What statistical physics sees in the heartbeat

The first post in this series analyzed an ECG with a linear signal processing pipeline: a band-pass filter, R-peak detection to form a tachogram of RR intervals, and Welch’s method to estimate the spectrum, yielding the low- and high-frequency bands and the LF/HF ratio.

That pipeline rests on an implicit assumption: that the information of interest resides in the amplitude and the spectrum of the variability, specifically how much it varies (the standard deviation of the intervals, SDNN), at what rates (the spectral bands), and in what balance (LF/HF). This post concerns what that assumption omits. A healthy heart is not a noisy metronome but a fractal, self-organizing system, and the salient property of its rhythm is not the magnitude of the wandering but the way the wandering is structured across time scales. Linear measures are largely blind to that structure; nonlinear measures are not.

The distinction is clinically consequential. In failing hearts, in autonomic neuropathy, and in the slow progression toward decompensation, the variability can appear unchanged to the linear measures while its temporal architecture deteriorates. A summary restricted to SDNN and LF/HF can therefore miss the change entirely.

Fractal physiology and self-similarity

Two physiological ideas frame the problem.

The first is a refinement of homeostasis: a healthy organism is not held at a fixed set-point but in a restless, fluctuating equilibrium. The fluctuations are not noise to be removed; they constitute the regulation. An excessively steady heart rate is a warning sign rather than a sign of health.

The second is fractal physiology [1]: these fluctuations are self-similar across time scales. A healthy RR series examined at the scale of seconds, of minutes, and of tens of minutes exhibits statistically similar structure at each, the signature of a control system with no single characteristic time, operating across many coupled loops simultaneously. This self-similarity appears as a 1/f (pink-noise) spectrum and a power-law scaling of fluctuations. Pathology tends to move the heart away from 1/f, either toward rigid, random-walk dynamics (loss of responsiveness) or toward uncorrelated noise (loss of integration).

The amplitude statistics of Part I cannot resolve this. The most direct demonstration is to destroy the structure deliberately.

What linear measures miss

Consider a healthy, 1/f-like RR series, and a second series formed by shuffling it: the same values in random order. Shuffling preserves every amplitude statistic that does not depend on order. The mean is unchanged, and so is SDNN, the most frequently reported heart-rate variability (HRV) measure [2]; the histogram is identical. By those measures the two recordings are indistinguishable. Dynamically they are not.

The upper trace shows slow undulations and drift that persist across hundreds of beats; the middle trace, the same values reordered, is uncorrelated and structureless. The mean and SDNN are identical (850 ms and 45 ms). Yet one is a correlated fractal process and the other is white noise. The lower panel shows the DFA scaling exponent (defined in full below): it falls from $\alpha_1 = 1.03$ for the healthy series to $0.55$ for the shuffled one.

This is the central motivation. The magnitude of the variability is conserved; its organization is destroyed; the standard measures do not respond. A measure of the organization is required.

One caveat sharpens the claim. The power spectrum is not itself blind to order: by the Wiener–Khinchin theorem it is the Fourier transform of the autocorrelation, so the shuffle that whitens the autocorrelation also flattens the spectrum. Two things, however, escape the linear pipeline of Part I. SDNN is order-independent and registers nothing; and the reported LF/HF summary compresses the spectrum into a few fixed bands, discarding the broadband 1/f slope in which the scaling actually lives. The nonlinear measures that follow recover that scaling directly and, beyond it, capture higher-order dependence that no power spectrum, a purely second-order quantity, can express.

Poincaré plots: beat-to-beat geometry

The simplest nonlinear view is geometric. Each interval is plotted against the next, $RR_n$ on the horizontal axis and $RR_{n+1}$ on the vertical axis. A perfectly regular rhythm collapses to a single point; a healthy rhythm forms an elongated cloud. The cloud’s shape is captured by fitting an ellipse [3]. SD1 measures the width perpendicular to the line of identity, picking out the beat-to-beat differences; it is dominated by vagal (parasympathetic) activity. SD2 measures the length along the line of identity, capturing longer-range trends. Formally:

Their ratio SD1/SD2 is a compact descriptor of the balance between short- and long-term regulation.

The Poincaré description is intuitive and clinically common, but it remains essentially second-order: SD1 and SD2 are functions of variances. The method detects the presence of structure without quantifying it across scales. Before arriving at a measure that does, it is worth posing a simpler question: how predictable is the next beat from the ones before it? The answer illuminates exactly what a scale-resolved measure must add.

Sample entropy: the predictability of the next beat

Sample entropy (SampEn) [4] poses an information-theoretic question: given two short segments of the rhythm that match for $m$ beats, how often do they continue to match for $m+1$ beats? Frequent continuation indicates a regular, predictable signal; infrequent continuation indicates a complex, irregular one. With tolerance $r$ (commonly $0.2$ times the standard deviation) and embedding length $m$ (commonly $2$),

where $B$ is the number of template pairs (length-$m$ windows that agree beat-by-beat within tolerance $r$) and $A$ the subset that still agree when extended to length $m+1$. Higher SampEn corresponds to lower predictability.

A caution is warranted, because the point is easily misread: irregularity is not health. A Brownian (random-walk) rhythm is smooth and highly predictable, giving low SampEn, and is pathological. White noise is maximally unpredictable, giving high SampEn, and is also pathological. Health lies between the two, at the 1/f regime. SampEn is therefore informative but non-monotonic; greater entropy is not uniformly better. This non-monotonicity motivates a measure that locates the rhythm on the continuum between rigid and random, which is what DFA provides.

