Characterization of chaotic dynamics in the human menstrual cycle
© Derry and Derry; licensee BioMed Central Ltd. 2010
Received: 11 June 2010
Accepted: 5 October 2010
Published: 5 October 2010
The human menstrual cycle is known to exhibit a significant amount of unexplained variability. This variation is typically dismissed as random fluctuations in an otherwise periodic and predictable system. Given the many delayed nonlinear feedbacks in the multiple levels of the reproductive endocrine system, however, the menstrual cycle can properly be construed as the output of a nonlinear dynamical system, and such a system has the possibility of being in a chaotic trajectory. We hypothesize that this is in fact the case and that it accounts for the observed variability.
Here, we test this hypothesis by performing time series analyses on data for 7749 menstrual cycles from 40 women in the 20-40 year age range, using the database maintained by the Tremin Research Program on Women's Health. Both raw menstrual cycle length data and a formal time series constructed from this data are utilized in these analyses. Employing phase space reconstruction techniques with a maximum embedding dimension of 12, we find appropriate scaling behavior in the correlation sums for these data, indicating low dimensional deterministic dynamics. A correlation dimension of Dc ≈ 5.2 is measured in the scaling regime. This result is confirmed by recalculation using the Takens estimator and by surrogate data tests. We interpret this result as an approximation to the fractal dimension of a strange attractor governing chaotic dynamics in the menstrual cycle. We also use the time series to calculate the correlation entropy (K2 ≈ 0.008/τ) and the maximal Lyapunov exponent (λ ≈ 0.005/τ) for the system, where τ is the sampling time of the series.
Taken collectively, these results constitute significant evidence that the menstrual cycle is the result of chaos in a nonlinear dynamical system. This view of the menstrual cycle has potential implications for clinical practice, modelling of the endocrine system, and the interpretation of the perimenopausal transition.
The prevailing biomedical view of the female reproductive system, exemplified by the menstrual cycle, has traditionally been that changes in various hormone levels cause further well-defined changes in a cyclically repeating pattern . Despite this widespread view of menstruation, however, the empirical data show a high degree of variability that no current model accounts for . Such variability is usually discounted as being due to random factors of no theoretical interest, but consideration of the dynamics inherent in this system suggests another explanation. The endocrine system governing the menstrual cycle has multiple nonlinear feedback loops involving at least six hormones produced by the ovaries, the pituitary gland, and the hypothalamus. This system can be modelled as a set of coupled nonlinear delay differential equations. Envisioning the menstrual cycle in this way as the output of a nonlinear dynamical system, chaotic solutions that would account for the observed variability are a distinct possibility. This paper presents the results of an experimental test of the hypothesis that the human menstrual cycle is in fact the output of such a chaotic regime in a nonlinear dynamical system, including a characterization of this regime by measurements of various relevant parameters using time series analysis.
The study of physiological systems using such techniques has been widespread, but includes only a small number of reports on endocrine physiology or the menstrual cycle. Prank et al.  analyzed the variation of parathyroid hormone levels in the blood over 24 hours (with a 2 minute sampling time) in three subjects, measuring the correlation dimension, Lyapunov exponents, and correlation entropy for these variations. Noguchi et al.  measured the blood levels of growth hormone and prolactin over 24 hours in six subjects (with a 30 minute sampling time), while Ilias et al.  measured blood levels of growth hormone and cortisol over 24 hours in ten subjects (again at 30 minute intervals). The latter two studies report values of attractor dimension estimates extracted from their data before and after sleep deprivation. All of this work, however, used relatively small numbers of time series data points, limited by the great difficulty of obtaining large amounts of physiological data for hormone concentrations in the blood and of obtaining such data over an extensive period of time. In this paper, we use a novel source of data to circumvent these problems. No studies, to our knowledge, have used nonlinear dynamical methods to study the time variation of female reproductive system hormones, but Bai et al.  did use such methods to examine the effects of ovarian hormones and thyroid-related hormones on heart rate variability during different parts of the menstrual cycle.
