Medical Hypotheses (1998) 51, 367-376© Harcourt Brace & Co. Ltd 1998Fractal organization of the pointwise correlationdimen...
368                                                                                              MEDICAL HYPOTHESESthe mat...
FRACTAL ORGANIZATION IN HEART RATE                                                                                36920 ra...
370                                                                                                  MEDICAL HYPOTHESESthe...
FRACTALORGANIZA~ONINHEARTRATE                                                                                             ...
372                                                                                                                       ...
FRACTAL ORGANIZATION IN HEART RATE                                                                                        ...
374                                                                                                        MEDICAL HYPOTHE...
FRACTAL ORGANIZATIONIN HEART RATE                                                                                         ...
376                                                                                                                 MEDICA...
Upcoming SlideShare
Loading in …5



Published on

Published in: Health & Medicine, Technology
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. Medical Hypotheses (1998) 51, 367-376© Harcourt Brace & Co. Ltd 1998Fractal organization of the pointwise correlationdimension of the heart rateE. NAHSHONI, E. ADLER*, S. LANIADO*, G. KEREN*Department E, The Gehah Psychiatric Hospital, Petah-Tiqva, and Sackler School of Medicine, Tel Aviv,Israeb *Department of Cardiology, Tel Aviv Medical Center, Tel Aviv, and Sackler School of Medicine,Tel Aviv, Israel. Correspondence to: E. Nahshoni, POB 102, 49100 Petah-Tiqva, Israel(Phone: +972 3 9258258; Fax: + 972 3 9241041)Abstract - - Objective: To depict and quantify the degree of organization of the heart ratevariability (HRV) in normal subjects. Methods: A modified algorithm was created to estimateseries of "point-dimensions" (PD2) from interbeat (R-R) interval series of 10 healthy subjects(21-56 years). Our innovation is twofold: (i) we quantified instances of low-dimensional chaos,random fluctuations, and those for which our method failed to provide either (due to poorstatistics); (ii) consecutive subepochs of PD2s underwent a relative dispersion (RD) analysis,yielding an index (D) which quantifies the dynamical organization of the heart rate generator. Results: The mean values of PD2 series varied between 4.58 and 5.88 (mean +_SD=5.21 +_0.41, n = 10). For group 1 (21-30 years, n = 6) we found an averaged PD2 of 5.49 _+0.27,while for group 2 (47-56 years, n = 4) PD2 averaged 4.79 +_.0.17. The RD analysis performedfor subepochs of PD2s yielded both instances obeying fractal scaling (D < 1.5) andstochasticity (D > 1.5). The average D for group 1 was 1.39 + 0.04 (14 subepochs) and forgroup 2, 1.20 _+0.008 (8 subepochs). Paired t-test and Hartley F-max test for comparisonbetween D values and homogeneity of variance between the two groups were performed,yielding P-values 0.004 and 0.02, respectively. Conclusions: The complexity of the HRV seems to be modulated by a non-random fractalmechanism of a hyperchaotic system, i.e. it can be hypothesized to contain more than oneattractor. Also, our results support the chaos hypothesis put forth recently, namely, thecomplexity of the cardiovascular dynamics is reduced with aging. The index of relativedispersion of the dimensional complexity has to be tested in various clinico-pathologicalsettings, in order to corroborate its value as a potential new physiological measure.Introduction frequency and phases of biological oscillators, or to the coupling of various regulatory feedback loops,Physiological systems have long been recognized to thus engaging nonlinear mechanisms for elucidationdisplay complex temporal fluctuations, even during of the dynamics. Although, as in the physical sciences,steady state conditions. Attempts were made to attri- solutions have resulted in linearizations, only duringbute them to random influences, which perturb the the last decade has a natural link been drawn betweenReceived 28 April 1997Accepted 12 June 1997 367
