism, deterministic chaos or different stochastic properties. A quantitative description of corresponding signal properties is needed, in particular for the classification of different pharmacological states. Theiler et al. [I I] proposed a test for non-linearity by comparing particular characteristicssuch as the correlation dimension, the forecasting error or the leading Lyapunov exponent of the original time serieswith those of phaserandomized versions of the original time series. Kennel and Isabella [ 121proposeda distinction between chaos and coloured noise on the basis of a forecasting

Corresponding author. Fax: (+ 49-3641)631212; 3641)631764, e-mail:iclh@hpux.rz.uni-jena.de.


l

Tel.: ( + 49Time for primary review 26 days.

0008-6363/%/$15.00
SSDI 0008-6363(95)00086-O

0 1996 Elsevier Science B.V. All rights reserved

D. Hoyer

et al./ Cardiovascular

Research

31 (19%)

434-440

435

function of a low-dimensional chaotic system and testing surrogate data similar to Theiler et al. [ 111. Sugihara and May [ 131 found that the development of the forecasting error is different in chaotic and uncorrelated stochastic time series, but that it cannot necessarily distinguish between chaos and coloured noise. Tsonis and Elsner [14] calculated a prediction error parameter which depends linearly on the prediction time, but only in chaotic time series. However, in correlated stochastic (fractal) time series they found a double logarithmic dependency between the prediction error parameter and the prediction time. This dependency is related to the Hurst coefficient, which is a scaling measure of fractal signals. Kaplan [ 15,161 proposed a search for the particular orientation of the trajectory and coarsed-grained embedding as tools for indicating determinism in phase space sets. In spite of these impressive approaches a clear differentiation between chaotic and non-linear correlated stochastic processes seems to be problematic in measured time series such as the HRF. Furthermore, it can be expected that the underlying physiological processes do not match exclusively with only one of these mathematical aspects. The aim of the present study was to analyse different non-linear characteristics of normal and pharmacologically disturbed HRF. Corresponding to the approaches referred to above, the representation of the time series in a multi-dimensional phase space forms an appropriate basis to investigate complex non-linear dynamics. We investigated prediction error development in the multi-dimensional phase space sets for different simulated signals with typical characteristics in comparison to measured heart rate data. Surrogate data sets were used to test for non-linearities and orientation in the phase space.

2. Methods
2.1. Experimental design and measurement

Six adult rabbits (White New Zealand, body weight 2.6-3.8 kg) were anaesthetized by i.m. injection of 22 mg/kg ketamine and 4 mg/kg xylazine. A selective vagal blockade was performed by S.C. injection of 3 mg/kg atropine. The corresponding states investigated were: Slconscious, S2-general anaesthesia, and S3-vagal blockade (atropine during general anaesthesia). Body temperature (rectal thermocouple) and arterial pressure were recorded. continuously. Arterial blood samples were withdrawn for a blood gas and acid-base state check. The ECG was obtained with a standard AC amplifier, and care was taken to record an analog signal with a prominent and positive R-wave that decreased the likelihood of errors in the following steps. After suitable amplification, the ECG was fed through an analog-to-digital (A/D) converter (DT2821, @, Data Translation, Inc., Marlboro, MA, USA) with a sample rate of 4096 Hz to the mass memory of the personal computer. Then the individual R-waves, with the steepest R-wave rise as the fiducial point, are sequentially recognized, and thus the series of R-R intervals is stored as a function of the beat sequence. This series constitutes the tachogram. The reciprocal of this series represents the instantaneous heart rate. Artifact-corrupted date were rejected. Only time series which were stationary during at least 8192 heart beats were investigated. In order to show the different time structures of HRF in Sl-S3 in more detail, parts of 100 heart beats are presented in Fig. 1. The mean heart rate did not differ significantly between the states (Sl 200 f 25, S2 175 f 48, S3 206 f 47 beats/min). The experimental investigations conform with the Guide for the Care and Use of Laboratory Animals published by

I 1 I

150

1 --I I

--

11

21

31

41
heart

51
beats

61

71

81

91

107

Fig. 1. Heart rate fluctuations, partsof 100 from 8192 heart beats. Sl-conscious; S2-general anaesthesia; S3-vagal block&e during aaesthesia (examples).

