Skip to content

Signal Quality Index (SQI)

Key: sqi · Family: morphology · Better: higher ↑ · Per-lead: yes

SQI asks: do the synthetic ECGs have the typical shape of the real ECGs? It builds a template, the average real recording, and measures how well each synthetic recording correlates with it.

Try it

The bold line is the template from 25 real beats. Thin lines are synthetic beats. Every beat is first scaled to the range 0–1, so only shape matters, not amplitude.

Interactive demo loading… (needs JavaScript)

What to notice:

  • Wide QRS and inverted T lower SQI because the shape differs from the template.
  • Misaligned beats: SQI also drops when beats are not at the same position, because correlation is computed sample by sample. On whole recordings with different heart rates the real template itself becomes blurred.
  • The real self-SQI is the ceiling: real beats are not perfectly correlated with their own average either.

How it works

For each lead:

  1. Scale every real recording to unit range and average them: this is the template.
  2. Scale every synthetic recording to unit range.
  3. Compute the Pearson correlation of each synthetic recording with the template.
  4. Average. extra["real_self_sqi"] gives the same average for the real recordings.

The math

\[ \tilde x = \frac{x - \min x}{\max x - \min x}, \qquad \mathcal T = \frac{1}{n}\sum_{i=1}^{n} \tilde x_i^{\,r}, \qquad \mathrm{SQI} = \frac{1}{m}\sum_{j=1}^{m} \rho\!\left(\tilde x_j^{\,s},\ \mathcal T\right) \]

with \(\rho\) the Pearson correlation.

Reading the score

  • Range [−1, 1]; compare with real_self_sqi. A synthetic SQI close to the real self-SQI means typical morphology.
  • SQI rewards resemblance to the average ECG. A generator that outputs the mean beat every time scores very well while having no diversity, so pair SQI with Precision & Recall or a distribution metric.
  • Works best on aligned, single-beat or fixed-rhythm data. For multi-beat recordings with varying heart rates the template is blurred, and SQI mostly reflects rhythm alignment.

Use it with ECGEN-Eval

ecgen-eval eval --real real.npy --synthetic synth.npy -m sqi --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)

result = AVAILABLE_METRICS["sqi"]().compute(real.data, synth.data, lead_names=real.lead_names)
print("SQI:", result.score, "real self-SQI:", result.extra["real_self_sqi"])

Parameters are in the Metric Reference.