  2. 2. 368 MEDICAL HYPOTHESESthe mathematico-physical field of nonlinear dynamics theses, which motivated ongoing research effortsand physiology. This has triggered an ongoing trend meant to quantify the dynamical characteristics ofof paradigm shift in the medical sciences and in the heart rate dynamics under the assumption thatbiological thinking in general. it evolves on a low-dimensional strange attractor. Since the advent of digital processing, the heart rate These attempts were based mainly on dimensionalbecame the most accessible and reliable signal for analysis, which resulted in correlation dimensionsanalysis among cardiovascular variables. The heart (interpreted as a static measure of the number ofrate variability (HRV) is traditionally assessed using independent variables necessary to specify the state offrequency (spectral analysis) and time (standard de- the system under study), ranging between 3.6 and 5.2viations, interval occurrence histograms, etc.) domain in normal subjects (17). This was supported later, bytechniques. Using such techniques a complex coupling introducing another measure of deterministic chaos,with respiration, baroreceptors, the nervous system, i.e. the largest Lyapunov exponent which yielded abody temperature, metabolic rate, hormones, sleep finite positive value, thus demonstrating the propertycycles, etc. was revealed. For example, spectral analysis, of sensitivity to initial conditions, which is the hall-which exposed activity bands in the frequency mark of chaotic behavior (18). But later estimatesdomain comprising thermoregulation (~0.05Hz), of the correlation dimension were found to be muchbaroreflex control of peripheral resistance (~ 0.1 Hz) higher (-8.5) than previously reported, thus pre-and respiratory control (~ 0.2 Hz), was combined cluding firm conclusions as to the true nature of thewith pharmacological blockade to attribute the lower- heart rate generator (19).band fluctuations (0.04--0.15 Hz) to the joint influence Recently, other modified measures of dynamicalof the sympathetic and parasympathetic arms of the complexity, mostly suited for nonstationary, noisy,autonomic nervous system, while the higher fre- and limited record length signals, have been intro-quency band (0.15-0.4 Hz) was shown to be purely duced. Among them is the estimation of the pointwiseparasympathetically mediated. The spectral signature correlation dimension (PD2), which provides moreof HRV was also related to various physiological information about the temporo-spatial evolution ofand pathophysiological settings, such as standing, the dominant complexity of the heartbeat (20,21).hemorrhage and hypotension, which enhance the low This technique was applied to a very limited numberfrequency fluctuations, while exercise and standing of subjects, from which no firm conclusions coulddecrease the respiratory fluctuations. From the clinico- be drawn, except for one clinical study which corre-pathological viewpoint, patients with heart failure lated a reduced dynamical complexity hours beforehave diminished power spectrum at frequencies above the occurrence of lethal arrhythmias in high-risk~ 0.02 Hz (1-7). The other arm of traditional analysis, patients (22).the time domain analysis, has related decreased HRV In the light of the open questions and computa-in diabetes mellitus, ischemic heart disease, conges- tional restrictions in this growing field of research,tive heart failure (8), and also associated an increased we addressed the issue of the irregular nature of themortality in patients after acute myocardial infarction HRV in 10 healthy subjects. We computed correlation(9). dimensions and the series of pointwise dimensions. Taken together, these techniques have several short- We also introduced a modified version of the point-comings. For example, spectral analysis is a method wise dimension algorithm, which, we believe, canmostly suited for linear systems, while physiological depict both instances of low-dimensional chaos andsystems are inherently nonlinear. Also, time domain stochasticity. The complex relation between themanalysis, which is basically an averaging technique, was investigated using fractal techniques. The physio-overlooks the dynamical nature. Thus, it appears that logical and clinical importance of the measure wethese techniques are often insufficient to characterize introduced is still unknown.the complex behavior of the heart rate generator. Since the last decade, nonlinear methods of analysis,based on the paradigm of deterministic chaos (10), Methodshave permeated the realm of biomedical signalanalysis (11). This was motivated by the observation Subjectsof an inverse power-law scaling (also called 1/f Ten volunteers, aged 21-56 (6 males, 4 females)spectrum), which some chaotic systems may display, without symptoms or history of heart disease andand by its association to the fractal concept (mani- under no medication, were recruited for the study.fested by self-similarity over multiple orders of Their surface ECG, which showed no signs of patho-temporal magnitude) (12-16). In the case of heart rate logy, was recorded at rest in a supine position, duringdynamics, these observations heralded new hypo- quite spontaneous breathing (~ 15 breaths/min) for
