Updated September 26, 2022; text and figure added in Does HRV index autonomic tone?, and text added in Respiratory sinus arrhythmia.
Updated August 10, 2022; text added in “Measuring RSA” section.

Heart rate variability (HRV) is a fascinating metric. It’s been around for over a century in the clinical and research space, and has exploded in popularity both in research and household settings because it can be measured easily and noninvasively, providing a window into the complexities of physiological and psychological processes without perturbing the underlying systems. Athletes use it to optimize training and recovery, doctors use it to predict patient outcomes, and psychologists use it to measure a whole host of cognitive activities…yet still today, nobody seems to be able to agree on precisely what these measurements actually mean.

As I’ll explain below, HRV is not a single measurement but rather a biological phenomenon that can be quantified in a number of different ways, which one needs a fair bit of technical and mathematical expertise to correctly interpret. A thorough understanding of the physiology underlying HRV has lagged well behind enthusiasm for its potential, which has led to a number of very persistent misunderstandings and a glut of unhelpful or misleading research. Reading through recent reviews and editorials, you often get the sense that the experts in the field are sighing with exasperation at the fact that controversies that ought to have been laid to rest in the late 1990s are still alive and well today.

But I’m getting ahead of myself. In this article, I’ve attempted to give an understandable but still thorough and detailed review of HRV, including both the physiology underlying this phenomenon and the ways it’s measured. I’ve done my best to pull together the most current and well-supported theories from a number of different research papers regarding how each metric ought to be interpreted, but I want to be very clear that the field is still far from consensus and I am not an expert! I expect to periodically update this article as research advances and my understanding improves.

I also want to warn you that this is not a practical guide to HRV. This is a scientific primer on the nature of HRV as a biological phenomenon and a statistical entity. I embarked on this journey because I needed to understand all of this in order to be able to make sense of the cornucopia of fascinating HRV research that exists, so consider this a foundation upon which future (more practical and perhaps more interesting) articles will be written.

One final note, regarding references. I cannot overstate the number of papers I read in the process of writing this article. It was truly absurd. As such, most of the things I say here are based on an understanding of the material gleaned over the course of much reading and cross-referencing and following rabbit holes, as opposed to direct citations from a specific paper. I’ve included some specific sources for specific bits of info, but any statement that doesn’t have a direct citation was drawn from some combination of the list of papers in the “references” section. (I tried to only include papers I deemed worthwhile sources, rather than including every single paper I read.)

Table of Contents
What is heart rate variability?
        A very brief history of heart rate variability
        Phenomenological vs physiological analysis of HRV
Arbiters of heart rate control
        Sinoatrial node
        Autonomic innervation
        Intrinsic cardiac nervous system
        Humoral influences
Major physiological contributors to rhythmic heart rate fluctuations
        Respiratory sinus arrhythmia
        The baroreflex
        Mayer waves
Measuring heart rate variability
        Time-domain measurements
        Frequency-domain measurements
        Nonlinear measurements
        Mathematical and technical challenges
Interpreting heart rate variability
        HF: the respiratory band
        LF: Mayer waves and the baroreflex
        VLF and ULF: circadian and other rhythms
        SDNN and SDANN: overall and long-term variability
        RMSSD and pNN50: short-term variability
        Measuring respiratory sinus arrhythmia (RSA)
Controversies and drama in HRV research
        Is HRV hopelessly confounded by heart rate?
        Does HRV index autonomic tone?
        Does the LF/HF ratio index sympathovagal balance?
Summary and conclusions
References

What is Heart Rate Variability?

Heart rate variability (HRV) is simply the normal beat-to-beat variation in heart rate observed in humans and many other animals. If you look at an ECG of a healthy human heart, you’ll see that the time between beats is not constant; that’s the biological phenomenon known as HRV.

This alone can be a surprise to people unfamiliar with the concept, since for a long time there was a common misconception that the heart beats at a steady, constant speed – like a metronome.

But when you think about it from the perspective of homeostasis, it makes perfect sense. Our internal and external environments are constantly changing, and a healthy heart can quickly adapt to best meet the demands of the moment. And indeed, higher HRV is generally indicative of better health (with a few exceptions), because it demonstrates physiological responsiveness and resiliency.

A Very Brief History of Heart Rate Variability

Most studies investigating HRV have occurred during the past 40 years. In the early days, many researchers thought HRV was simply a technological or experimental error; the idea that the heart kept a stable rhythm still prevailed, despite the fact that ancient Greek, Chinese, and Indian medicine acknowledged and used pulse variations as a common diagnostic.

But once the phenomenon was acknowledged and technology was adequately precise, HRV rapidly became a metric of interest across clinical medicine and psychology. Its initial popularity was in large part because it could be measured easily and noninvasively, but it quickly proved itself a valuable metric on its own merits.

Evidence started accumulating that HRV might be able to index cardiac autonomic function. Certain HRV parameters were observed to correlate very well with known vagal activity, and others were observed to correlate with sympathetic activity.

In the 1960s and 70s, researchers also began to discover that HRV was useful for predicting certain clinical outcomes, with reduced HRV reliably presaging fetal distress, diabetic autonomic neuropathy, and mortality risk after a heart attack. And now, we know that HRV is a relevant predictive or associative variable in an extraordinarily wide range of contexts and disciplines related to human health.

For instance, HRV is a good metric for a person’s level of physical fitness, and it’s becoming more and more common for athletes to use HRV as a recovery metric to inform their training schedule and intensity.

HRV also appears to be a good metric for a person’s level of mental fitness, with higher HRV being strongly associated with contemplative practices such as mindfulness meditation. Higher HRV is also associated with better emotional regulation, and with positive emotional states in general.

On the flip side of the same coin, HRV tends to be inversely correlated with states of disease, with lower average HRV observed in nearly every disease state (both physical and mental) in which it has been measured. And within a given patient population, risk stratification by HRV often predicts disease outcomes.

All of these observations make sense when HRV is viewed as an indication of adaptability and resilience. The ability of the body to maintain homeostasis in the face of internal and external perturbations is dependent on the give and take of innumerable autonomic regulatory systems and feedback loops, and HRV gives us a glimpse into how that dynamic process is playing out.