A number of mathematical models for the menstrual cycle have been developed. Bogumil et al. [7, 8] developed an early but sophisticated model and simulation. Although nonlinearities are naturally built into this model, the simulation results required the addition of stochastic elements in order to exhibit the empirically observed variability in the menstrual cycle. This work was performed prior to a significant portion of our present physiological knowledge, however, and also prior to our present appreciation for the possibilities of chaotic dynamics in such systems. Increasingly accurate and sophisticated models have been developed and implemented more recently by Grigoliene and Svitra  and by Clark et al.  These models have been augmented by Reinecke and Deuflard  through the inclusion of more detail in the GnRH pulse generation, but this model has not been implemented yet. Although the recent models that have been implemented are able to reproduce measured hormone levels reasonably well over a single cycle, the observed variability of the cycles has not yet emerged naturally from these simulations. Other models have focused more on particular mechanisms such as the binding of hormones to receptors and calcium ion pumping through membranes [12, 13]. The focus of attention has so far generally been on periodic solutions and predictability of the modelled variables. We hope to influence the course of future modelling investigations with the presentation of the results herein.
where Δti is the ith menstrual cycle time length in the sequence. Equation (1) is essentially a device by which to create a formal time series using the information content available in the menstrual cycle time data. The rationale that underlies the validity of this device is based on the fact that the onset of menstruation is a discrete event that occurs when the body's endocrine system achieves some specific state, and that this state is by hypothesis deterministically related to all preceding and following states, since we are assuming that the endocrine system is indeed a dynamical system. It then follows that the time separating this state from the state of the system at time tn is likewise a function f(t) of the state of the system, and it is this function that we have sampled the nth measurement of with our definition. Although this is an unconventional sort of time series, being a time difference itself rather than some other separately measured variable, the reasoning that underlies it is basically the same as that which underlies the validity of the much-used phase space reconstruction techniques . For each individual woman, the sampling time was set equal to the average cycle length to eliminate stationarity. The final step in this construction is the concatenation of the data together into the time series used for analysis. A similar concatenation procedure has also been used in previous physiological studies . An example of the resulting time series for a subset of the data is shown in figure 1B.
We perform analyses using both this formal time series that we constructed and also the raw inter-event time data, Δti, treated as a series. The work of Castro and Sauer  has shown that such inter-event time data can be used to calculate correlation dimensions for two different models of event generation from the dynamical behavior of a system. The disadvantage of this method is that we don't know the details of the event generation mechanism for our present experimental case of a real physiological system. The disadvantages of the formal time series are that it may artificially introduce autocorrelation into the series (which, however, can be corrected for) and the justification for it is not truly rigorous. We argue, however, that concordance between the results of the analysis using the Δti and the results using the f n strongly suggests that both are valid.
A number of analyses were performed on this menstrual cycle data. Our objective, in part, was to look for evidence of chaos governing the dynamics of this system by examining the data for signs of a strange attractor. This was done using standard techniques, namely by performing a state space reconstruction with time delay embeddings  for the time series with time delay ΔT = 7τ for all of the analyses reported here, and simply using ΔT = τ for the cycle time sequence as outlined in Castro and Sauer .
can then be found from the slope of a log[C(R)] versus log[R] plot over some appropriate scaling region, and the Dc obtained by this method is known to be a good approximation to the fractal dimension of a strange attractor that has generated the time series data. This procedure was repeated for embedding dimensions ranging from d = 1 to d = 12. Since a strange attractor only fills a limited volume of the available state space, values of Dc that remain roughly constant as d increases indicate the existence of a low dimensional attractor governing the dynamics, and hence the likelihood of a deterministic chaotic system. Since data points closer in time are forced to be near each other by the time series construction protocol, we did not use the first 70 data points (nearest the point in question for that term of the summation) in the correlation sums, employing a correction for this autocorrelation problem suggested by Theiler . For the inter-event timing data (i.e. menstrual cycle lengths), a similar procedure is used, in this case constructing the embedding vectors out of consecutive values in the sequence. Correlation sums are computed in the same way and values of Dc found from the slopes of log[C(R)] versus Log[R] plots. Because the amplitude of this sequence is smaller than that of the time series, the scaling range of R is also considerably smaller, but the scaling is quite good within that range. Autocorrelation is minimal for this sequence, so no Theiler correction is needed (this fact was verified empirically).
To test for potential artifacts due to non-randomness in the data from correlated noise (which can masquerade as deterministic chaos by yielding low correlation dimensions), we recalculated Dc using surrogate data sets generated from the inter-event time sequence . The surrogate data was generated by randomizing the phases of the Fourier transform of the real data and then inverse transforming the resulting series. For the real data, the Takens estimator for d = 10 and R = 5.5 is computed to be ≈ 4.5, whereas for the surrogate data the Takens estimator is computed to ≈ 8.8. This procedure was repeated using two 2500 point subsets of the data, with similar results. Hence, the surrogate data tests indicate that the low correlation dimensions are the result of deterministic chaotic dynamics, not artifacts.