  3. 3. FRACTAL ORGANIZATION IN HEART RATE 36920 rain. All recordings were done between 10 and series were time-delayed for successive embedding12 a.m., and each subject was allowed to adjust dimensions (from m = 1 to m = 16). Within a givencomfortably for 10 min in a supine position before the embedding dimension, the distance (r) of each pointdata were collected. They all gave informed consent to every other point was calculated. Their absoluteto the protocol. values were rank-ordered from the smallest to the largest, and the range from the smallest to the largestData aquisition value was broken up into discrete intervals. Then, the number of times a distance fell within an interval wasThe ECG signals were continuously recorded using a counted. A cumulative histogram was then formedlaptop-based HIPEC ANALIZER HA-200/AH system by summing the number of instances for which a(Aerotel - - computerized systems, Israel) with a distance was less than or equal to the upper boundarysampling rate of 1000 Hz, and 16 bit signal resolu- of the interval. This is the correlation integral C(r).tion. The ECG records were transferred to a personal C(r) was then plotted as a function of r on a log-logcomputer for off-line analysis which started with a representation, resulting in a sigmoid-shaped curvequality control procedure: visual inspection, baseline (in this case implying chaotic dynamics). The slopeshift evaluation and a moving average (four points over the largest linear range (if there is one) wasaveraging in succession along the record) for signal to measured, using linear regression (with a regressionnoise ratio improvement. Then the interbeat intervals coefficient R 2 > 0.98). In this scaling range the local(R-R) were computed using an algorithm developed exponent is constant and ~ d InC(r)/d In(r). Then, thein our laboratory, with which an R wave threshold embedding dimension was advanced and its corre-detection was combined with first derivative and QRS sponding slope was calculated. These slopes werewidth considerations, for an accurate R wave detection. then plotted versus the embedding dimension, looking for a saturation region, i.e. a region in which theAttractor reconstruction slopes no longer grow. This plateau region was thenUsually the experimentalist is confronted with in- considered as the correlation dimension (D2), and itsability to gain access to m simultaneous recordings value was calculated with a weighted average tech-necessary to describe the systems trajectory in nique (each value in the plateau region was weightedm-dimensional phase space. Thus, only one scalar by the variance of its underlying slope calculation).observable can be monitored as a function of time. This process was also performed for randomizedFortunately, it has been shown that certain properties versions of the R - R series (with similar mean andof the dynamics are feasible through the method of variance) in order to provide confidence limits for ourtime delays using Takens theorem, as follows (23). calculations.Consider a single time series regularly spaced intime: xi = x%), i = 1..... N. Then a time lag "~is intro- Pointwise dimension of R-R intervalsduced, such that m-dimensional vectors are created:xi= [x(ti), x(ti + "c)..... x(ti+ (m-1)x)]. This process is The point estimate of the correlation dimensiontermed embedding, and m is called the embedding (PD2) begins with the time lag calculation, followeddimension. Through this reconstruction a phase by the embedding procedure, as described is spanned and dynamical and metric measures Then, starting with the initial point in the series, its(Lyapunov exponents, dimensions) may be accessible. local correlation integral is calculated, i.e. the dis- tances are taken with respect to this point and ranked- ordered as usual, for each embedding dimensionCorrelation dimension of R-R intervals (m = 1..... 