As McCraty and Shaffer put it in their 2015 review paper, “an optimal level of HRV within an organism reflects healthy function and an inherent self-regulatory capacity, adaptability, or resilience. Too much instability, such as arrhythmias or nervous system chaos, is detrimental to efficient physiological functioning and energy utilization. However, too little variation indicates age-related system depletion, chronic stress, pathology, or inadequate functioning in various levels of self-regulatory control systems.”

Phenomenological vs Physiological Analysis of HRV

Before getting into the meat of this article, I want to briefly introduce a theme that occurs throughout the body of HRV research. A framework, if you will. There are two main ways one can view HRV metrics in clinical and research settings:

As an observed variable that can statistically correlate with or predict various states of health and disease
As a physiological phenomenon that is the result of a number of complex homeostatic feedback mechanisms

All this means is that it’s very possible to observe something, and see that it correlates with and even predicts other things, without actually understanding the thing. For instance, epidemiologists and statisticians can observe that low HRV reliably predicts all-cause mortality, without knowing why (or even without having any clue what HRV is).

This type of analysis can be extremely valuable if taken for what it is, but the lack of complete understanding can easily lead to false conclusions and misinterpretations, from the classic “causation from correlation” mistake to more subtle ones.

If you’re interested, this paper does a good job highlighting how the historical approach of treating HRV as an observed variable without thinking through the physiological mechanisms at play has led to a number of missteps and limitations in the field of HRV research (but note that the namesake of the paper, the polyvagal theory itself, is questionable…more on this in another article).

The preeminent 1996 HRV task force report addresses this issue outright, stating:

Efforts should be made to find the physiological correlates and the biological relevance of various HRV measures currently used….The use of markers of autonomic activity is very attractive. However, unless a tenable mechanistic link between these variables and cardiac events is found, there is an inherent danger of concentrating therapeutic efforts on the modification of these markers. This may lead to incorrect assumptions and serious misinterpretations.
Heart rate variability: standards of measurement, physiological interpretation and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology (1996)

So in the interest of taking a physiological approach to understanding HRV, let’s start by looking at the physiology of heart rate control.

Arbiters of Heart Rate Control

Sinoatrial Node

First things first: the heart has a default rate. In the absence of external influences, the average human heart will beat at around 100 bpm, a pace initiated and maintained by the sinoatrial (SA) node (the intrinsic “pacemaker”).

Some stimuli affect heart rate by directly affecting the SA node. For instance, the SA node contains mechanoreceptors that are sensitive to stretching. When you take a breath, pressure in your ribcage increases, activating these stretch receptors and temporarily increasing heart rate. This happens on every breath. Boom – our first source of HRV.

Autonomic Innervation

The two branches of the autonomic nervous system – sympathetic and parasympathetic – each have neurons that terminate at the SA node. Sympathetic neurons increase heart rate, while parasympathetic neurons decrease it.

At rest, parasympathetic control of the SA node predominates, such that healthy humans have resting heart rates significantly below the intrinsic rate of ~100 bpm.

Of note, sympathetic impulses to the heart act more slowly than parasympathetic impulses: it can take up to 5 seconds before a sympathetic stimulus induces a heart rate increase, and the effects of even a brief stimulus can last up to 10 seconds. On the other hand, impulses from the (parasympathetic) vagus nerve affect heart rate almost instantaneously. Keep this in mind for later.

Intrinsic Cardiac Nervous System

These autonomic branches don’t always exert their effects directly on the SA node. The heart also has its own intrinsic nervous system, a web of interconnected neurons that can be likened to the enteric nervous system of the gut. Many regulatory mechanisms and feedback loops that affect heart rate (among other cardiac functions that we don’t care about in this article, such as contractility) are present within this intrinsic cardiac nervous system, and it also receives and integrates external signals – such as additional sympathetic and parasympathetic neurons.

In this context, sympathetic activation still generally increases heart rate while parasympathetic activation decreases it, but since these signals are received and processed first by the intrinsic cardiac nervous system before reaching the SA node, there’s a bit more complexity involved in the collection of effects resulting from a given nerve impulse.

Humoral Influences

The heart is also responsive to humoral signals – “humoral” meaning simply any messengers that reach the heart in body fluids (such as blood) rather than neural impulses. These effects are much slower than nerves, on the scale of 30 seconds or more.

Major Physiological Contributors to Rhythmic Heart Rate Fluctuations

One explanation for HRV that I’ve seen tossed around is that the variability in heart rate is created by the competing influence of the sympathetic and parasympathetic nervous systems on heart rate. It’s easy to see how someone could come to this conclusion, given that heart rate is primarily under sympathetic/parasympathetic control, and that the two branches of the nervous system have opposite effects on heart rate. But a tug-of-war between the parasympathetic and sympathetic nervous systems is not actually what’s going on.

To help better understand subsequent sections, recall that the heart has an intrinsic rate of ~100 bpm, and that it’s the tonic influence of the vagus (parasympathetic) nerve that keeps resting heart rate in humans lower than that. So if the body wanted to increase heart rate, it could fire some sympathetic neurons…or, it could just reduce vagal influence. Take the foot off the brake, if you will, rather than stepping on the accelerator.

And indeed, “taking the foot off the break” appears to be a far greater contributor to HRV than a back-and-forth between parasympathetic and sympathetic influence. Below, I’ll describe the two main sources of short-term fluctuation in heart rate: respiratory sinus arrhythmia (RSA), and the baroreflex associated with Mayer waves. And at least one, but possibly both, of these processes are primarily parasympathetically mediated.

Respiratory Sinus Arrhythmia

The best-understood contributor to HRV is respiratory sinus arrhythmia, or RSA: the increase in heart rate during inhalation and decrease in heart rate during exhalation, creating a rhythmic variability in heart rate at the same frequency as respiration.

This phenomenon was observed and noted in the literature long before we had any fancy techy ways to measure HRV, simply by feeling the pulse of a human or animal and noticing it increase and decrease in time with their breathing. Early HRV research focused almost exclusively on RSA, often with little or no distinction made between RSA and the concept of HRV in general.

You might recall that I already mentioned one component of RSA – the stretch receptors in the SA node. However, this is only a very tiny contributor to the overall amplitude of RSA in most people; most of the variance in heart rate from inhale to exhale is mediated by the vagus nerve, with nerve traffic decreasing during inhalation (thus increasing heart rate) and increasing during exhalation (thus decreasing heart rate). This effect has been described as a “respiratory gate.”