One of the interesting features of the results reported here is the comparatively low dimensionality of the system. The phenomenological details of the human menstrual cycle are extraordinarily complicated. Reinecke and Deuflhard , for example, have devised a model for the human menstrual cycle consisting of 43 differential equations with 191 parameters, and yet an attractor with a fractal dimension of ≈5.2 characterizing the dynamics of the human menstrual cycle suggests that only 6 degrees of freedom may be needed to describe these dynamics. This will hopefully stimulate a rich array of novel approaches to identify the agents that appear to govern the dynamics of the menstrual cycle, and it offers the possibility that simpler systems of equations may then be possible to analyze these dynamics.
The second important implication for modeling results is that the variability of the menstrual cycle, which has been well documented for many years, is a natural feature of the dynamics itself. This variability is not merely due to random fluctuations or external interference, and thus any results generated by implementations of model systems that do not exhibit this variability imply that the models have not captured some important part of the system's dynamics. Variability in the menstrual cycle should not be ignored by only considering average behaviors as is typical, nor imposed by means of ad hoc stochastic additions as Bogumil et al. [7, 8] did. Instead, the variability is intrinsic to the behavior of the chaotic system and valid models should reproduce comparable variability as a natural outcome of their implementations. Moreover, such models can in principle be tested, and parameters optimized, by comparing the various nonlinear measures reported in the present paper (Dc, K2, and λ) to the output of the models. We would assert that any model producing only perfectly periodic menstrual cycles is, at best, incomplete.
Beyond the details of particular models, the discovery of chaotic dynamics in the menstrual cycle has implications for the more general paradigmatic approach that is taken with regard to its behavior and the interpretation thereof. For example, an increase in variability might be construed as a natural consequence of the dynamics of the system rather than as a pathological deviation from normal behavior. Clinical goals associated with control and predictability may not be appropriate for a system that is known to be chaotic, and pharmacological interventions that impose regularity might need to be re-examined in light of this new context. The prevailing view of menopause as a senescent breakdown of the system should also be reconsidered in light of the idea that the menstrual cycle is the output of a nonlinear dynamical system and therefore might have a variety of possible regimes characterized by different values of relevant control parameters. Reinterpretations of this sort will be considered in more detail elsewhere, but far more analytical work is needed before any definitive conclusions are possible.
We are particularly interested in the issue of menopause and the perimenopausal transition. We are analyzing menstrual cycle data for the perimenopause in order to characterize the chaotic dynamics in that regime and compare it to the results of data from 20-40 year old women presented in this report. Significantly different values of the relevant measures (such as Dc, K2, and λ) during the perimenopause would indicate that the dynamical system (i.e. reproductive endocrine physiology) had undergone a phase transition to some new attractor, a conclusion that seems more consistent with existing evidence than the usual notions of senescence, breakdown, and pathology. However, it is likely that refinements in the analysis to improve the precision of these measurements are needed in order to enact this agenda.
The major conclusion of the present paper is that the human menstrual cycle is in fact the output of a nonlinear dynamical system in a chaotic regime, even in the most comparatively regular phase of its development during the 20-40 year age range. A quantitative characterization of this trajectory is provided by its correlation dimension Dc = 5.2 ± 0.7, correlation entropy K2 ≈ 0.008/τ, and largest positive Lyapunov exponent λ ≈ 0.005/τ (where τ represents the sampling time of the data). We believe that the evidence presented here for the chaotic nature of the menstrual cycle is persuasive and that our quantitative measures of its dynamics are reliable.
We would like to thank Phyllis Mansfield and the Tremin Research Program on Women's Health for permission to download and utilize their database of women's menstrual histories in this work.