16). The slopes for each m were evaluatedThe correlation dimension was calculated using the using a linear regression (R2> 0.98), and a slopemethod of Grassberger and Procaccia (G-P) as follows values, corresponding to m = 8 ..... 16 were stored in(24). First, for each R - R interval series, the normal- a file. The algorithm steps to the next point in theized autocorrelation function given by: series, and the whole procedure is repeated until the whole file is exhausted. Then comes the procedure~g(~) = {(I/N) £[R-R)i - < (R-R) >][(R-R)i+x - that we call slope convergence, which calculates the < (R-R) >] }/{(I/N) Z[R-R)i - < (R-R) >]2} slope of the 9 slopes versus the embedding dimensionswhere (m = 8 ..... 16) using linear regression. Our innova- tion was to subject this to the imposition of three< (R-R) > = (l/N) Z (R-R)/ conditions as follows: Ill if the slope was less thanwas constructed, and its first zero crossing was calcu- 0.5 and larger than -0.23, we considered this as goodlated to provide the time lag (x) in beats. Then, the convergence and the PD2 could be estimated using
  4. 4. 370 MEDICAL HYPOTHESESthe weighted average technique as described before; dimensions was needed. Homogeneity of variance(ii) if the slope was equal or larger than 0.5, we was tested by the Hartley F-max test. Statisticalconsidered it as if no saturation existed, and at significance was assumed if the null hypothesis couldthis point (or time), the system probably manifested be rejected at the 0.05 probability level.a random fluctuation. In order to incorporate sucha behavior into the sequence of PD2s, we decided,quite arbitrarily, to take the average of the two highest Resultsslope estimates, as the point-dimension, when sucha condition appears; (iii) if the slope was equal to Thc correlation dimension (D2) of R-R intervalsor less than --0.23, we considered it as if no slope varied from 3.29 to 5.16, with an overall mean ofconvergence existed, and the dimensional estimate 4.01 ± 0.54 (Table 1). Fig. la illustrates one of theat this point was excluded, possibly because of poor series of R-R intervals. This corresponding normal-statistics. ized autocorrelation function is shown in Fig. lb. The The results of the PD2 series were assigned first zero crossing (x), in this case was equal to 6according to the three conditions mentioned above. beats. The correlation integral (C(r)) for embeddingThis provided us with the ability to discriminate the dimensions (m = 2,4,6,9,12,16) is shown in Fig. 2a,points which manifested low-dimensional chaos and while the calculated slopes in the linear regions ofrandom behavior, from those for which a dimensional the log-C(r) representation, versus the embeddingestimate could not be achieved. From the above dimension, is shown in Fig. 2b. Note the convergenceoutput files we extracted sequences of dimensional towards a dimensional value of 4. Randomized ver-subepochs, which were then suited for the relative sions of the R-R intervals have demonstrated, asdispersion analysis. expected, non-convergence (Fig. 2c). A sequence of pointwise dimensions (PD2s) versus the reference point is shown in Fig. 3a. Note threeDispersion analysis regions in the dimensional complexity plot, i.e. high values (PD2 > 6), low-dimensional region (3 > PD2 < 6)There are three basic methods of dispersion analysis and zero-valued reference points, corresponding tothat can be applied to temporal observations (25). One non-convergence due to poor statistics. This can beof them, adopted in our study for each sequence of seen from the histogram (Fig. 3b) showing the distrib-calculated pointwise dimensions, is called relative ution of the rounded dimensional values, includingdispersion (RD) analysis. Our intention was to try and the points corresponding to stochasticity and to non-see if the temporal evolution of PD2 series obeys any calculability at both ends of the figure. For the subjectscaling properties. Thus for each subject, this simple shown in the figure the average PD2 was 5.37 ± 0.93.algorithm goes as follows: first, the mean, standard In Fig. 4, four subepochs, each comprising ~150 PD2deviation (SD), and RD% (= 100 x SD/mean) of the values (corresponding to an average of about 2.5original PD2 series were calculated. Then, pairs of minutes-record-length each) are shown. In Fig. 5 theadjacent PD2s were averaged and their RD% values logarithmic plot of the RD(%) versus the intervalwere calculated, thus doubling the interval length.Recursive pairing with doubling of each