There are a few different theories as to why this happens, and it remains a source of debate and exploration (good review in the introduction here). The two main positions in the debate are that RSA is created by the blood pressure changes associated with breathing (summarized here), and that a central mechanism in the medulla controls both respiration and the associated variability in heart rate, with blood pressure changes being secondary rather than causal (summarized here).

The two papers cited above present the two positions debate-style, with researcher Dwain Eckberg arguing for the latter stance (with a touch of the academic sass that can be found many places throughout the HRV literature):

Before I accept the view that R-R interval fluctuations at respiratory frequencies are baroreflex responses, someone must explain 1) how an arterial pressure change can, within 0.1 s, speed or slow the appearance of the next P wave; 2) how the kinetics of sinoatrial nodal responses to acetylcholine are modulated systematically by body position and breathing frequency; and most difficult of all, 3) how baroreflex R-R interval responses can occur before the arterial pressure changes that provoke them. I submit that the last possibility, that effects can precede causes, turns logic on its head.”
Dwain Eckberg. Point:counterpoint: respiratory sinus arrhythmia is due to a central mechanism vs. respiratory sinus arrhythmia is due to the baroreflex mechanism. (2009)

It appears that the general consensus is indeed for the involvement of a central mediator, but that other mechanisms (including baroreflex physiology) are likely at play as well. John Karemaker, taking the other side in the debate, attempts to highlight the complexity involved and suggests that the debate is oversimplifying complex physiology.

Either way, there is broad consensus that the final heart rate effects seen with RSA are the result of fluctuating impulses from the vagus nerve. However, it’s important to note that RSA is only a good marker for the final vagal effects upon the heart; it is not a good marker for central vagal outflow or tone, since certain factors (such as sympathetic tone) can decrease RSA while central vagal outflow remains constant (per Grossman and Taylor 2007).

The Baroreflex

Another well-characterized contributor to HRV is the baroreflex: the feedback loop responsible for regulating arterial pressure. Baroreceptors are neurons found mainly in the aortic arch and the carotid sinuses of the neck that are sensitive to the stretching of blood vessels. Their activation (indicating an increase in blood pressure) elicits several compensatory responses, including reductions in stroke volume and blood vessel dilation, but the response that is relevant to HRV is a reduction in heart rate.

Mayer Waves

One very interesting thing about blood pressure and the baroreflex is that researchers have observed consistent, rhythmic oscillations in arterial pressure. These oscillations are coherent with oscillations in sympathetic nervous activity to the blood vessels, and they occur at a characteristic frequency within a given species. In humans, the frequency of Mayer waves is 0.1 Hz, equating to a period of 10 seconds.

Thanks to the the baroreflex, these regular fluctuations in arterial pressure create corresponding fluctuations in heart rate at that same frequency, as the heart rate oscillations act to buffer the Mayer waves.

Mayer waves are a fascinating area of research in themselves, but to avoid going down a rabbit hole, I’ll reference just two papers for a broad overview: one excellent review from 2006, and an update by the same author in 2019 essentially confirming that no significant advances in understanding have been made in the preceding 13 years. This paper about rhythms in sympathetic nerve activity in general also looks fascinating.

Now, as to whether baroreflex heart rate changes are sympathetically or parasympathetically mediated, there appears to be no consensus. Most of the preeminent HRV researchers seem to refer to baroreflex heart rate effects as primarily (or even exclusively) vagally mediated. For instance, “Direct evidence in sino-aortic baroreceptor denervated rats and indirect evidence in healthy humans support the conclusion that LF oscillations of HR are a vagal baroreflex response to underlying Mayer waves.” (Source: comments found in the supplemental materials of this article)

However, I’ve seen others describe the baroreflex somewhat nonspecifically as reciprocal control by the vagus and sympathetic neurons.

My two cents is that at rest, it wouldn’t make much sense for the body to press on the accelerator (activate sympathetic neurons) every 10 seconds when it could simply ease up on the brake, especially since vagal nerve responses are significantly quicker than sympathetic responses. But unfortunately we don’t currently have the ability to directly measure nerve traffic to the SA node in humans, so the precise autonomic physiology remains to be elucidated.

Measuring Heart Rate Variability

With that background in physiology out of the way, we can now boldly go forth into the measurement and interpretation of HRV. The way HRV is talked about in both popular and academic contexts might lead one to believe that it’s measured one way, as if the answer to “what’s your HRV?” is as simple as the answer to “what’s your heart rate?”

But think about it for a moment – if you were handed a 24-hour ECG printout, including the inter-beat intervals (aka RR intervals) in milliseconds for each heartbeat as depicted below, and somebody asked you what that person’s HRV was, how would you go about answering?

Your first instinct might be to use some form of statistical math. If that’s the case, you may arrive at something like one of the time-domain measurements discussed below. However, other HRV parameters (such as frequency-domain and nonlinear measures) aren’t quite so intuitive.

Most of the parameters and standards that are still used today were laid out in the 1996 report from the Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. The report summarized and provided practical recommendations for collecting time-domain and frequency-domain measures of HRV, discussed what was known at the time about the physiology behind the various HRV metrics, and reviewed clinical applications.

A subsequent report from the task force was published in 2015, but rather than providing any updates, it introduced a new third category of HRV metrics: nonlinear measures.

The next three sections will review these three categories (time-domain, frequency-domain, and nonlinear) in more detail.

Time-Domain Measurements

Time-domain measurements analyze HRV as a series of data points over time. These data points can either be RR interval (those millisecond measurements in the image above), or instantaneous heart rate. Either way, you end up with a collection of numbers, and you can perform various statistical calculations on that dataset to quantify the variability. The most common are:

SDNN: the standard deviation of NN* intervals
SDANN: the standard deviation of the average NN intervals for each 5 min segment of a 24-hour HRV recording
RMSSD: the root mean square of the differences between successive NN intervals. Each interval is subtracted from the one next to it to get the difference, each difference is squared, all those squared values are averaged, then the square root is taken
pNN50: the percentage of adjacent NN intervals that differ from each other by more than 50 ms

*An NN interval is the same as an RR interval, but excludes heartbeats classified as “abnormal,” i.e. not initiated by the SA node.

For other examples of time-domain measurements (including some that use geometric rather than statistical calculations), you can check out Table 1 of this review paper.

Mathematically inclined readers may notice that the calculation of each of these metrics requires defining a time period. In other words, what’s our sample? Are we finding the standard deviation of 100 numbers or 100,000?