436

D. Hoyer

et al./ Cardiovascular

Research

31 (19961434-440

US National Institutes of Health (NIH 85-23, revised 1985).


2.2. Simulated data

publication

No.

:i E4I 21
0
7 3 5 7 9 77 13 75
At Fig. 3. Prediction error (I?,,,) in dependence on the prediction time interval (At), Large filled squares-original data (covered by the empty squares); small tilled squares-phase-randomized data; empty squares-randomly shuffled data. Simulated uncorrelated stochastic (random) signal (D,)= 16.

The measured heart rate data may consist of a mixture of different deterministic and stochastic parts. It cannot be presumed that only one kind of theoretically well-defined signal property will exist. In order to evaluate the proposed method of non-linear prediction and to find relations between the HRF and theoretical characteristics, different signals with typical well-defined characteristics were simulated. Data length (N = 8192) and amplitude normalization were chosen identical to the heart rate data.
Regular deterministic data. Regular deterministic

time se-

corresponding phases (cp):


1

scaled amplitudes

[P(o)]

and random

ries were simulated by sinus waves.


z(t) = ~~~cos(wit+pi) Deterministic chaotic data. As a deterministic chaotic time series z(t) the x-values of the Henon attractor (leading
i-

(2)

Lyapunov exponent = 0.419) were used:


X,+ 1

= 1 - 1.4x,*-yt,

y,+1 =0.3x,

(1)

with P(oi) = l/w,, (Y= 0.5 scaling parameter; wi = 2rrii/Ndt, N = data length, dt = sampling rate; 0 < p < 2~, cp uniformly distributed, m = N/2. This special kind of correlated stochastic signal is called fractal.
2.3. Phase space reconstruction

Uncorrelated stochastic (random) data. Uncorrelated

stochastic time series were simulated according to Knuth H71.


Correlated stochastic data. The l/f

power law spectra which are found in heart rate data [5,7] indicate statistically self-similar behaviour which is a special case of correlated stochastic signals. We simulated a corresponding time series [ z( t>] by superposition of harmonic waves with

The phase space sets were constructed from the time series z( t> according to Takens theorem [ 181. The coordinates of the points P( t> in the phase space are the time-delayed values of the time series; PC)
t = 1,2,3 . ..(N-D.) (3)

N is the number of data points P in the phase space. The embedding dimension (0,) is the number of coordinates of the phase space which is equal to the number of time-delayed values of the time series. A one-to-one map of a smooth time series needs D, 2 20 + 1 (D = dimension of the process) [19]. We used a time delay (7) equal to the time of the decay of the autocorrelation function to l/e according to Albano et al. [20]. The embedding dimension CD,) was increased until the estimated parameters became approximately constant.
2.4. Estimation of the meanprediction error (E,,,)

Fig. 2. Simulated deterministic (periodic) trajectory in the two-dimensional phase space. Z&l panel: Deterministic orientation of the original simulated trajectory. Right panel: Orientation of the randomly shuffled sequence of the identical points.

We investigated the prediction errors in dependency on the prediction time in the multi-dimensional phase space. Beginning from a starting point P(t,), the nearest Euclidean neighbour is searched for in the phase space. Its distance is t(t,). The distance L!(t, + At) between the developed points (after passing the prediction time interval At) is taken into account in order to estimate the prediction

D. Hoyer

et al./ Cardiovascular

Research

31 (1596)

434-440

431

7 3

5 7 9 77 13 15
At

8. U-X-6 I 64 2 1 = 0 7

;
At

_ 5

_ 7

Fig. 4. E, versus At, plot as in Fig. 3. Simulated chaotic signal (Henon attractor), D, = 8.

