5  HRV Metrics: Time Domain, Frequency Domain, Nonlinear

There are dozens of HRV metrics, but many are strongly correlated with one another. This chapter organizes what the commonly used metrics compute, and what they are thought to reflect physiologically.

Below, the series of NN intervals is denoted x_1, x_2, \dots, x_N (in ms).

5.1 Time-domain metrics

5.1.1 SDNN: the size of the overall fluctuation

\mathrm{SDNN} = \sqrt{\frac{1}{N-1}\sum_{i=1}^{N}(x_i - \bar{x})^2}

  • The standard deviation of NN intervals
  • Includes fluctuations of all frequency components within the recording (equivalent to the square root of total power)
  • Strongly dependent on recording length. A 24-hour SDNN and a 5-minute SDNN are different things
  • 24-hour SDNN is one of the best-validated metrics for prognosis after myocardial infarction (Chapter 7)

5.1.2 RMSSD: differences between adjacent beats

\mathrm{RMSSD} = \sqrt{\frac{1}{N-1}\sum_{i=1}^{N-1}(x_{i+1} - x_i)^2}

  • The root mean square of differences between adjacent NN intervals
  • Captures beat-to-beat fast changes → mainly reflects regulation by the vagus nerve. Strongly correlated with HF power
  • Less affected by slow trends, and less affected by respiratory rate than HF
  • Relatively stable even in short recordings → the most widely used metric in wearables and sports. Because its distribution is right-skewed, the natural log, lnRMSSD, is often used

5.1.3 pNN50

  • The percentage of differences between adjacent NN intervals exceeding 50 ms (%)
  • Strongly correlated with RMSSD and likewise a vagal index. However, in people with fast heart rates and small HRV it is nearly 0%, making differences hard to detect

5.1.4 Others

  • SDANN: divide a 24-hour recording into 5-minute segments and take the standard deviation of the mean NN interval of each segment (slow fluctuations)
  • HRV triangular index: the total number of NN intervals in the histogram divided by the count of the modal bin. Robust to abnormal beats

5.2 Frequency-domain metrics

The time series of NN intervals is converted to a power spectral density (PSD), and the power (area, ms²) of each band is computed.

Figure 5.1: Schematic of a resting HRV power spectrum. The peak due to respiration (0.25 Hz) appears in the HF band and the peak due to the baroreflex (about 0.1 Hz) appears in the LF band.
Metric Band Physiological interpretation (current mainstream understanding)
HF 0.15–0.4 Hz An index of cardiac vagal activity (respiratory sinus arrhythmia). However, strongly affected by respiratory rate and tidal volume
LF 0.04–0.15 Hz Reflects both vagal and sympathetic nerves + baroreflex function. Not an index of sympathetic activity
VLF 0.0033–0.04 Hz Much about its origin is unknown. Low reliability in short recordings
Total power ≤ 0.4 Hz The overall fluctuation. Equivalent to SDNN²
LF/HF — Widely used as “sympathetic–parasympathetic balance”, but its physiological basis is weak (see below)
CautionLF/HF ratio = “sympathetic/parasympathetic balance” does not hold

In the 1980s–90s, the interpretation spread that LF reflects sympathetic and HF parasympathetic activity, and that their ratio LF/HF represents “sympathetic–parasympathetic balance”. Today, this interpretation is not supported, for the following reasons (Billman 2013, etc.).

  1. LF has a large vagal contribution: atropine (vagal blockade) also greatly reduces LF (Chapter 2)
  2. LF does not correlate with cardiac sympathetic activity: LF power hardly correlates with direct measures such as noradrenaline release into the heart (spillover). One view is that LF rather reflects baroreflex function (Goldstein et al. 2011)
  3. The two nervous systems are not in a simple tug-of-war: there are many situations in which both rise or fall at the same time (exercise, emotion, cold stimulation, etc.)
  4. Effect of respiration: when breathing slows, the respiratory fluctuation moves from HF to LF, and LF/HF rises greatly even though the autonomic nerves have not changed (checked in Chapter 8)

If you must use LF/HF, it is safest to record and control the respiratory rate and to avoid the word “sympathetic” in your description.

5.2.1 Absolute values, normalized units, and logarithms

  • Absolute values (ms²): inter-individual differences are very large and the distribution is right-skewed → statistical analysis is often done after taking the natural log (ln)
  • Normalized units (n.u.): LF n.u. = LF / (total power − VLF) × 100. Because LF n.u. and HF n.u. sum to 100, they carry only the same information as LF/HF mathematically

5.3 Nonlinear metrics

5.3.1 Poincaré plot

A scatter plot with x_i on the horizontal axis and x_{i+1} on the vertical axis. It lets you see the nature of the fluctuations intuitively.

flowchart LR
  P["Poincaré plot<br/>(x_i, x_i+1)"] --> SD1["SD1: spread perpendicular to the diagonal<br/>= fast beat-to-beat fluctuation<br/>SD1 = RMSSD / √2 (approx.)"]
  P --> SD2["SD2: spread along the diagonal<br/>= slow fluctuation<br/>SD2² ≈ 2·SDNN² − SD1²"]

  • SD1 carries almost the same information as RMSSD (mathematically almost \mathrm{SD1} \approx \mathrm{RMSSD}/\sqrt{2})
  • SD2 is the fluctuation on longer time scales
  • From the pattern of the point cloud’s shape (comet, torpedo, fan, etc.), ectopic beats and arrhythmias are easy to spot visually

5.3.2 Other nonlinear metrics

  • DFA α1 (short-term scaling exponent of detrended fluctuation analysis): the “correlation structure” at scales of 4–16 beats. About 1 at rest, decreasing as exercise intensity rises. Its application to estimating the aerobic threshold during exercise (α1 ≈ 0.75) is being studied
  • Sample entropy (SampEn): the “unpredictability/complexity” of a series. Applied to early detection of neonatal sepsis, among others (Chapter 7)

5.4 Reference values

Values from resting 5-minute recordings in healthy adults have been reported, for example, as follows (approximate mean ± SD from the systematic review by Nunan et al. 2010).

Metric Value
Mean NN interval about 930 ± 90 ms
SDNN about 50 ± 16 ms
RMSSD about 42 ± 15 ms
LF about 520 ± 290 ms²
HF about 660 ± 780 ms²
ImportantHow to use reference values
  • Inter-individual differences (age, sex, fitness, genetics) are very large, so they are not suited to judging an individual by a “normal range”
  • Values change with the measuring device, recording length, preprocessing, and spectral estimation method, so do not directly compare values from different studies or devices
  • In healthcare, it is more meaningful to look at changes in the same person under the same conditions (deviation from the individual’s own baseline) (Chapter 7)

5.5 Summary of the relationships between metrics

flowchart TB
  subgraph Fast["Fast fluctuations (mainly vagal)"]
    RMSSD --- pNN50
    RMSSD --- HF
    RMSSD --- SD1
  end
  subgraph All["Overall fluctuation"]
    SDNN --- TP["Total power"]
    SDNN --- SD2
  end
  subgraph Mixed["Mixed / hard to interpret"]
    LF
    LFHF["LF/HF"]
    VLF
  end

5.6 Summary

  • RMSSD / pNN50 / HF / SD1 are a group of strongly correlated metrics that mainly reflect cardiac vagal activity
  • SDNN / total power are the overall fluctuation and depend on recording length
  • LF is not an index of sympathetic activity. Avoid reading LF/HF as “autonomic balance”
  • Reference values have a very wide range; looking at within-individual change is more useful