The monitoring period could range anywhere from less than a minute to 24 hours, with longer sampling times unsurprisingly leading to greater variability. Most time-domain HRV analysis is done in 5 minute segments, because the ability to derive meaningful physiological interpretations from the metrics declines substantially over longer spans of time.

As one applied example, the Oura ring uses the RMSSD metric to measure nighttime HRV, and performs this calculation for each 5-minute chunk of time while you’re asleep. The average HRV number it reports for the night is the average of all of the RMSSD measurements for all of the 5-minute chunks of time that night. Oura also reports the maximum measurement, and gives you a graph of each measurement plotted against time for the whole night.

Frequency-Domain Measurements

A less intuitive way to measure HRV is through frequency-domain measurements, which use power spectral density analysis (usually Fourier transform or autoregression) to show how total power (variance) is distributed as a function of frequency. This type of signal analysis is also used in other fields such as geology (think seismic waves) and communications (think radio waves).

This method of analysis admittedly pushes the boundaries of my mathematical comprehension, but I’ll try to provide a simplified explanation. This guide from BetterExplained.com was fairly helpful for understanding the basics of Fourier transform, and also links to quite a few additional resources on the topic.

Imagine a wave on a graph, like a sine wave, but rather than a regular oscillation it’s going up and down seemingly at random. Spectral analysis takes that data and finds the collection of regular oscillations that, when combined, result in the observed signal. Each of those separate oscillation “ingredients” has a set frequency (measured in Hertz) at which it oscillates, and each contributes a different amount of power, or variance, to the overall observed variability in the signal.

One way to think about it is to picture a boat bobbing in the ocean. It doesn’t bob in a regular rhythm; the waves hitting it vary in size and shape and frequency, tossing the boat this way and that. But if you zoom out, you can see that a series of regular waves has been generated a jetski that just whizzed by on the left, and another series of waves is coming from a giant yacht that is slowly passing on the right, and still other slower waves are coming from behind as the tide rolls in. The sum of all of those (rhythmic) influences makes up the final (random) bobbing motion of the boat.

For the purposes of the HRV signal, researchers generally group these variability ingredients into four categories:

HF: high-frequency (between 0.15 Hz and 0.4 Hz, or 2.5 and 7 seconds)
LF: low-frequency (between 0.04 Hz and 0.15 Hz, or 7 and 25 seconds)
VLF: very-low-frequency (between 0.0033 and 0.04 Hz, or 25 and 333 seconds [5.6 min])
ULF: ultra-low-frequency (below 0.0033 Hz, or 333 seconds [5.6 minutes])

The HRV metrics themselves are either the total absolute power within a frequency band, or the relative power within a frequency band after being normalized to total power. “Total power” in a spectral analysis can be thought of as essentially the same thing as SDNN – a nonspecific quantification of overall variability in heart rate for a given sampling time.

As with time-domain measurements, spectral analysis can be conducted on heart rate data over different spans of time, and to accurately analyze a given frequency band the sample should be long enough to allow several complete oscillations. Table 1 of this review paper summarizes the possible HRV parameters that can be measured using a given time sample. For a complete analysis of all spectral components of HRV, a 24-hour ECG is necessary.

That said, most people studying HRV from a physiological perspective are interested in short-term variability in the LF to HF range.

Those longer ULF and VLF frequencies also tend to dominate the frequency spectrum of 24-hour HRV recordings, obscuring the smaller HF and LF signals. Additionally, the analysis of HF and LF bands over such long sampling intervals isn’t very useful anyway, because the frequencies associated with various physiological processes aren’t perfectly stable across different body positions and activities.

The practical upshot of all this is that the vast majority of HRV spectral analysis is done on samples of 5 minutes or less, and HF and LF bands are normalized to the total signal after subtracting the influence of frequencies below the LF range. (In other words, HF and LF are normalized to the sum of HF + LF.)

Spectral analysis is far more conducive to physiological interpretation than time-domain analysis, the idea being that each of the frequency ingredients represents a different physiological process that can be empirically identified.

Nonlinear Measurements

If you’re slightly math-ed out from the previous two sections, don’t worry – we’re going to take a break from math in favor of an entertaining glimpse into the nature of clinical and academic research.

Almost two decades after the initial 1996 task force report on HRV metrics, a new report was published in 2015. Somewhat shockingly, it did not provide any updates or amendments to their previous summaries and recommendations, despite such updates and advancements being sorely needed.

Instead, the report reviewed several new-fangled, mathematically-intensive HRV metrics collectively known as “nonlinear measurements;” these include fractal analysis, entropy, nonlinear dynamical systems, and chaotic behavior. A selection of these is listed in Table 3 of this paper.

Recounting this progress in HRV research, the report paints a picture of cerebral mathematicians and signal processing engineers on one side, excitedly feeding HRV signals through ever more cutting-edge mathematical models, and doctors and clinical researchers on the other side, plodding forward with the outdated standards from 1996 because no practical advances have been made in the field since then.

In the words of the task force:

Available data indicate that these novel methods have provided more information on the complexity and mathematical or physical characteristics of the variability signal than on sympathetic or parasympathetic neural control mechanisms, even if they were originally developed to clarify the physiological correlates of HRV…Signal processing HRV techniques that do not reflect unmet clinical needs are of little value irrespective of the depth of their mathematical apparatus.
Advances in heart rate variability signal analysis: joint position statement by the e-Cardiology ESC Working Group and the European Heart Rhythm Association co-endorsed by the Asia Pacific Heart Rhythm Society. (2015)

Perhaps I should’ve added a third category to the “phenomenological vs physical” analysis framework – that of treating HRV as a fun mathematical puzzle with no relationship to biology at all.

Mathematical and Technical Challenges

Without getting too far into the weeds, I’d like to give you a taste of some of the significant limitations and challenges related to quantifying HRV. Keep in mind that these challenges are primarily relevant to clinical and especially research settings, where accurate and consistent quantification is important for experimental integrity.

First of all, you’ll recall that HRV analysis uses NN rather than RR intervals. This requires careful selection of ECG segments, and often manual “cleaning” of ECG data to remove irregular beats.

Once you have your NN intervals, you still can’t perform any analysis yet, because all the common statistical calculations and mathematical models require data points to be equally spaced in time. But since heartbeats are different lengths (thwarted by the very phenomenon we’re trying to study!), some level of fancy re-sampling that I don’t fully understand must take place to derive an evenly-spaced dataset.