previousinterval length was carried out while its correspond- Table 1 Correlation dimension (D2) of 10 healthying RD% was calculated. This was done until the subjects at restwhole record was exhausted. By plotting the RD%against the interval length on a logarithmic scale, the Gender Age (years) HR ± SD Correlation dimensionslope was estimated using a least-squares linear fit. (beat/rain) (D2 ± ZkD2)The fractal dimension (D) could thus be extractedfrom the slope (slope --- l-D). In order to confirm the M 21 69.1 ± 5.6 4.51 ± 0.13temporal organization of the PD2 series, randomized F 25 65.5 ± 3.0 3.58± 0.07versions based on similar statistical characteristics M 26 65.2 ± 2.5 3.93 ± 0.09 F 28 67.6± 3.5 5.16 ± 0.02(number of points, mean and standard deviation) were M 30 57.6 ± 4.6 3.99 ± 0.16generated, and their RD analysis was also performed. M 30 54.2±2.3 4.50±0.19 M 47 71.4± 3.9 3.87 ± 0.21 M 56 60.9 ± 3.0 3.29 ± 0.19 F 56 64.7 ± 2.3 3.48 ± 0.21Statistical analysis F 56 66.8 ± 2.6 3.77 ± 0.03All data are expressed as mean ± SD. A paired t-test mean ± SD 37.5 ± 13.7 64.3 ± 5.0 4.01 ± 0.54was performed when comparison between fractal
  5. 5. FRACTALORGANIZA~ONINHEARTRATE 371 1.25 0.75 e- •- 0.5- I e¢- 0.25 0 1 500 1000 Beat number 0.7 0.6 0.5 0.4-- 0.3-- 0.2 0.1- 0 -0.1 .v , , .... IW ,v, V V_ -0.2 m 100 200 300 b T Fig. 1 (a) R - R intervals for one of the subjects (1296 intervals, 20 min). (b) The normalized autocorrelation function of R - R intervals shown in (a). The first zero crossing was found to be 6 beats.length (measured in beat number) for one of the 850. The overall mean values of the PD2 series variedsubepochs is shown. Its slope provides the fractal between 4.5 and 5.88 (mean = 5.21 ± 0.41, n = 10),dimension of the dimensional complexity at a parti- but the mean PD2s of the various subepochs werecular subepoch. smaller than the overall average, at least during one Table 2 summarizes the results of the fractal subepoch for each subject. We divided the subjectsdimensions (D) of the subepochs of series of PD2s. into two groups according to their age. For group 1The shortest subepoch consisted of 80 consecutive (21-30 years) the average PD2 varied between 5.19dimensional values, while the longest consisted of and 5.88 (mean = 5.49 ± 0.27, n = 6), while the
  6. 6. 372 MEDICAL HYPOTHESES 0 -1, -2. -3 to t- -4 -5 -6 -7 15 2~0 . . . . 2~5 . . . . Inr Emb 9 --K--- Emb 12 ~ Emb 16 5 8C C0 Or r 4 r? ?6 i 3 . i0 On n 4di 2 ?m me e 2n 1 [I0 On 0 n . . . . . . . . . . . . . . 0 . . . . . . . . 110 111 I i I4 I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 2 3 4 5 6 7 8 9 1 13 1 1 16 b Embedding dimension c Embedding dimensionFig. 2 The correlation integral C(r) versus r on a logarithmicplot. Seen are embeddingdimensions m = 2, 4, 6, 9, 12, 16. (b) The slopeof the scaling region as a function of the embeddingdimension (m), for a healthy subject (26 years). Note the saturation towards acorrelation dimension (D2) estimate of ~4. (c) For a randomlygenerated version of R-R intervals, the slope estimates of In C(r) versuslnr, as a functionof the embeddingdimension, do not saturate.relative dispersion analysis of their consecutive PD2 Discussionseries yielded both instances of fractal scaling(D < 1.5) and stochasticity (D > 1.5). The averaged The concept of fractals, first coined by B. Mandelbrotfractal dimension for this group was 1.39 4-0.04 (26), and its association with chaos theory, heralded(14 subepochs). In group 2 (47-56 years), the PD2 novel insights into the realm of structural andmean values ranged between 4.58 and 5.03 (mean = dynamical variability in the medical sciences (27)4.79 ± 0.17, n = 4). Note that in group 2 the fractal and biology in general (11). During the last decade,estimates ranged between 1.09 and 1.33 (mean = the intimate connection between deterministic chaos1.20 ± 0.008, 8 subepochs), i.e. never exceeded 1.5. and fractal geometry has stimulated ongoing research The t-test and the F-test showed statistical signi- efforts to quantify the dynamical aspects of the heartficance when the means and variances of the fractal rate generator. Babloyantz and Destexhe were the firstdimensions of the PD2 subepochs were compared to quantify its dynamical measures using chaos theory(P values: 0.004 and 0.02, respectively). Note that the techniques (17). Their results (correlation dimensions,overall results of the RD analysis were indicative Kolmogorov entropies and the largest Lyapunovof fractal scaling (D < 1.5) in about 80% of all exponent), were supportive to the contention that thesubepochs tested. heart rate generator evolves on a low-dimensional
