ECGEN-Eval Report

Synthetic ECG Quality Evaluation — Real: 00000 | Synthetic: 00000

Generated: 2026-03-19 11:19:09 | ecgen-eval v0.1.0

Dataset Summary

Type	Label	Samples	Leads	Timepoints	Fs	Duration	Lead Names
Real	00000	987	12	5000	500.0 Hz	10.0 s	I, II, III, AVR, AVL, AVF, V1, V2, V3, V4, V5, V6
Synthetic	00000	987	12	5000	500.0 Hz	10.0 s	I, II, III, AVR, AVL, AVF, V1, V2, V3, V4, V5, V6

ECG Sample Gallery

Comparison mode: overlay | Calibration pulse: 1 mV × 0.2 s prepended | Paper: 25 mm/s · 10 mm/mV

Sample 0

Sample 1

Sample 2

MMD — Maximum Mean Discrepancy (MMD): measures the distance between the distributions of real and synthetic ECGs in a reproducing kernel Hilbert space. Score of 0 means identical distributions. Lower is better.

Reference: Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). A kernel two-sample test. Journal of Machine Learning Research, 13, 723–773. https://jmlr.org/papers/v13/gretton12a.html

Dataset	Score
00000	0.000000

Lower is better (0 = identical distributions).

DTW — Dynamic Time Warping (DTW): measures morphological shape similarity between real and synthetic ECG waveforms by finding the optimal non-linear alignment along the time axis. Mean pairwise DTW distance over randomly sampled pairs. Lower is better.

Reference: Berndt, D. J., & Clifford, J. (1994). Using dynamic time warping to find patterns in time series. KDD Workshop, 10(16), 359–370. | Goldberger, A. L., et al. (2000). PhysioBank, PhysioToolkit, and PhysioNet. Circulation, 101(23), e215–e220. https://doi.org/10.1161/01.CIR.101.23.e215

Dataset	Score
00000	4.985142

Lower is better (0 = identical distributions).

PRD — Percent Root-mean-square Difference (PRD): measures waveform distortion of the nearest synthetic neighbour for each real ECG. PRD = 100 × ‖real − synth‖₂ / ‖real‖₂ (%). Lower is better.

Reference: Zigel, Y., Cohen, A., & Katsevman, A. (2000). The weighted diagnostic distortion measure for ECG signal compression. IEEE Transactions on Biomedical Engineering, 47(11), 1422–1430. https://doi.org/10.1109/10.871205

Dataset	Score
00000	58.851376

Lower is better (0 = identical distributions).

PSD — PSD Divergence: Jensen–Shannon divergence between the mean normalised power spectral densities of real and synthetic ECGs per lead (frequency domain similarity). Range: [0, ln 2]. Lower is better.

Reference: Welch, P. D. (1967). The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms. IEEE Transactions on Audio and Electroacoustics, 15(2), 70–73. https://doi.org/10.1109/TAU.1967.1161901

Dataset	Score
00000	-0.000000

Lower is better (0 = identical distributions).

FD — Fréchet Distance (FD): measures the distance between multivariate Gaussians fitted to ECG feature embeddings of real and synthetic sets per lead. Analogous to FID for images. Lower is better.

Reference: Heusel, M., et al. (2017). GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NeurIPS, 30. https://arxiv.org/abs/1706.08500 | Thambawita, V., et al. (2021). DeepFake electrocardiograms. Scientific Reports, 11, 21896. https://doi.org/10.1038/s41598-021-01295-2

Dataset	Score
00000	0.000000

Lower is better (0 = identical distributions).

Overall Quality Summary

Radar chart: each axis is one metric, normalised so the outer ring = best quality.

References

MMD: Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). A kernel two-sample test. Journal of Machine Learning Research, 13, 723–773. https://jmlr.org/papers/v13/gretton12a.html
DTW: Berndt, D. J., & Clifford, J. (1994). Using dynamic time warping to find patterns in time series. KDD Workshop, 10(16), 359–370. | Goldberger, A. L., et al. (2000). PhysioBank, PhysioToolkit, and PhysioNet. Circulation, 101(23), e215–e220. https://doi.org/10.1161/01.CIR.101.23.e215
PRD: Zigel, Y., Cohen, A., & Katsevman, A. (2000). The weighted diagnostic distortion measure for ECG signal compression. IEEE Transactions on Biomedical Engineering, 47(11), 1422–1430. https://doi.org/10.1109/10.871205
PSD: Welch, P. D. (1967). The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms. IEEE Transactions on Audio and Electroacoustics, 15(2), 70–73. https://doi.org/10.1109/TAU.1967.1161901
FD: Heusel, M., et al. (2017). GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NeurIPS, 30. https://arxiv.org/abs/1706.08500 | Thambawita, V., et al. (2021). DeepFake electrocardiograms. Scientific Reports, 11, 21896. https://doi.org/10.1038/s41598-021-01295-2

Generated by ecgen-eval v0.1.0 — GitHub