For spectral analysis, that evenly-spaced dataset can be presented in a number of different styles of tachogram, which ought to be thoughtfully selected based on experimental goals. This source demonstrates the considerations surrounding which type of tachogram to use.

Once a clean and appropriate dataset is acquired, a mind-boggling array of math can be applied to it in an attempt to glean meaningful information. Truly, I’ve been amazed as I’ve read article after article that I simply do not have the mathematical chops to understand. But as the above-mentioned 2015 report points out, fancy math can only get us so far.

Let’s now finally take a look at what our different HRV metrics can tell us, and what controversies and gaps in the field still exist.

Interpreting Heart Rate Variability

Because the frequency-domain metrics have better-defined physiological correlates, we’ll start there and work backwards. (I didn’t switch up the order just to confuse you, promise!)

HF: The Respiratory Band

HF is the best-characterized and least controversial of the frequency-domain metrics. The HF band is conventionally defined as 0.15 Hz to 0.4 Hz, which equates to periods between 2.5 and 7 seconds. Most researchers agree that HF is almost purely vagally mediated – the 0.15 Hz cutoff between HF and LF is understood to effectively separate pure vagal influence from combined influence because sympathetic signaling is too slow to act at higher frequencies. A great deal of evidence involving vagotomies and autonomic blockades corroborates this.

Remember the first major contributor to HRV that we discussed – RSA? Well, HF is sometimes called the “respiratory band” because spontaneous respiration frequencies usually fall within this range. Thus, short-term spectral analyses for HRV (excluding VLF and ULF frequencies, as discussed earlier) typically reveal a major peak in the HF band that corresponds to RSA.

However, it’s very important to understand that the HF band does not magically capture RSA. RSA will only be found within the HF band if a person is breathing at that frequency (i.e. at least one inhale every 7 seconds). If someone is breathing more slowly than that, RSA will be found not in the HF band, but the LF band!

In practice, HF is used frequently as an index of cardiac vagal modulation, although as you’ll see below, it probably doesn’t capture the entirety of vagal influence. HF is also often conflated with RSA, although this is not good practice; one should confirm the respiratory rate to ensure it actually falls with the HF band.

LF: Mayer Waves and the Baroreflex

The physiological interpretation of LF has been fraught with controversy. For a long time, the commonly held belief was that while HF indexes vagal influence on the heart, LF indexes sympathetic influence. And wouldn’t this be oh-so-convenient? Part of me wonders whether wishful thinking contributed in some part to the persistence of this idea.

The belief that LF indexes cardiac sympathetic influence came about primarily from tilt-table testing, where researchers observed that as subjects were tilted from horizontal to vertical – a maneuver that we know increases sympathetic activation – LF power increased relative to HF power. The belief was bolstered by further observations where an increase in relative LF power occurred in concert with other situations or activities known to increase sympathetic activation, such as standing, mental stress, and moderate exercise.

However, researcher David Goldstein deftly countered all of this “evidence” in his 2011 review Low-frequency power of heart rate variability is not a measure of cardiac sympathetic tone but may be a measure of modulation of cardiac autonomic outflows by baroreflexes. (Great title, right?)

Recall from our discussion that Mayer waves, oscillating at 0.1 Hz in humans, elicit coherent oscillations in heart rate. What frequency band does 0.1 Hz fall into? The LF band! And indeed, in addition to the respiratory peak usually seen in the HF range in short-term spectral analyses, another main peak is typically observed in the LF range – right around 0.1 Hz.

Goldstein argues succinctly and convincingly that this LF peak (and therefore most LF power) simply represents baroreflex function. (What’s funny is that researchers were arguing this back in 1997 and even earlier, yet apparently the controversy was still alive and well in 2011.)

As I explained in the baroreflex section, there doesn’t appear to be consensus over the extent to which baroreflex heart rate changes are vagally or sympathetically mediated, but either way, it’s safe to say that LF is far from being a measure of purely sympathetic cardiac activity.

VLF and ULF: Circadian and Other Rhythms

As mentioned earlier, VLF and ULF dominate 24-hour HRV recordings, representing a good 95% of total power. However, their physiological mediators are understood poorly or not at all.

In addition to the two main peaks corresponding to RSA and the Mayer wave baroreflex, spectral analyses that include VLF frequencies will sometimes show an additional small peak around 0.03 Hz, which is thought to reflect fluctuations of slower humoral influences on the heart. (Source) Cycles related to thermoregulation and fluid balance (via the renin-angiotensin system) may also contribute, and we have several bits of experimental evidence suggesting that some oscillations in the VLF band may be generated by the heart’s intrinsic nervous system. (Source)

Although the ULF band captures any oscillations with a period of greater than 5 minutes, most of its power comes from the significant circadian oscillations in heart rate, which occur on the scale of hours rather than minutes.

Ultimately, interest in these two metrics in the field of physiology has been mild at best, because these longer cycles don’t appear to tell us much about the near-instantaneous sympathetic and parasympathetic modulation that everyone is so interested in. Arguably, the very concept of “heart rate variability” – as far as it is characterized by the idea of “beat-to-beat” variability – scarcely even applies to lower frequencies.

Despite this, both metrics still have excellent utility as predictive or correlative variables in clinical settings. Both the ULF and VLF bands have stronger associations with all-cause mortality and other adverse outcomes than the LF or HF bands, making them useful for risk stratification and other predictive measures. (Source)

SDNN and SDANN: Overall and Long-Term Variability

Now, we turn back to our time-domain metrics! The most straightforward of all HRV measurements is SDNN – the standard deviation of all the NN intervals in a given sample. As with any measure of standard deviation, it’s a good way of determining how spread out your data is from the mean, and SDNN measured over 24 hours is often used to quantify overall HRV. SDNN essentially supplies the same information that “total power” does in spectral analysis – a measure of total variance without any attempt at parsing out its component parts.

Unsurprisingly, SDNN correlates strongly with ULF and VLF, since those two frequency bands make up the lion’s share of total power. As such, SDNN also has very good predictive power as a variable in clinical settings (it’s considered the gold standard for cardiac risk stratification), but is not useful for any sort of physiological analysis.