- Goodman HM: Basic Medical Endocrinology. 2009, Amsterdam: Elsevier Academic Press, 4Google Scholar
- Treloar A, Boynton R, Behn B, Brown B: Variation of the human menstrual cycle through reproductive life. International Journal of Fertility. 1967, 12: 77-126.Google Scholar
- Prank K, Harms H, Brabant G, Hesch R, Dammig M, Mitschke F: Nonlinear dynamics in pulsatile secretion of parathyroid hormone in normal human subjects. Chaos. 1995, 5: 76-81. 10.1063/1.166089.View ArticleADSGoogle Scholar
- Noguchi T, Yamada N, Sadamatsu M, Kato N: Evaluation of self-similar features in time series of serum growth hormone and prolactin levels by fractal analysis: effects of delayed sleep and complexity of diurnal variation. Journal of Biomedical Science. 1998, 5: 221-225. 10.1007/BF02253472.View ArticleGoogle Scholar
- Ilias I, Vgontzas AN, Provata A, Mastorakos G: Complexity and non-linear description of diurnal cortisol and growth hormone secretory patterns before and after sleep deprivation. Endocrine Regulations. 2002, 36: 63-72.Google Scholar
- Bai X, Li J, Zhou L, Li X: Influence of the menstrual cycle on nonlinear properties of heart rate variability in young women. American Journal of Physiology Heart and Circulatory Physiology. 2009, 297: 765-774. 10.1152/ajpheart.01283.2008.View ArticleGoogle Scholar
- Bogumil RJ, Ferin M, Rootenberg J, Speroff L, Vande Wiele RL: Mathematical studies of the human menstrual cycle. I. Formulation of a mathematical model. Journal of Clinical Endocrinology & Metabolism. 1972, 35: 126-142.View ArticleGoogle Scholar
- Bogumil RJ, Ferin M, Vande Wiele RL: Mathematical studies of the human menstrual cycle. II. Simulation performance of a model of the human menstrual cycle. Journal of Clinical Endocrinology & Metabolism. 1972, 35: 144-156.View ArticleGoogle Scholar
- Grigoliene R, Svitra D: Mathematical model of the female menstrual cycle and its modifications. Informatica. 2000, 11: 411-420.MathSciNetMATHGoogle Scholar
- Clark LH, Schlosser PM, Selgrade JF: Multiple stable periodic solutions in a model for hormonal control of the menstrual cycle. Bulletin of Mathematical Biology. 2003, 65: 157-173. 10.1006/bulm.2002.0326.View ArticleGoogle Scholar
- Reinecke I, Deuflhard P: A complex mathematical model of the human menstrual cycle. Journal of Theoretical Biology. 2007, 247: 303-330. 10.1016/j.jtbi.2007.03.011.View ArticleMathSciNetGoogle Scholar
- Blum JJ, Reed MC, Janovick JA, Conn PM: A mathematical model quantifying GnRH-induced LH secretion from gonadotropes. American Journal of Physiology Endocrinology and Metabolism. 2000, 278: 263-272.Google Scholar
- Washington TM, Blum JJ, Reed MC, Conn PM: A mathematical model for LH release in response to continuous and pulsatile exposure of gonadotrophs to GnRH. Theoretical Biology and Medical Modelling. 2004, 1: 9-10.1186/1742-4682-1-9.View ArticleGoogle Scholar
- Mansfield P, Bracken S: Tremin: A History of the World's Oldest Ongoing Study of Menstruation and Women's Health. 2003, Lemont: East Rim PublishersGoogle Scholar
- Packard NH, Crutchfield JP, Farmer JD, Shaw RS: Geometry from a time series. Physical Review Letters. 1980, 45: 712-716. 10.1103/PhysRevLett.45.712.View ArticleADSGoogle Scholar
- Castro R, Sauer TD: Forecasting and dimension calculations from event timing data. Nonlinear Phenomena in Complex Systems. 1999, 2: 42-51.Google Scholar
- Grassberger P, Procacia I: Characterization of strange attractors. Physical Review Letters. 1983, 50: 346-349. 10.1103/PhysRevLett.50.346.View ArticleADSMathSciNetGoogle Scholar
- Theiler J: Spurious dimension from correlation algorithms applied to limited time-series data. Physical Review A. 1986, 34: 2427-2432. 10.1103/PhysRevA.34.2427.View ArticleADSGoogle Scholar
- Takens F: On the numerical determination of the dimension of an attractor. Dynamical Systems and Bifurcations. Edited by: Braaksma B, Braer H, Takens F. 1985, Berlin: Springer-Verlag, 99-106. full_text.View ArticleGoogle Scholar
- Theiler J, Eubank S, Longtin A, Galdrikian B, Farmer JD: Testing for nonlinearity in time series: the method of surrogate data. Physica D. 1992, 58: 77-94. 10.1016/0167-2789(92)90102-S.View ArticleADSMATHGoogle Scholar
- Grassberger P, Procaccia I: Estimation of the Kolmogorov entropy from a chaotic signal. Physical Review A. 1983, 28: 2591-2593. 10.1103/PhysRevA.28.2591.View ArticleADSGoogle Scholar
- Rosenstein MT, Collins JJ, De Luca CJ: A practical method for calculating largest Lyapunov exponents from small data sets. Physica D. 1993, 65: 117-134. 10.1016/0167-2789(93)90009-P.View ArticleADSMathSciNetMATHGoogle Scholar
- Kantz H: A robust method to estimate the maximal Lyapunov exponent of a time series. Physics Letters A. 1994, 185: 77-87. 10.1016/0375-9601(94)90991-1.View ArticleADSGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.