  7. 7. FRACTAL ORGANIZATION IN HEART RATE 373 10 I N of points round(D2)- alp-)0.50 0 N o! p- 4z m 0 | round(O2)- I N of p- 0 i I round(D2}- round(D2)- 2 H o f p- 3 H of rouncl(D2)m 4 N o f roun4(D2)- S N of roarl4(D2)- 6 N o f p- pm 246 p- Se2 pm 146 0 "k roun4(D2)- 7 X o f p - 64 round(D2)- 8 g of p,, ]. round(D2)- 9 N o f p- 0 round(D2)-20 N of p- 0 i roun4(D2)-22 N of p- 0 round(D2)-12 N of p- 0 round(D2)-13 N of p- 0 round(D2)-14 N of p- 0 round(D2)-25 N of p- 0 round(D2)-26 H of p" 0 round(D2)-17 N of p,, 0 round(D2)-28 N of p- 0 round(D2)-lg N Of p- 0 round(D2)-20 N of p- @ H of exclu sip<-0.23 2v • a 0 250 500 750 1000 1250Fig. 3 (a) Serial pointwise dimensions (PD2s) as a function of the beat number (nref) for one of the subjects. The zero-valued PD2s areonly sign of the instances for which a dimensional estimate could not be derived. (b) A histogram showing the distribution of the beatnumber as a function of the rounded dimensional estimates. At the two extremes of the diagram we note the number of points for which arandom fluctuation is supposed to take place (slope > 0.5), and on the other side the number of point for which an estimate could not befound (slope < -0.23).chaotic attractor. Later, other groups provided sup- periodic behavior was seen in normal subjects, withportive evidence to this hypothesis (18,28), although an increase in complexity during sleep (21). Recentrecently Kanters et al found weak evidence in favor studies found the dimensional complexity during(19). Thus, the existence of low dimensional chaos in experimental myocardial infarction in pigs to declinecardiac activity, at least at the whole organ level of significantly prior to the occurrence of ventricularactivity, is still an open question. Most of the estima- fibrillation (31). This has motivated Skinner et altions were based on dimensional estimations of the to evaluate lethal arryhythmias in various groups ofwidely used Grassbeger-Procaccia algorithm. Imple- high-risk patients. It was found (with high specificitymentation of this algorithm needs several precondi- and sensitivity) that the dimensional complexitytions to be observed (29,30): an adequate choice of is reduced hours before the occurrence of lethalembedding dimension, a suitable choice of the time arrhythmias (22).delay needed to span the attractor, low level of noise Our results support the contention of low-present in the system, stationarity, and the data set dimensional chaos, as proposed by others. The corre-should not be too short. Some of these requirements lation dimension was found to vary between 3.29are not attainable in biology and in physiology in and 5.16, with a mean of 4.01 + 0.54 (n = 10). Theparticular. Moreover, the G-P algorithm provides a average pointwise dimension ranged between 4.58dimensional estimate which averages out possible and 5.88, with a mean of 5.21 + 0.41 (n = 10). Further-relevant dynamical features. Recently, the intro- more, we noticed that our subjects could be dividedduction of the pointwise dimension algorithm, which into two groups according to age as follows: group 1provides series of point dimensions, has provided (21-30) years, n = 6) had a higher average, 5.49 + 0.27,some solutions to the limitations of the G-P algo- than group 2 (47-56 years, n = 4 ) at 4.79+ 0.17.rithm, namely, non-stationarity and record length. Our innovation in this study was twofold. First, weThis method was implemented for heart transplant included in the dimensional complexity algorithmrecipients and the dimensional complexity was found means to include both instances of low-dimensionalto oscillate almost periodically (20). Also, a roughly chaos and stochastic bursts in a sequential manner,