SDANN, by averaging together the NN intervals within each 5-minute chunk of time prior to finding the standard deviation, essentially “smooths out” short-term (<5 min) HRV. Therefore, SDANN essentially eliminates the influence of RSA and the baroreflex, again providing a picture of frequencies in the VLF and ULF ranges.

RMSSD and pNN50: Short-Term Variability

On the other hand, RMSSD quantifies the short-term variation in heart rate. Because the calculation involves finding the differences between successive RR intervals, it captures beat-to-beat variation, which – as discussed previously – can only be vagally mediated due to the time scales involved.

Several other time-domain metrics provide similar information and are highly correlated with RMSSD, including pNN50. These metrics are also both highly correlated with HF, which makes sense, since all of these metrics quantify the shortest-term (i.e. highest-frequency) modulations in heart rate.

Thus, like HF, RMSSD is often used to quantify vagal influence on the heart. (As mentioned earlier, this is the metric that Oura uses, and likely other wearables as well.)

Measuring Respiratory Sinus Arrhythmia (RSA)

As the oldest and most well-studied contributor to HRV, I thought RSA deserved its own little section here about how we measure it. This is especially relevant since I’ve frequently observed authors refer to RSA in an experiment without saying how it was measured, and I’ve seen a fair bit of terminology conflation between HF, RSA, and even sometimes HRV itself.

This 2012 paper does a nice job reviewing three different commonly-used ways of quantifying HRV:

HF
Peak-to-trough
The Porges-Bohrer method (RSA_P-B)

You’re already familiar with HF and why it can be a good measure of RSA under the right circumstances.

Peak-to-trough finds the difference between the shortest RR interval during inhalation and the longest RR interval during exhalation, which is a fairly intuitive way to measure RSA.

Finally, the Porges-Bohrer method basically does a lot of data cleaning and statistics to precisely extract just the respiratory frequency. So, similar to HF analysis, but rather than being bounded by a set frequency range, the method targets the specific respiratory frequency in any given sample.

All their experiments and validation procedures found the third method to be the most accurate. That said, in practice, all three metrics highly correlate, so can often be used interchangeably if the correct conditions are met. [Edit Aug 10, 2022: Researcher Paul Grossman, in his extended critique of Porges’ polyvagal theory, has pointed out that the Porges method has “misestimation problems,” citing this source. In any case, it’s unsurprising that Porges would be biased in favor of his own method in his own paper.]

All this to say – if you’re reading something that cites RSA values, remember that not all RSA metrics are created equal, particularly if HF is used without confirming respiratory rate.

Controversies and Drama in HRV Research

Having now spent the better part of two months immersed in HRV papers, I can say it’s one of my favorite topics I’ve written on. There are a lot of reasons for this, many of which I mentioned in the introduction; but in addition to the topic itself being fascinating, the field is just an absolute saga of controversy and drama. And when you’re deep in the trenches trying to understand spectral analysis, a little spice goes a long way to keep things interesting.

Now that I’ve (finally [I think]) gotten my arms around the big themes in the field, I wanted to share the big ones here. It would be disingenuous to present all the preceding information without acknowledgement of the many controversies that are still alive and well, and I think understanding the context in which research questions are being asked and answered is super helpful to understanding the topic itself.

Is HRV Hopelessly Confounded by Heart Rate?

In 2019, the Journal of Physiology hosted a written proponent-opponent “CrossTalk” debate about whether HRV is a useful metric in the assessment of autonomic nervous system responsiveness, generating two initial proponent/opponent papers, a rebuttal from each side, and 17 write-in comments from other researchers giving their two cents.

The opposition paper, which incidentally did not address the debate prompt (more on this in the next section), argued that HRV is simply a function of heart rate, and is therefore no more than a complicated and unnecessary way to glean the same information we could already glean from heart rate itself.

This stance was met with a flood of opposing comments citing study after study proving otherwise. For instance, both Cooke et al. (1998) and Brown et al. (1993) showed that mean heart rate was comparable across a wide range of breathing frequencies, whereas HRV differed significantly, tending to increase substantially with slower respiration.

The comments cite many other experiments besides, and also point out that post-exercise, heart rate returns to baseline within minutes, while HRV doesn’t return to baseline for hours or even days.

So we can all breathe a collective sigh of relief that we haven’t just wasted minutes (in your case), weeks (in my case), or years (in the case of these researchers) on a totally superfluous metric.

Nevertheless, heart rate IS intimately connected with HRV, and this is often not acknowledged sufficiently in the research, so the issue deserves some discussion.

Generally, increases in heart rate will lead to decreases in HRV (and vice versa). One very obvious reason is that shorter RR intervals (i.e. faster heart rate) essentially allow less time for variability to occur between beats.

This is partially a mathematical issue, since the confounding is much weaker when using datasets composed of instantaneous heart rate rather than RR interval, but basically all of the common HRV metrics are calculated using RR intervals and are therefore subject to this mathematical confounding. This includes all time-domain metrics, as well as non-normalized spectral components.

But there are also very real physiological reasons for the heart rate/HRV relationship involving the precise timing of autonomic nerve impulses in a given cardiac cycle. The first section of the opposition paper mentioned above explains this. This editorial also does a good job reviewing this and other relevant considerations, and also reviews the HRV metrics that are not confounded by heart rate.

Normalized HF and LF are independent of heart rate because they’ve been normalized to total power (and total power is what decreases with increasing heart rate), as are the relatively impractical nonlinear measurements I mentioned briefly. There are also methods by which other metrics can be corrected for heart rate. However, this is not routinely done, so these researchers do bring up a topic that deserves attention in the field.

Nevertheless, much of their argument seems to suffer from treating HRV as a solely phenomenological variable, rather than a physiological process. Here’s a sentence from their conclusion:

It is not even known whether autonomic nerve activity affects HRV – previously, this will have been tested by investigating the effect of autonomic blockade on HRV, but of course autonomic blockade will affect the heart rate and therefore HRV.
Boyett et al. CrossTalk opposing view: Heart rate variability as a measure of cardiac autonomic responsiveness is fundamentally flawed. (2019)

This is an excellent example of the types of faulty conclusions that can come about when one loses sight of the physiological underpinnings of HRV. It’s absolutely true that autonomic blockade affects both heart rate and HRV – but that’s because autonomic nerve activity determines both heart rate AND the variation in said heart rate. HRV is not a mathematical construct – it is a physiological phenomenon that is created by the very same autonomic pathways that dictate heart rate.

