Ningbo First Hospital ECG Database (Idiopathic Ventricular Arrhythmia)

334 twelve-lead ECGs recorded during catheter ablation, one per patient, each labelled with the outflow tract the ablation proved the arrhythmia came from. That is what makes 334 records worth having: the ground truth is invasive, so the intended task is to predict the origin from the surface ECG before the procedure, which is a decision an operator actually has to make.

Every patient had a successful ablation, defined as absence of spontaneous or induced outflow-tract arrhythmia 30 minutes after the last energy delivery, and patients whose frequent ectopy recurred within six months were excluded. Only single-source arrhythmias are included.

Three properties will surprise anyone who assumes this looks like a resting 12-lead dataset:

The lead order is alphabetical. The CSV header reads aVF,aVL,aVR,I,II,III,V1…V6. So signal[0] is aVF, not lead I, and signal[4] is lead II, not aVL. Use leads=[...] to select by name — see the snippet below.

The samples carry no declared unit. The CSVs hold bare integers, and neither the paper nor figshare states a gain; the paper’s own Figure 8 plots the raw counts with a MATLAB ×10⁴ multiplier and no y-axis unit. ECGBench supplies a measured estimate rather than leaving the tensors uncalibrated — see “About the amplitude scale”.

The sampling rate is 2000 Hz, the highest in this catalogue and four times the usual 500 Hz, because the recordings come from an EP-lab acquisition system (Abbott EP-WorkMate) rather than a diagnostic cart. A 10 s window is 20,000 samples here. Record length runs 2.9 s to 59.3 s in 317 distinct lengths over 334 records, because the providers cut an excerpt containing both sinus beats and ectopy out of each procedure recording.

Every figure on this page was recomputed from the shipped Diagnosis.xlsx and the 334 CSV files. Most agree with the paper exactly: 334 patients, 257 RVOT (77%) / 77 LVOT (23%), 104 male / 230 female, and the per-class male split (65 of 257 RVOT, 39 of 77 LVOT). Two figures do not.

The 40 blank sub-sites are explained by the paper, but ECGBench does not fill them in. Table 2 assigns 45 RVOT patients to “RVOTOther” and one LVOT patient to “NA”. The shipped spreadsheet has 6 explicit RVOTOther plus 39 blanks among the RVOT patients, and the single LVOT blank — 6 + 39 = 45, and the arithmetic closes on both tracts. The inference is almost certainly right, but it is our inference, not the providers’, so the cells stay blank in the labels and the table above shows both columns.

The presentation type disagrees, and the shipped file wins. Table 1 reports 325 frequent PVC / 9 sustained VT (RVOT 251/6, LVOT 74/3); Diagnosis.xlsx reads 329 PVC / 5 VT (RVOT 254/3, LVOT 75/2). Four patients are VT in the paper and PVC in the file. There is no changelog, and figshare shows a single revision, so which is correct cannot be established from the release. ECGBench reports the file. Do not quote the paper’s 9.

No age ships at all. The paper reports a cohort mean of 46.1 ± 13.1 years and per-tract means, but Diagnosis.xlsx has five columns — HospitalID, Type, LeftRight, Sublocation, Gender — and none of them is age. It is not recoverable from the data.

Sub-site is not usable as a split target. 12 values over 334 patients, five of them under ten cases, plus 40 blanks. Ten-fold stratification is not well defined on it, so ECGBench stratifies on left_right and exposes sublocation for analysis.

Sex leaks the label. 230 of 334 patients are female overall, but 192 of 257 RVOT cases against 38 of 77 LVOT ones. Sex alone separates the classes better than chance, so a model handed it as a feature can learn it instead of the ECG.

signal_unit_scale for this dataset is an ECGBench estimate, not a declared value. The release ships integers with no unit anywhere: the paper states the sampling rate and the acquisition system and never states a gain, and its Figure 8 plots the counts directly.

Rather than hand users uncalibrated tensors — which would make units="mV", amplitude_range_mv and any cross-dataset comparison meaningless — the scale was measured. Method: median lead-II R-peak prominence (Butterworth 0.5–40 Hz, scipy.signal.find_peaks) over all 334 records, compared sex for sex against sph, whose samples are millivolts by declaration. Sex matters because this cohort is 69% female and QRS amplitude differs by sex.

16,384 sits between the two estimates and is the nearest binary-clean gain of the kind acquisition systems export. Treat it as accurate to roughly ±20%: every waveform’s shape is exact, its absolute millivolt calibration is not. The same measurement confirmed the 2000 Hz rate independently — R-peak intervals give a median 81.5 bpm at 2000 Hz, which would be an impossible 20 bpm at 500 Hz.

If you would rather calibrate it yourself, divide the millivolt values by 6.1035e-05 to recover the shipped integers exactly, or pass units="uV" and scale from there.

The release ships the same 334 recordings twice: PVCVTRawECGData/ (raw integers) and PVCVTECGData/ (wavelet-denoised, coif5 with a SURE threshold, per the providers’ MATLAB script). ECGBench points at the raw copy, matching chapman_shaoxing, which likewise takes ECGData/ over ECGDataDenoised/.

The denoised copy is not a drop-in substitute, for two measured reasons:

• The denoiser ran on each lead independently, so the leads no longer agree with each other. III − (II − I) is under 1% of III’s RMS in the raw files and up to 14% in the denoised ones; the Goldberger relations degrade similarly. Anything that derives one lead from others, or checks internal consistency, will see the difference.
• 106 of the 334 denoised files are shorter than their raw counterparts, by up to 7×, even though the published denoising script preserves row count exactly. The two directories are therefore not sample-aligned, and a window computed on one does not transfer to the other.

ecgbench.labels.ningbo_iva exposes the denoised path as signal_path_denoised for anyone who wants it anyway.

Ten folds stratified on left_right, which is already one label per patient, so nothing is reduced. No patient grouping is configured, and none is needed: HospitalID is simultaneously the record and the patient identifier — 334 patients, 334 files, an exact one-to-one match in both directions — so patient_id_column stays null rather than asserting a grouping nothing exercised.

What that produced: folds of 33–34 records, each holding 25–26 RVOT and 7–8 LVOT, against the 77/23% cohort ratio. Fold membership is identical between original/ and clean/, which here is trivially true since nothing was excluded.

With 33 records per fold, single-split results on this dataset are noisy. Use the ten folds as cross-validation — ECGDataset(split=None, fold_numbers=[...]) selects by fold across the default layout — rather than quoting one 33-record test set.

amplitude_range_mv is left at [-10, 10] deliberately even though the largest sample anywhere in the release is 9.45 estimated mV, so the check is a guard rather than a filter. Tightening it would turn a 20% error in the estimated scale into hundreds of spurious exclusions.

Ningbo First Hospital contributed to two unrelated public releases, and they are easy to confuse. The PhysioNet ecg-arrhythmia dataset (catalogued here as Chapman-Shaoxing Arrhythmia) merges the Chapman-Shaoxing and Ningbo cohorts into 45,152 records, of which 34,905 are Ningbo’s — routine diagnostic 10 s ECGs at 500 Hz under JS… names.

This dataset is a different thing entirely: 334 intra-procedural EP-lab recordings at 2000 Hz, of variable length, named by hospital number. No recording can be shared between the two. Whether the same patients appear in both is unverifiable — the releases publish no common identifier, and this one publishes no age or date to join on — so the overlap edge on the Chapman-Shaoxing Arrhythmia page stays verified: false.