Skip to content

Heart-Rate Variability (HRV)

Key: hrv · Family: physiological · Better: lower ↓ · Per-lead: yes · Uses: fs

A healthy heart does not beat like a metronome. Intervals between beats lengthen and shorten with breathing and nervous-system activity. hrv asks: do synthetic ECGs show the same amount and kind of beat-to-beat variation as real ECGs?

Try it

Top: the tachogram (RR interval of each beat over time) of one real and one synthetic recording. Bottom: the distributions of three HRV statistics over 200 recordings each.

Interactive demo loading… (needs JavaScript)

What to notice:

  • Metronome: a generator that copies one beat at a fixed interval has almost zero SDNN and RMSSD, which is easy to spot here even when the heart rate is right.
  • Erratic: random beat-to-beat jumps inflate RMSSD and pNN50 much more than SDNN.
  • 2-minute strips: with longer recordings the histograms become narrower and the statistics more reliable. On 10-second strips (about 12 beats) HRV statistics are very noisy.

How it works

For each lead:

  1. Detect R peaks with the slope-energy QRS detector described on the Heart rate page, and compute the RR intervals.
  2. Per recording, compute SDNN, RMSSD and pNN50.
  3. For each statistic, compare the real and synthetic distributions with the Jensen–Shannon divergence (50-bin histograms).
  4. The lead's score is the mean of the three divergences; the overall score is the mean over leads.

The math

With RR intervals \(RR_1..RR_N\) in ms:

\[ \mathrm{SDNN} = \sqrt{\frac{1}{N-1}\sum_{i=1}^{N}(RR_i - \overline{RR})^2}, \qquad \mathrm{RMSSD} = \sqrt{\frac{1}{N-1}\sum_{i=2}^{N}(RR_i - RR_{i-1})^2}, \qquad \mathrm{pNN50} = \frac{\#\{i : |RR_i - RR_{i-1}| > 50\,\mathrm{ms}\}}{N-1} \]
\[ \text{score} = \tfrac13\left[\mathrm{JS}_{\mathrm{SDNN}} + \mathrm{JS}_{\mathrm{RMSSD}} + \mathrm{JS}_{\mathrm{pNN50}}\right] \]

SDNN captures overall variability; RMSSD and pNN50 capture fast, beat-to-beat changes. ECGEN-Eval reports pNN50 as a fraction (0–1), not a percentage.

Reading the score

  • Range [0, ln 2]. Look at the means in extra (real_sdnn_mean, synth_sdnn_mean, …) to see the direction of the difference: too regular or too erratic.
  • Short-term HRV norms assume at least 5 minutes of recording. On 10 s strips, compare only real and synthetic sets of the same length, and expect noisy scores.
  • Missed or extra R peaks distort HRV heavily; check R-peak detection on noisy data.

Use it with ECGEN-Eval

ecgen-eval eval --real real.npy --synthetic synth.npy -m hrv --fs 500 --output report.html

from ecgen_eval import load_ecg_npy, AVAILABLE_METRICS

real = load_ecg_npy("real.npy", fs=500)
synth = load_ecg_npy("synth.npy", fs=500)

hrv = AVAILABLE_METRICS["hrv"](fs=real.fs)
result = hrv.compute(real.data, synth.data, lead_names=real.lead_names)
print(result.score)
print(result.extra["II"])   # real/synth SDNN, RMSSD, pNN50 means for lead II

Parameters are in the Metric Reference.

Learn more

  • Task Force of the ESC and NASPE, Heart rate variability: standards of measurement, physiological interpretation and clinical use, Circulation 1996.