But like I said, the issue of heart rate confounding is absolutely still relevant to interpretation of HRV metrics. So what’s the practical upshot here?

Of the metrics we’ve discussed, SDNN, SDANN, RMSSD, pNN50, and total power in any of the frequency bands (HF, LF, VLF, ULF) are intimately tied to mean heart rate, so a chunk of any observed differences in these HRV measurements (whether between subjects or between groups or between two different timepoints for the same person) might be explained away by differences in heart rate. So if one wants to draw any conclusions about HRV using any of those metrics, one must experimentally or mathematically account for heart rate.

Alternatively, one could use normalized HF and LF as their HRV metrics.

What about the interpretation of HRV measurements from wearables? Is Oura’s RMSSD readout confounded by heart rate? Well – yes. But since they also measure resting heart rate overnight, you can fairly easily look at both your HRV and heart rate trends and get a sense for whether changes in your HRV appear to track closely with changes in resting heart rate, or whether HRV is actually giving you unique information.

Does HRV Index Autonomic Tone?

This is a big one, and I struggled with where to place this section in the article because while an understanding of this issue colors the understanding of basically every aspect of HRV, a basic understanding of HRV is also necessary in order to understand this issue. So I’m trusting you, dear reader, to retroactively apply this information to everything you’ve learned thus far.

Like the rest of humanity, researchers often find themselves mired in issues of semantics. One such issue that cropped up early in HRV research is the distinction between tone and fluctuations.

It is the position of most of the experts in the field that HRV is unequivocally a measure of responsiveness, not tone, except in specific circumstances. Yet still, statements equating HRV metrics with “autonomic tone” continue to proliferate in research journals and popular websites alike. Indeed, the CrossTalk debate discussed above was plagued by this issue. I’ll let Malik et al, the “proponents,” speak for themselves:

The link of heart rate variability (HRV) to the autonomic nervous system (ANS) encompasses different facets. When this journal invited the present proponent–opponent debate, both sides agreed to discuss the usefulness of HRV in the assessment of ANS responsiveness rather than in the assessment of ANS tone. The difference is obvious to all even marginally involved in ANS studies. While assessing ANS tone means estimating the sympatho/vagal nerve traffic, the assessment of ANS responsiveness means estimating ANS responses to provocations. (The difference is as the difference between assessing water waves due to wind and measuring the depth of a lake.)

Since The Journal confirmed that our opponents agreed to address ANS responsiveness and not the ANS tone, we read the contribution by Boyett et al. (2019) with some bewilderment. Apart from the title, they do not mention responsiveness at all and write solely about ANS tone assessment. Is this because they agree with us and, similar to ourselves, cannot find any arguments against the use of HRV in the estimates of ANS reactions?
Malik et al. Rebuttal from Marek Malik, Katerina Hnatkova, Heikki V. Huikuri, Federico Lombardi, Georg Schmidt and Markus Zabel. (2019) [emphasis mine]

Shade thrown. Their whole response was pretty sassy, in fact (here’s another quote: “can our learned colleagues cite any article in a reasonable physiology/cardiology journal that interpreted SDNN values as those of sympathetic or of vagal tone?”)

However, things are not quite so simple as the lake analogy makes it seem, so although the consensus is indeed that HRV indexes variation rather than tone, I want to add a little nuance here.

For instance, we know that RSA is the rhythmic fluctuation of heart rate in response to respiration. On the face of it, this seems to be a clear measurement of variation in vagal traffic, not underlying vagal tone. But recall that in the case of RSA, respiration acts as an autonomic “gate” of sorts, where inhalation temporarily restricts vagal outflow, which resumes during exhalation (review here).

Using this analogy of a gate rather than waves, it becomes apparent that the absolute amount of vagal traffic (i.e. tone) may have an effect on the magnitude of the variations observed when that traffic is stopped and then started again. From a 1997 review by Berntson et al, “RSA represents the phasic inspiratory inhibition of vagal outflow, and hence its magnitude is correlated with mean vagal control of the heart.”

That said, they were quick to clarify that this relationship is only relevant within a given subject, and not for between-subject comparisons. They also note that inhalation doesn’t necessarily block vagal traffic completely, so by no means can RSA be used as a sensitive measure of vagal tone. This review also discusses the caveats to the association between RSA and cardiac vagal tone, and includes the below figure to illustrate how the same vagal tone could produce different RSA amplitudes at different breathing frequencies.

“The first three RSA cycles occur at a respiration period that is twice as long as the last three cycles. Note that the same quantity of vagal bursts occur per time unit (i.e. six bursts during one slow breath; six [bursts] during two rapid breaths), indicating the same mean level of vagal discharge per minute, but a different pattern of vagal discharge and impact on RSA at different respiration rates.” Grossman & Taylor. Toward understanding respiratory sinus arrhythmia: relations to cardiac vagal tone, evolution and biobehavioral functions. 2007.

Another review from 1997 (by Dwain Eckberg, who was a contributing author to the previous paper) discusses several additional situations where RSA and cardiac vagal tone were not correlated. For instance, the Brown et al. study I mentioned earlier found huge differences in HF (used as a proxy for RSA) across different breathing frequencies but no change in mean heart rate, and another found no change in HF in response to changes in arterial pressure. Eckberg concludes: “Thus, moderate changes of arterial pressure, which alter vagal-cardiac nerve activity, do not change HF RR-interval fluctuations, and changes of breathing frequency and depth, which profoundly alter HF RR-interval fluctuations, may not change vagal-cardiac nerve activity at all.”

So despite the persistence of the claim (both in research and mainstream) that HRV tells us something about sympathetic or vagal tone, that is a mischaracterization of what HRV actually measures. That’s not to say that higher HRV measurements don’t often correlate with higher vagal tone, but to claim that one actually measures the other is simply incorrect.

Does the LF/HF Ratio Index Sympathovagal Balance?

The LF/HF ratio is the magnum opus of sorts of the HRV controversies, incorporating several layers and levels of misunderstanding in order to achieve the ultimate misinterpretation. Adopting the erroneous notion that HRV measures autonomic tone at all, and that LF measures sympathetic tone specifically, the LF/HF ratio was born, widely used to index the ratio of sympathetic to parasympathetic tone; or, sympathovagal balance. And despite researchers explicitly dispelling this notion way back in 1993, the LF/HF ratio still makes its appearances to this day (for instance, in this paper using the LF/HF ratio to compare sympathovagal balance in ulcerative colitis patients to healthy controls).