Fig. 5. E,,, versus At, plot as in Fig. 3. Simulated regular deterministic signal (sinus wave), D, = 8.

error (Ei). This procedure is repeated for every point Z(t,>. The single prediction errors (Ei> are estimated by: Ei = log, L(ti+At)
Lt i>

relations between E, and At in the simulated signals with different ideal characteristics. Non-linearities and orientation in the phase space were investigated by means of special transformations of these data.
2.5. Testfor non-linearities (phase randomization,J

The mean prediction error (E,,,) can be estimated as the mean value of all prediction errors (Ei) by:
1

N-At

E,,, = N-,*ZEi

The mean prediction error (E,), normalized to the prediction time interval (At), is an approximation of the leading Lyapunov exponent, if N, D,, and At are optimal [21]. We investigated E,,, for data sets of 8192 points in dependence on the prediction time interval At (Figs. 2-5, filled squares). In uncorrelated stochastic data the prediction errors are uniformly distributed over the simulated region of data values (Fig. 3). Consequently, no dependency of E, on At was found. The linear relation between E,,, and At over a particular prediction time region is caused by the exponential divergence of infinitesimal neigbouring pathways. This is typical of the short-time prediction of deterministic chaotic signals (Fig. 4). In the regular deterministic signal, there is obviously a vanishing prediction error (Fig. 5). In the simulated correlated stochastic signal, the dependency between E,,, and At is logarithmic (Fig. 6). According to Sugihara and May [ 131 the linear dependency between the logarithmic prediction error (E,, Eq. 4) and the logarithm of At could be caused by the scaling properties of this fractal signal. We found an increasing prediction error (E,,,> depending on a decreasing scaling parameter (01) (Eq. 2). In this section we could show different typical

Theiler et al. [ 111 proposed a comparison of the analysis of the original (measured) data with that of phase-randomized surrogate data sets which have the same linear properties but removed non-linear relations. If appropriate measures were statistically different, there would be evidence of non-linearity in the original data. In the present investigation the differences between the prediction errors (E,) of the original and the corresponding surrogate data sets were tested. If it is supposed that the surrogate data sets are linear and stochastic, the rejected null hypothesis indicates non-linear deterministic (regular deterministic or deterministic chaotic) or non-linear stochastic data [9]. If the estimated prediction error (E,,,> corresponds to a positive leading Lyapunov exponent, the possibility of regular determinism can be excluded. Consequently, non-linear dynamics such as chaotic or non-linear correlated stochastic processes are present. The small filled squares in Figs. 3-6 show the E, values of the phase-randomized time series in comparison with the original time series. A significant difference indicating particular non-linearities could be found only in the chaotic signal. This is due to the linear properties of the simulated regular deterministic signals and the simulated correlated stochastic (fractal, superposition of scaled sinus waves) signals. The results of the tests for non-linearities of all simulated data are summarized in Table 1.

Table 1 Mean values of 1Lorlninof - Em.surrol/(Em.oriPinar + J%.~~~,~), short-time prediction (At = 1). long-time prediction (AI = 16). data length N = 8192 Reg. deterministic Determ. chaos Uncorr. stochast. Corr. stochast.(0 16 1 16 At I 16 1 16 1 Non-linearities Orientation
l

0.14 < 0.01

0.14 < 0.01

< 0.01 0.94 *

< 0.01 0.58


l

< 0.01
>0.99 *

< 0.01
> 0.99 *

0.68 *
0.96 l

0.45 * 0.07

P < 0.05 Paired Wilcoxon rank test [26] (n = 6 simulations).

438

D. Hoyer

et al./ Cardiovascular

Research

31 (19%)

434-440

2:
0. 7 mm

. l m..

IO

7 3 5

9 11 13 15
At

A t (logarithmic scale) Fig. 6. E, versus At, plot as in Fig. 3. Simulated correlated stochastic signal (fractal, superposition of l/f scaled sinus wave with random phases), D, = 8.