Detrended fluctuation analysis

DFA [5] estimates the spectral exponent of the rhythm. A fractal time series has a power spectral density that obeys $S(f) \propto f^{-\beta}$, where $\beta$ is the spectral exponent: $\beta \approx 0$ for white noise (flat spectrum), $\beta \approx 1$ for a healthy 1/f rhythm, $\beta \approx 2$ for Brownian motion. In principle one could estimate $\beta$ directly from the periodogram. In practice, physiological recordings are short and non-stationary, and the spectral estimate is correspondingly unreliable. DFA estimates the equivalent exponent $\alpha$ (related to $\beta$ by $\alpha = (\beta + 1)/2$) from the time domain, robustly even when the signal drifts. The method was not developed in a cardiology laboratory: introduced in the early 1990s by Peng, Havlin, Stanley, and Goldberger at Boston University to detect long-range correlations in non-stationary DNA nucleotide sequences [6], it reached cardiac rhythm by the same mathematical route. The procedure has three steps.

First, integrate. The RR series is converted into a profile, the cumulative sum of the mean-removed intervals:

Second, detrend in windows. The profile $y(k)$ is divided into boxes of size $n$; within each box a local polynomial trend (linear here) is fitted and subtracted, which is what makes DFA robust to nonstationarity. The root-mean-square of the residuals is

Third, scale. The computation is repeated across box sizes, and $F(n)$ is examined against $n$ on logarithmic axes. For a self-similar process the relationship is a straight line, $F(n) \sim n^{\alpha}$, whose slope $\alpha$ is the scaling exponent. The short-term exponent over $4 \le n \le 16$ beats is denoted $\alpha_1$.

On the synthetic series DFA recovers the planted exponents, $0.59$, $1.03$, and $1.54$, confirming the estimator before it is applied to real data. It also accounts for the shuffling result: shuffling whitens the spectrum and drives $\alpha_1$ from approximately $1.0$ toward $0.5$, while leaving the mean and SDNN unchanged.

The clinical relevance is that a reduced or drifting $\alpha_1$ has been associated with autonomic dysfunction and with adverse outcomes in heart failure [7]: the failing heart tends to lose its 1/f signature, moving toward either over-regular ($\alpha_1 \to 1.5$) or disorganized ($\alpha_1 \to 0.5$) dynamics before the linear summaries change. This is the kind of early structural drift that a longitudinal monitor would aim to detect.

Sensitivity to artifacts

The preceding methods all assume a clean RR series. In practice that assumption frequently fails. DFA, SampEn, and the Poincaré indices all read the fine temporal structure of the intervals, which makes them acutely sensitive to the two artifacts that dominate real-world heart-rate data: ectopic beats (a premature beat followed by a compensatory pause, which injects a large artificial step) and dropped beats (a missed detection that merges two intervals into one). The effect of even a small rate of each on $\alpha_1$ is severe.

The clean rhythm has $\alpha_1 = 1.03$. Ectopic beats at $1.5\%$ of the total reduce it to $0.35$; a $6\%$ drop rate reduces it to $0.64$. A measurement that should indicate healthy fractal organization instead indicates disorganization, not because the underlying dynamics changed but because a small number of intervals are incorrect. The signal is intact; the acquisition corrupts the estimate.

This has a direct implication for consumer wearables. The beat-to-beat timing these methods require is reliably obtained from a chest-strap ECG under controlled conditions. The wrist-worn optical sensor (photoplethysmography, PPG) in a consumer watch, the sensor that most patients possess, is precisely the source that produces ectopic-like spikes and dropped beats under motion, poor perfusion, or low peripheral temperature. The most prognostic HRV measures are thus the most fragile, and they depend on the least reliable hardware available to the patient. This is the constraint that frames the engineering problem downstream.

Summary

The heart is not a metronome, the variability is the regulation, and its fractal organization, quantified by $\alpha_1$, carries information that the linear summaries cannot retrieve. The remaining difficulty is measurement: obtaining the artifact-free interval sequences that the nonlinear measures require.

References

[1] A. L. Goldberger, L. A. N. Amaral, J. M. Hausdorff, P. Ch. Ivanov, C.-K. Peng, and H. E. Stanley, “Fractal dynamics in physiology: alterations with disease and aging,” Proceedings of the National Academy of Sciences, vol. 99, suppl. 1, pp. 2466–2472, 2002.

[2] Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, “Heart rate variability: standards of measurement, physiological interpretation, and clinical use,” Circulation, vol. 93, no. 5, pp. 1043–1065, 1996.

[3] M. Brennan, M. Palaniswami, and P. Kamen, “Do existing measures of Poincaré plot geometry reflect nonlinear features of heart rate variability?,” IEEE Transactions on Biomedical Engineering, vol. 48, no. 11, pp. 1342–1347, 2001.

[4] J. S. Richman and J. R. Moorman, “Physiological time-series analysis using approximate entropy and sample entropy,” American Journal of Physiology-Heart and Circulatory Physiology, vol. 278, no. 6, pp. H2039–H2049, 2000.

[5] C.-K. Peng, S. Havlin, H. E. Stanley, and A. L. Goldberger, “Quantification of scaling exponents and crossover phenomena in nonstationary heartbeat time series,” Chaos, vol. 5, no. 1, pp. 82–87, 1995.

[6] C.-K. Peng, S. V. Buldyrev, A. L. Goldberger, S. Havlin, F. Sciortino, M. Simons, and H. E. Stanley, “Long-range correlations in nucleotide sequences,” Nature, vol. 356, pp. 168–170, 1992.

[7] H. V. Huikuri et al., “Fractal correlation properties of R-R interval dynamics and mortality in patients with depressed left ventricular function after an acute myocardial infarction,” Circulation, vol. 101, no. 1, pp. 47–53, 2000.