Eckberg tackles this issue in great detail in his 1997 paper Sympathovagal Balance: A Critical Appraisal, systematically examining each of the premises necessary for the construction of the LF/HF ratio. I love how he introduces the topic:

This review, which examines the physiological foundations for the concept of sympathovagal balance, was undertaken because of concern that this new measure has entered the mainstream of cardiovascular research without having been critically vetted.
Dwain Eckberg. Sympathovagal Balance: A Critical Appraisal. (1997)

Oof.

Then later, regarding the “tone vs variability” semantics issue:

These concerns about language reflect a larger concern that because the words are so familiar that the concepts engendered by these words may also be regarded as familiar. As this review suggests, I find many of the concepts foreign. In the end, however, it is not philosophy or language that must justify sympathovagal balance; it is physiology, which I have reviewed here.
Dwain Eckberg. Sympathovagal Balance: A Critical Appraisal. (1997)

This 1999 paper from Jeffery Goldberger provides a very cogent summary of the debate surrounding measures of sympathovagal balance, and specifically using LF/HF ratio or other HRV parameters as such a measure. It then goes on to attempt to identify the best, most accurate index of cardiac sympathovagal balance, according to the following criteria: “the response of the index to various autonomic conditions should reflect the underlying changes in the physiological state and have a meaningful interpretation. Specifically, the index should provide information on the net effects of the sympathetic and parasympathetic nervous systems.” (Like…obviously, right?)

Hilariously, after a great deal of sincere and thoughtful analysis, he found that the index that best fits these criteria is basically…heart rate. Technically something they termed vagal-sympathetic effect (VSE) is the ideal measurement, and is defined as the ratio of average RR interval (heart period) to intrinsic heart period (determined under total autonomic blockade). Intrinsic heart period varies based on physical fitness, levels of thyroid hormone, and other factors, but for the same person over a period of time where no significant changes in fitness or thyroid status have occurred, one can get a decent idea about changes in sympathetic/parasympathetic balance from changes in resting heart rate.

I gotta tell you, reading this paper was a real head-smack moment for me – if you think about it, the conclusions are so obvious! Sympathovagal balance is a relative measure, a summed effect measure, a net. We know that sympathetic impulses increase heart rate, and vagal impulses decrease heart rate. So OBVIOUSLY heart rate is the best index for sympathovagal balance. In all our eager searching for a sophisticated metric, the answer was right there in front of us the whole time.

And so, have we now come full circle, back to foregoing HRV in favor of lowly heart rate? Not quite. Here’s what Goldberger concludes:

It is possible that cardiac autonomic evaluation needs to occur in a continuum with the R-R interval serving as a global measure of sympathovagal balance, heart rate variability as a measure of autonomic modulation by respiratory and vasomotor activity, and baroreflex sensitivity as a measure of autonomic responsiveness.
Jeffrey Goldberger. Sympathovagal balance: how should we measure it? (1999)

Now, lest you think this issue was laid to rest in 1999, here’s a series of three letters to the editor from 2012, through which you may continue to follow the thread of this debate if you wish:

But what I can tell you from reading dozens of these back-and-forths is that often, everyone just ends up roughly where we were back in the 90s.

Summary and Conclusions

We’ve covered a lot of ground in this article (to say the least). So here are the main takeaways (in bullet form, since you’ve probably had enough of paragraphs by now):

The term heart rate variability (HRV) describes the fact that in a healthy human heart, the amount of time between subsequent heart beats is not constant; i.e. the heart is not a metronome
The phenomenon of HRV is created mostly by fluctuations in nerve traffic, primarily from parasympathetic (vagal) neurons but also, to a lesser extent, sympathetic neurons
Contrary to popular belief, HRV does not measure autonomic tone
There are a number of different ways to quantify HRV, including time-domain metrics derived using statistical calculations (SDNN, RMSSD) and frequency-domain metrics derived by breaking down the HRV signal into its various frequency components (HF, LF)
The two main physiological sources of rhythmic nerve traffic fluctuation to the heart are respiratory sinus arrhythmia (RSA) and the baroreflex associated with Mayer waves, and these influences are roughly reflected in the HF and LF HRV measurements, respectively
There are a great number of mathematical and technical challenges and considerations when choosing and interpreting HRV metrics, with some metrics being useful primarily as statistical correlates and predictors in a given context, while others (specifically HF and LF) appear to provide a window of sorts into the physiological activities of the autonomic nervous system
The conflation of these two paradigms (specifically, inferring precise autonomic activities from observed correlations) has led to a great deal of debate and controversy over the correct interpretation of HRV metrics
Heart rate is a much better measurement of “sympathovagal balance” than HRV, because mean heart rate is the final output of combined sympathetic (which increases heart rate) and parasympathetic (which decreases heart rate) influence on the heart
Heart rate also gives a better idea of autonomic tone, while HRV measures autonomic fluctuation
HRV is confounded by heart rate, naturally increasing with decreasing heart rate, so researchers must (but often don’t) tease apart these effects for accurate characterization of HRV as a variable
For practical purposes, higher HRV values are generally associated with better health across the board, whether that effect is primarily from lower heart rate (indicating greater vagal influence on the heart) or from greater autonomic modulation

To sum up, HRV has become an enormously popular metric in clinical research for its broad association with different states of health and disease and its ability to predict disease outcomes. It has also grown in popularity as an at-home metric for athletes and other health-conscious people, because as in clinical research, it has significant practical value as a generalized health metric.

However, a detailed and accurate understanding of the physiological mechanisms behind these observed phenomena continues to lag significantly behind the adoption of these metrics in research. As such, a great deal of discretion is warranted when attempting to infer physiological processes like autonomic tone or function from HRV metrics.

The field of HRV research is fascinating to me. It sits at the intersection of math, theoretical physics, and biology, inviting discussions of power spectral density and chaos theory alongside discussions of autonomic physiology and cardiovascular mortality. Within biology, it sits at the intersection of physiology and psychology, inviting discussions of cognitive function and emotional regulation alongside discussions of athletic training and blood pressure control.

We can only hope that current and future researchers in the field heed the counsel of their forbears, enabling advances in understanding that give us true glimpses into the regulation of our autonomic systems, providing us with new avenues for exploring and bringing together all the facets of human health: physical, mental, emotional, and social.