Fig. 8. I& versus Ar, plot as in Fig. 3. Measured heart rate sequence during general anaesthesia; example; D, = 8.

2.6. Test for orientation space data)

frandomly shufling

the phase

In the original data, the orientation of the points ( P) in the phase space is organized by the running time parameter (t). To construct another kind of surrogate data set this parameter t is randomly shuffled. The resulting surrogate data sets consist of the identical points in the phase space but in randomized sequence. Therefore the statistical distribution of the data is not altered. The null hypothesis is now that there is no orientation in the original data. Orientation in the phase space data was found in all data sets with the exception of the uncorrelated stochastic data. The typical prediction characteristics of the phaserandomized and the randomly shuffled data in comparison with the original data are shown in Figs. 2-5. The results of the tests for orientation of all simulated data are summarized in Table I.

3. Results

We analysed the heart rate fluctuations of the states Sl-S3 according to the analysis of the simulated time series. Particularly in the analysis of measured data the choice of appropriate embedding parameters is important in order to avoid misleading results. We found that D < D, < (2 D + 1) is advantageous if the data length available is limited. Lower embedding dimensions are not able to consider the essential process characteristics and higher embedding dimensions can lead to statistical problems because of lack of data. With regard to the estimated heart rate dimensions (Sl, D = 7.15 f 0.8; S2, D = 2.4 f 0.7;

S3, D = 1.4 f 0.5 [7,26]), the data length of N = 8192 heart beats was necessary to obtain reproducible investigation results (N > 1O(o/2), [22]). In Figs. 7-9 the dependencies of E, on At for original, phase-randomized, and randomly shuffled data are shown by typical examples of Sl-S3. The significant differences between the original and the surrogate data (Table 2) gave evidence of non-linearities and orientation with regard to the 6 investigated rabbits in all states. We found a continuous increase of E,,, versus At during the short time prediction (At: 1 4 2) in all states. The mean slopes E,,, versus At were largest in Sl (2.6 f 0.4/heart beat), reduced in S2 (1.5 + 0;3/heart beat) and once more reduced in S3 (0.82 + 0.3/heart beat). Also with regard to the long-time prediction (At: 15 + 16) a transition to a lower increase in E,,, was found in all states. There, the mean values found of the slopes E,,, versus At were 0.26 f 0.06 /heart beat in S 1, 0.1 rt O.O2/heart beat in S2, and 0.01 f O.OOl/heart beat in S3 (Figs. 7-9). During Sl the highest values for non-linearities and orientation (Table 2) as well as the E, versus At characteristics (Fig. 7) of heart rate characteristics were mostly similar to that of the chaotic signal. A linear dependency of E,,, versus At can be approximated for 3-4 heart beats. This corresponds to the prediction characteristics of ideal chaotic signals [23]. During S2 the characteristics of E,,, versus At show a similar relation to that in the conscious state. But the predictibility was significantly increased with regard to the short-time and the long-time prediction (Fig. 8 vs. Fig. 7). The number of non-linearities was significantly decreased. The measure of orientation was not changed during the short-time prediction (Table 2). Corresponding to Sl, we

9 11 13 15 At Fig. 7. E, versus At, plot as in Fig. 3. Measured heart rate sequence of conscious rabbits; example; D, = 16.
7

1 3

9 17 13 75 At Fig. 9. Em versus At, plot as in Fig. 3. Measured heart rate sequence during vagal blockade and anaesthesia; example; D, = 8.
5 7

7 3

D. Hoyer

et al./ Cardiovascular

Research

31 (1996)

434-440

439

Table 2 Mean values of I En.origtnat - Enr.surrol/(Enr,origi~t + &.,,A Shorttime prediction (At = 11, long time prediction (At = 161, data length N=8192 At Conscious Anaesthesia Vagal blockade 1 16 1 16 1 16 Non-linearities Orientation 0.9 * 0.81 * 0.25 * 0.29
l

