ecgen-eval 0.4.0

Evaluation metrics and interactive reports for synthetic ECG quality assessment

ECGEN-Eval

Evaluation metrics and interactive HTML reports for synthetic ECG quality assessment. Part of the ECGEN research suite (ECGEN-FM, ECGEN-VAE, Pulse2Pulse).

Installation

pip install -e .           # core
pip install -e ".[dev]"    # + pytest, black, ruff

Optional dependencies:

pip install wfdb           # for loading PTB-XL / WFDB records
pip install scipy          # for Welch PSD, resampling, and R-peak detection (recommended)
pip install dtaidistance   # fast Cython DTW (falls back to pure-Python)

# Clinical feature extraction (Task 1.2)
pip install -e ".[features]"   # installs neurokit2 + scipy
# or individually:
pip install neurokit2 scipy

Quick start

import numpy as np
from ecgen_eval import load_ecg_npy, AVAILABLE_METRICS

# Load model outputs (shape: N × L × T)
real  = load_ecg_npy("real_ecgs.npy",  label="PTB-XL")
synth = load_ecg_npy("synth_ecgs.npy", label="ECGEN-FM")

# Compute a metric
mmd = AVAILABLE_METRICS["mmd"]()
result = mmd.compute(real.data, synth.data, lead_names=real.lead_names)
print(f"MMD = {result.score:.6f}")
print(result.per_lead_scores)

Data loaders

Function	Input	Use case
load_ecg_npy(path)	.npy / .npz (N, L, T)	ECGEN-FM, ECGEN-VAE, Pulse2Pulse outputs
load_ecg_wfdb(dir)	WFDB .hea/.dat records	PTB-XL and PhysioNet datasets
load_ecg_folder(path)	auto-detected	Any of the above + CSV

load_ecg_folder auto-detects the format based on file extensions.

Metrics

All metrics are available via AVAILABLE_METRICS["key"]() and return a MetricResult with .score, .per_lead_scores, and .extra.

Original metrics

Key	Name	Lower / Higher	Paper
mmd	Maximum Mean Discrepancy	Lower ↓	Gretton et al., JMLR 2012
dtw	Dynamic Time Warping	Lower ↓	Berndt & Clifford, KDD 1994
prd	Percent Root-mean-square Difference	Lower ↓	Zigel et al., IEEE TBME 2000
psd	PSD Divergence (Jensen-Shannon)	Lower ↓	Welch, IEEE Trans Audio 1967 · Applied: Golany & Radinsky, AAAI 2019
fd	Fréchet Distance	Lower ↓	Heusel et al., NeurIPS 2017 · Applied: Thambawita et al., Sci Rep 2021

New metrics (literature review 2018 – 2025)

Key	Name	Lower / Higher	Paper
psnr	Peak Signal-to-Noise Ratio	Higher ↑	Huynh-Thu & Ghanbari, Electron Lett 2008
ssim	Structural Similarity Index (1-D)	Higher ↑	Wang et al., IEEE TIP 2004 · ECG: Hayn et al., Physiol Meas 2018
swd	Sliced Wasserstein Distance	Lower ↓	Rabin et al., LNCS 2011 · Kolouri et al., NeurIPS 2019
mae	Mean Absolute Error (nearest-neighbour)	Lower ↓	Zigel et al., IEEE TBME 2000
hrv	Heart Rate Variability (SDNN/rMSSD/pNN50)	Lower ↓	Task Force ESC/NASPE, Circulation 1996
pr_dist	Distribution Precision & Recall	Higher ↑	Kynkäänniemi et al., NeurIPS 2019
sqi	Signal Quality Index (QRS template)	Higher ↑	Clifford et al., CinC 2017
nn_dist	Nearest-Neighbour Distance (memorisation)	Lower ↓	Meehan et al., AISTATS 2020
heart_rate	Heart Rate Distribution (JS divergence)	Lower ↓	Pan & Tompkins, IEEE TBME 1985 · Applied: Golany & Radinsky, AAAI 2019
spectral_entropy	Spectral Entropy Divergence	Lower ↓	Inouye et al., EEG Clin Neurophysiol 1991

Feature Extraction

Extract 10 clinical ECG features per recording and compare their distributions between real and synthetic datasets.

from ecgen_eval import load_ecg_npy
from ecgen_eval.features import extract_features

real  = load_ecg_npy("real_ecgs.npy",  label="PTB-XL")
synth = load_ecg_npy("synth_ecgs.npy", label="ECGEN-FM")

# Returns (beat_df, summary_df)
_, real_feat  = extract_features(real)
_, synth_feat = extract_features(synth)

print(real_feat[["rr_mean_ms", "hr_bpm", "qrs_duration_ms",
                  "qt_interval_ms", "hrv_sdnn_ms"]].describe())

Extracted features (one value per recording):

Feature	Column	Description
RR interval	rr_mean_ms, rr_std_ms	Mean and std of peak-to-peak intervals
Heart rate	hr_bpm	60 000 / mean(RR)
PR interval	pr_interval_ms	P-onset to QRS-onset
QRS duration	qrs_duration_ms	Width of QRS complex
QT interval	qt_interval_ms	QRS-onset to T-end
QTc (Bazett)	qtc_bazett_ms	QT / √RR(s)
ST deviation	st_deviation_mv	Mean ST voltage vs isoelectric line
T-wave amplitude	t_amplitude_mv	Signed T-peak amplitude
P-wave duration	p_wave_duration_ms	P-onset to P-offset
HRV	hrv_sdnn_ms, hrv_rmssd_ms, hrv_pnn50	SDNN, RMSSD, pNN50