  8. 8. 374 MEDICAL HYPOTHESES 10 2O g tJL, JJl| RF Jd l I0 10 i1. D ill Ir ,I Fig. 4 Four subepochs of sequential PD2s from the series shown in Fig. 3a. Each subepoch comprises about 150 PD2s.i.e. as a function of the beat number. Second, this cant decline in the complexity of the cardiovascularenabled us to apply a fractal technique (relative system (blood pressure and heart rate). Such findingsdispersion analysis) to explore the different subepochs may reflect the breakdown and decoupling of inte-of dimension series for scale independence. We found grated physiologic regulatory systems with agingthat the older group manifested fractal scaling and may signal an impairment in the cardiovascular(D < 1.5) in all subepochs tested. As for the younger ability to adapt efficiently to internal and externalgroup, only in 64% of tested subepochs did we find perturbations. This is contradictory to the sacredfractal scaling (D < 1.5), while the rest was indicative principle of homeostasis, which was developed byof a random control (D > 1.5). Moreover, the differ- Walter Cannon, and postulates that with disease andences in the averages and variances of the fractal aging the body is less able to maintain a constantdimensions between the two groups were found to steady state, as a result of breakdown of its regulatorybe statistically significant. This is in contention with systems. Our findings support the chaos hypothesis ofresults from other chaos-derived techniques imple- a homeokinetic principle in physiology, namely,mented by Kaplan et al on old versus young subjects physiological systems in young healthy subjects tend(32). They found that the older group showed signifi- to fluctuate between a set of metastable states, thus
  9. 9. FRACTAL ORGANIZATIONIN HEART RATE 375 of the heart rate variability, which can be quantified 20 using fractal techniques, seems to contain mixing of both chaotic and random fluctuations. The nature 10 of such a behavior is not yet understood, but one 5 may hypothesize that an increase in the dimensional complexity (D2 can be thought of as a measure --,.. of independent variables necessary to describe the 2 system), may correspond to recruitment of several 1 subsystems influencing the heart rate generator, or to the activation of more independent control loops.0.50 A reduced complexity, on the other hand, may mani-0.20 fest deactivation of control loops, or maybe increased self-organization of some of these systems. Also, the abrupt changes in the dimensional complexity may0.10 represent shifts between different attractors of the system. Such hypotheses, may better be resolved by 140 74 37 19 10 5 3 comparing the fractal dimensions of the dimensionalFig. 5 Plot of RD% (relative dispersion) versus interval length on complexity (and other measures of nonlinear tech-a logarithmic scale. The fractal dimension (D) is derived from the niques) under different physiological and clinico-slope (slope = l-D). pathological settings. Currently, we are in the process of obtaining longer data records from heart transplant recipients, in ordermaking the system more adaptable to its internal and to gain more insight regarding the value of the fractalexternal surroundings (15,33). estimate of the dimensional complexity of the heart We thus propose that the dimensional complexity rate generator, as a potential new dynamical measure.T a b l e 2 Subepochs o f pointwise dimension (PD2) series, averaged P D 2 s for each subepoch, fractal dimension (D) foreach subepoch, and averaged P D 2 s for w h o l e recordsGender Age Nref PD2 ± SD average over subepochs D(RD) ± SD PD2 ± SD average over total recordM 21 1-150 5.28 ± 0.88 1.42 + 0.07 5.37 ± 0.93 300--450 5.09 ± 0.09 1.27 + 0.03 570-720 5.61 ± 0.73 1.49 ± 0.06 850-1000 4.35 ± 0.69 1.09 ± 0.04F 25 1-90 4.45 ± 1.05 1.21 ± 0.09 5.19 ± 1.04 200-350 4.86 + 0.75 1.54 ± 0.06M 26 130-196 6.01 ± 0.97 1.72 ± 0.11 5.83 ± 1.02 250-850 5.62 ± 0.97 1.19 ± 0.10F 28 1-300 5.90 ± 0.90 L28 + 0.09 5.34 ± 0.72 301--600 5.03 ± 0.35 1.36 ± 0.12M 30 1-200 4.91 ± 0.69 1.55 ± 0.09 5.30 ± 0.11 600--995 5.09 ± 1.10 1.13 ± 0.06M 30 30-110 5.92 ± 0.83 1.57 ± 0.09 5.88 ± 1.08 125-295 5.34 ± 0.82 1.63 ± 0.09M 47 1-350 3.92 ± 0.51 1.11 ± 0.04 4.69 ± 0,77 400-1250 4.99 ± 0,63 1.21 ± 0.06M 56 1-300 4.95 ± 0,68 1.18 ± 0.07 5.03 ± 0,71 500-1000 5.11 ± 0.69 1.32 ± 0.05F 56 100-300 4.63 ± 0.72 1.24 ± 0.07 4.87 ± 0.96 500-750 4.32 ± 1.21 1.09 ± 0.05F 56 1-80 4.39 ± 0.69 1.33 ± 0.08 4.58 ± 0.64 200-380 4.63 ± 0.79 1.14 ± 0.08mean ± SD 5.02 ± 0.54 1.32 ± 0.18 5.21 ± 0.41M, male; F, female; Nref, sequences of consecutive data points subepochs; PD2 + SD, averaged pointwise dimension + standard deviation;D(RD), fractal dimension of each subepoch, derived from relative dispersion analysis.