0.25 0.81 *
l

0.06 * 0.16
l

0.07 0.69
*

0.05

0.22

* P < 0.05 paired Wilcoxon rank test (261 (n = 6 rabbits).

would emphasize the possibility of a deterministic chaotic underlying process that generates the heart rate fluctuations in S2. During S3 the number of non-linear characteristics was once more decreased (Table 2) but was nevertheless significant. The decreased amount of orientation in the short-time prediction and the constant Em in the long-time prediction (as in the uncorrelated random signal) indicates a particular stochastic part in the heart rate fluctuations. It could be concluded that the non-linearities and orientation found in S3 may be caused more probably by a regular deterministic, noise-corrupted process than by a low-dimensional chaotic process.

4. Discussion

The starting point of the present study was that non-integer correlation dimensions and positive leading Lyapunov exponents of heart rate fluctuations were found which were significantly decreased during different pharmacological disturbances. In order to characterize the underlying processes in more detail we investigated the orientation and predictability of their trajectories in the multidimensional phase space. The typical behaviour of this dependency of uncorrelated stochastic (random), correlated stochastic (using the example of a fractal time series), regular deterministic (sinus wave), and deterministic chaotic (Henon attractor) signals was shown using simulated signals. The mean prediction errors are quantitative parameters of the time behaviour of infinitesimal neighbouring points in the phase space. The investigation of surrogate data which indicated non-linearities and orientation in the multidimensional phase space enabled a furthermore specific description. The quantitative heart rate characteristics found and their relations to typical simulated signals allowed some essential conclusions about the underlying processes and their pharmacological disturbance. Essential non-linear characteristics and their changes due to anaesthesia and vagal blockade could be described quantitatively. Common consideration of orientation in the phase space, non-linearities and prediction characteristics gives essential indications of participating chaotic processes. The approach presented here for analysis of orientation in the phase space is partly

related to the test for determinism proposed by Kaplan and Glass [15,16]. The assumption of chaos could furthermore be supported by Glass and Malta [25]. They demonstrated a deterministic process which corresponds to the heart rate fluctuations using multi-looped negative feedback. Determination of the underlying dynamic process of a measured time series without any other information is methodically limited. For instance, a deterministic dynamic system can be approximated by a stochastic Markov chain. Furthermore, Barlett [25] showed a special autoregressive model that produces a stochastic data set as time increases but which, when reversed in time becomes deterministically chaotic. In such cases, it would seem to be a meaningless task to distinguish chaos from correlated noise. Corresponding differences between the underlying physiological processes cannot be distinguished by the resulting measured time series. Furthermore, it should be taken into consideration that the physiological process which generates the heart rate fluctuations cannot be completely separated from any other undefined influences. It is also obvious that the physiological system does not match just one mathematical model exclusively. Nevertheless, the quantitative measures investigated of the multi-dimensional prediction characteristics in connection with estimation of the amount of non-linearities and orientation gave some impressive information about the complex organization of heart rate fluctuations. It could be shown that there are well-organized non-linear processes. Pharmacologically induced changes in these processes could be characterized by means of appropriate non-linear parameters.
Acknowledgements

This work was supported by grants from the Federal Ministry for Research and Technology (BMFT 01 202 F8) and the Deutsche Forschungsgemeinschaft (Ho 1634/l- 1).
References
[ll Mandell AJ. Dynamical complexity and pathological order in the cardiac monitoring problem. Physica D 1987:27;235-242. 121 Ganz RE, Weibels G, Sticker K-H. Ihe Lyapunov exponent of heart rate dynamics as a sensitive marker of central autonomic organization: an exemplary study of early multiple sclerosis. Int J Neurosci 1993:71;29-36. [3] Hoyer D, Schmidt K, Lilthke B, Zwiener U. Correlation dimension and Lyapunov exponent, essential characteristics of chaos in heart rate variability and respiration in different states of rabbits. Pfhlgers Arch 1994:426;Rl02. [4] Skinner JE, Carpeggiani C, Landisman DE, Fulton KW. Correlation dimension of heartbeat intervals is reduced in conscious pigs by myocardial &hernia. Circ. Res. 1991:68(4);966-976. [5] Zwiener U, Hoyer D, Llithke B, Schmidt K, Batter R. Relations between parameters of spectra1 power densities and deterministic chaos of heart rate variability. J Autonom Nerv Syst 1995:25;195198.