2D and 3D distribution plots

from ecgen_eval.visualization.feature_plots import (
    plot_feature_2d,
    plot_feature_pair_grid,
    plot_feature_3d,
    plot_feature_triple_grid,
)

# Single 2D scatter + KDE contour
fig = plot_feature_2d(real_feat, synth_feat, "rr_mean_ms", "hr_bpm")
fig.show()

# Grid of default feature pairs
fig = plot_feature_pair_grid(real_feat, synth_feat)
fig.show()

# Interactive 3D scatter
fig = plot_feature_3d(real_feat, synth_feat, "rr_mean_ms", "hr_bpm", "qrs_duration_ms")
fig.show()

CLI

# Run all metrics and generate interactive HTML report
ecgen-eval eval \
  --real   /path/to/real_ecgs.npy \
  --synthetic /path/to/model1_ecgs.npy \
  --synthetic /path/to/model2_ecgs.npy \
  --output report.html

# Include clinical feature distributions in the report (requires neurokit2)
ecgen-eval eval \
  --real /path/to/real_ecgs.npy \
  --synthetic /path/to/synth_ecgs.npy \
  --feature-plots \
  --output report.html

# Extract features only — save summary CSV
ecgen-eval features \
  --real /path/to/real_ecgs.npy \
  --synthetic /path/to/synth_ecgs.npy \
  --output features.csv

# List available metrics
ecgen-eval list-metrics

# ECG waveform comparison only (no metrics)
ecgen-eval visualize \
  --real /path/to/real.npy \
  --synthetic /path/to/synth.npy \
  --compare-mode overlay

Key options for eval:

• --metrics mmd prd psd psnr hrv — run a subset of metrics
• --feature-plots — add clinical feature extraction + 2D/3D distribution plots
• --fs 500 — sampling frequency
• --n-samples 5 — ECG samples shown in gallery
• --offline — embed Plotly.js for offline viewing
• --max-samples 500 — cap loaded recordings per folder

Report

The HTML report includes:

• Dataset summary table
• Interactive ECG gallery (overlay / side-by-side / grid)
• Per-metric: score table, per-lead bar chart, PSD overlay (PSD metric), violin plot (DTW)
• Radar chart summarising all metrics
• Metrics summary table with Download CSV button
• Clinical Feature Summary table (mean ± std, real vs synthetic) — with --feature-plots
• 2D feature distribution plots (scatter + KDE contours) — with --feature-plots
• 3D feature distribution plots (interactive rotation) — with --feature-plots
• References / citations

Running tests

pytest tests/ -v

Project structure

ecgen_eval/
├── data/
│   ├── dataset.py          ECGDataset container
│   └── loader.py           npy / WFDB / CSV loaders
├── features/               Clinical ECG feature extraction (optional: neurokit2)
│   ├── __init__.py         Exports FeatureExtractor, extract_features
│   ├── extractor.py        FeatureExtractor class → (beat_df, summary_df)
│   ├── qrs_detection.py    R-peak detection (neurokit2 primary, scipy fallback)
│   ├── interval_features.py  RR, HR, PR, QRS, QT, QTc, P-wave duration
│   ├── amplitude_features.py ST deviation, T-wave amplitude
│   ├── hrv_features.py     SDNN, RMSSD, pNN50
│   └── utils.py            Bandpass filter, peak helpers
├── metrics/
│   ├── base.py             BaseMetric + MetricResult
│   ├── mmd.py              MMD
│   ├── dtw.py              DTW
│   ├── prd.py              PRD
│   ├── psd.py              PSD divergence
│   ├── fd.py               Fréchet Distance
│   ├── psnr.py             PSNR
│   ├── ssim.py             1-D SSIM
│   ├── swd.py              Sliced Wasserstein Distance
│   ├── mae.py              MAE / RMSE
│   ├── hrv.py              HRV statistics (SDNN, rMSSD, pNN50)
│   ├── pr_dist.py          Distribution Precision & Recall
│   ├── sqi.py              Signal Quality Index
│   ├── nn_dist.py          Nearest-Neighbour Distance
│   ├── heart_rate.py       Heart Rate Distribution
│   └── spectral_entropy.py Spectral Entropy Divergence
├── visualization/
│   ├── ecg_paper.py        ECG graph-paper figure factory
│   ├── ecg_waveform.py     Waveform comparison plots
│   ├── metric_plots.py     Bar, violin, PSD overlay, radar
│   └── feature_plots.py    2D/3D clinical feature distribution plots
├── report/
│   └── html_report.py      HTML report generator
└── cli.py                  Click CLI entry point (eval, visualize, features, list-metrics)
tests/
├── conftest.py
├── test_dataset.py
├── test_loader.py
├── test_metrics.py
└── test_features.py
docs/
└── ecg_features_review.md  Literature review of the 10 clinical features