  10. 10. 376 MEDICAL HYPOTHESESReferences 15. Lipsitz L A. Age-related changes in the complexity of cardiovascular dynamics: a potential marker of vulnerability to 1. Appel M L, Berger R D, Saul J P, Smith J M, Cohen R J. Beat disease. Chaos 1995; 5(1): 102-109. to beat variability in cardiovascular variables: noise or music? 16. Kobayashi M, Musha T. 1/f fluctuation of heartbeat period. J Am Coil Cardiol 1989; 14:1139-1148. IEEE Trans Biomed Eng 1982; 29: 456-457. 2. Malliani A, Pagani M, Lombardi F, Cerutti S. Cardiovascular 17. Babloyantz A, Destexhe A. Is the normal heart a periodic neural regulation explored in the frequency domain. Circula- oscillator. Biol Cybernetics 1988; 58:203-211. tion 1992; 84: 482-492. 18. Rigney D R, Mielus J E, Goldberger A L. Is normal sinus 3. Malliani A, Pagani M, Lombardi F. Physiology and clinical rhythm chaotic? measurement of Lyapunov exponents. implications of variability of cardiovascular parameters with Circulation 1990; 82(4) Suppl. III, abstr. focus on heart rate and blood pressure. Am J Cardiol 1994; 19. Kanters J K, Holstein-Rathlou N-H, Agner E. Lack of 73: 3C-9C. evidence for low-dimensional chaos in heart rate variability. 4. Akselrod S, Gordon D, Ubel F A, Shannon D C, Barger A C, J Cardiovasc Electrophysiol 1994; 5: 591-601. Cohen R J. Power spectrum analysis of heart rate fluctuations: 20. Zbilut J P, Mayer-Kress G, Giest K. Dimensional analysis of a quantitative probe of beat-to-beat cardiovascular control. heart rate transplant recipients. Math Biosci 1988; 90: 40-70. Science 1981; 213: 220-222. 21. Mayer-Kress G, Yates F E, Benton L et al. Dimensional 5. Saul J P, Arai Y, Breger R D, Lilly L S, Colucci W S, Cohen analysis of nonlinear oscillations in brain, heart, and muscle. R J. Assessment of autonomic regulation in chronic congestive Math Biosci 1988; 90: 155-182. heart failure by heart rate spectral analysis. Am J Cardiol 1988; 22. Skinner J E, Pratt C M, Vybrial T. A reduction in the correla- 61: 1292-1299. tion dimension heartbeat intervals precedes imminent ventricular 6. Pomeranz B, Macaulay R J B, Caudill M A e t al. Assessment fibrillation in human subjects. Am Heart J 1993; 125: 731-743. of autonomic function in humans by heart rate spectral 23. Rand D, Young L-S, eds. Detecting strange attractors in fluid analysis. Am J Physiol 1985; 248 (Heart Circ Physiol 17): turbulence. In: Dynamical Systems and Turbulence. Berlin: H151-H153. Springer-Verlag, 1981. 7. Madwed J B, Snads K E F, Saul J P, Cojen R J. Spectral 24. Grassberger P, Procaccia I. Measuring the strangeness of analysis of beat-to-beat variability in HR and ABP during strange attractors. Physica 1983; 9D: 183-208. hemorrhage and aortic constriction. In: Lown B, Malliani A, 25. Glenny R W, Robertson H T, Yamashiro S, Bassinghwaighte. Prodocimi M, eds. Neural Mechanisms and Cardiovascular Applications of fractal analysis to physiology. J Appl Physiol Disease. Fidia Research Series. Padova: Liviana Press, 1986, 1991; 70(6): 2351-2367. Vol. 5: 291-301. 26. Mandelbrot B B. Thc Fractal Geometry of Nature. San 8. Takase B, Kurita A, Noritake Met al. Heart rate variability in Francisco: W H Freeman, 1982. patients with diabetes mellitus, ischemic heart disease, and 27. West B J. Fractal Physiology and Chaos in Medicine. New congestive heart failure. J Electrocardiol 1992; 25: 79-88. Jersey: World Scientific, 1990. 9. Kleiger R E, Miller J P, Bigger J T, Moss A J, and the Multi- 28. Ravelli F, Andolini R. Complex dynamics underlying the center Post-infarction Research Group. Decreased heart rate human electrocardiogram. Biol Cyb 1992: 67; 57. variability and its association with increased mortality after 29. Kantz H, Schreiber T. Dimension estimates and physiological acute myocardial infarction. Am J Cardiol 1987; 59: 256-262. data. Chaos 1995; 5(1): 143-154.10. Ott E. Chaos in Dynamical Systems. Cambridge: Cambridge 30. Schreiber T, Kantz H. Noise in chaotic data: diagnosis and University Press, 1993. treatment. Chaos 1995; 5(1): 133-142.11. Olsen L F, Degn H. Chaos in biological systems. Quarterly 31. Skinner J E, Carpeggiani C, Landisman C E, Fulton K W. Rev Biophys 1985; 10(2): 165-221. Correlation dimension of hcartbeat intervals is reduced in12. Goldberger A L, Rigney D R, West B J. Chaos and fractals conscious pigs by myocardial ischemia. Circulation Res 1991; in human physiology. Sci Am 1990; 262: 42--49. 68: 966-976.13. Goldberger A L, West B J. Fractals in physiology and 32. Kaplan D T, Furman M 1, Pincus S M, Ryan S M, Lipsitz L A. medicine. Yale J Biol Med 1987; 60: 421-435. Aging and the complexity of cardiovascular dynamics. Bioph14. Goldberger A L. Nonlinear dynamics, fractals and chaos: J 1991; 59: 945-949. applications to cardiac electrophysiology. Ann Biomed Eng 33. Lipsitz L A, Goldberger A L. Loss of complexity and aging. 1990; 18: 195-198. J Am Med Assoc 1992; 267: 1806-1809.