440

D. Hoyer

et al./

Cardiovascular

Research

31 (1996)

434-440

[6] Hoyer D, Schmidt K. Quantitication of chaos in heart rate variability and respiratory movements in different states by correlation dimension and Lyapunov exponent. In: Proceedings, IFAC Symposium, Modeling and Control in Biomedical Systems. Galveston, TX, 199419-20. [7] Hoyer D, Schmidt K, Zwiener U, Liithke B, Liebovitch LS. Estimation of stochastic and deterministic chaotic characteristics of heart rate variability. Proceedings, Int. Wiss. Koll., Ilmenau 1994;378383. [8] Osborne AR, Pr ovenzale A. Finite correlation dimension for stochastic systemswith power-law spectra. Physica D 1989:35;357-381. [9] Palus M. Chaotic time series analysis: identification and quantitication using information-theoretic functionals. Santa Fe Inst. Rep. 1993:93-03-017;1-30, [IO] Rapp PE, Albano AM, Schmab TI. Filtered noise can mimic low-dimensional chaotic attractors. Phys Rev 1993:E47(4);2289-2297. [ 111 Theiler J, Eubank S, Longtin A, Galdrikian B. Testing for nonlinearity in time series: the method of surrogate data. Physica D 1992:58;77-94. [12] Kennel MB, Isabel S. Method to distinguish possible chaos from colored noise and to determine embedding parameters. Phys Rev A 1992:46;3111-3118. [ 131 Sugihara G, May RM. Nonlinear forecasting as a way of distinguishing chaos from measurement error in time series. Nature 1990:344;734-741. [14] Tsonis AA, Elsner JB. Nonlinear prediction as a way of distinguishing chaos from random fractal sequences. Nature 1992:350;217.

[15] Kaplan DT. Direct test for determinism in a time series. Phys Rev Lett 1992:68;427-430. [16] Kaplan DT, Glass L. Coarse-granted embedding of time series: random walks, Gaussian random processes, and deterministic chaos. Physica D 1993:64;43 l-454. [17] Press WH, Teukolsky SA, Vetterling WT. Flamiery BP. Numerical Recipes in C. Cambridge University Press, 1992. [18] Takens F. Detecting strange attractors in turbolence. In: Rand D, Young LS, eds. Dynamical Systems in Turbulence, New York: Springer Verlag, 1980:366-381. [19] Sauer T., Yorke JA, Casdagli M. Embeddology. Santa Fe Inst Rep 1991:91-Ol-008;1-46. [20] Albano M, Muench J, Schwartz C, Mees AI, Rapp PE. Singular-value decomposition in the Grassberger-Procaccia algorithm. Phys Rev A 1988:38;3017-3026. [21] Wolf A, Swift JB, Swinney HL, Vastano JA. Determining Lyapunov exponents from a time series. Physica D 1985:16;285-317. [22] Eckmamt JP, Ruelle D. Fundamental limitations for estimating dimensions and Lyapunov exponents in dynamical systems.Physica D 1992:56;185-187. [23] Peitgen HO, Jiirgens H, Saupe D. Chaos and Fractals. New York: Springer-Verlag, 1992. [24] Glass L, Malta CP. Chaos in multi-loop negative feedback systems. J. Theor Biol 1990: 145;217-223. [25] Barlett MS. Chance or chaos. JR Stat 1992:A3;321-347. [26] Sachs, L. Angewandte Statistik. Berlin: Springer Verlag, 1992.

Você também pode gostar