Wilson Central Terminal ECG Database

540 ten-second segments from 92 patients admitted to Campbelltown Hospital, Sydney, recorded by the MARCS Institute at Western Sydney University with the left-arm, right-arm and left-leg electrode potentials brought out individually — so that the Wilson Central Terminal could be measured rather than assumed.

The WCT is the reference V1–V6 are measured against. Conventional electrocardiography treats it as 0 V; this release exists because it is not, and the authors report WCT amplitudes reaching 241% of lead II. With the WCT measured, the six chest electrodes can be referenced against a true zero instead, giving the true unipolar chest leads UV1–UV6 that ship alongside the conventional V1–V6. V = UV − WCT holds by construction, which is what makes the release usable both ways: to quantify the error the assumption introduces, and to develop methods that reconstruct or correct it.

So the WCT channel is the point of this dataset, not an artefact to filter out, and the “37 leads” are not a 37-lead ECG — see the next section.

The cohort is clinical and elderly: mean age 65.2 (SD 12.1, range 41–94), 27 of 92 patients female, mostly with a history of cardiac disease. Each patient contributed 1 to 31 segments, which is why folds are grouped on patient.

This is the only dataset in ECGBench that ships every channel twice, and the only 12-lead-family dataset with no aVR, aVL or aVF at all. Channel order is identical in all 540 headers:

“Filtered” means DC removal plus a 0.05–150 Hz band-pass. There is no WCT-Raw.

Three practical consequences:

• Do not mix the two families in one tensor. They are two preprocessing states of the same recording. Index 18 is filtered lead I; index 0 is raw lead I. Select by name (leads=[...]), never by index.
• The raw limb and unipolar channels carry several mV of DC offset, which is exactly what the DC-removal filter takes out — in patient001/seg01 the raw LA/RA/LL and UV1–UV6 channels sit around −5 mV while their filtered counterparts are centred near zero. That is why amplitude_range_mv is a deliberately wide ±20 rather than the usual ±10: an unreferenced potential is not a lead voltage.
• aVR, aVL and aVF must be derived if you want them: aVR = −(I + II)/2, aVL = I − II/2, aVF = II − I/2.

Also unusual: 800 Hz, and 8001 samples — 10.00125 s, not a round number. All 540 records are identical in geometry, so the truncation check is exact.

Eight segments from five patients (007, 008, 010, 014, 031) have one or two precordial channels that were reconstructed as V = UV − WCT rather than recorded, because the recorded channel was unusable. The release notes this per record in a #Reconstruct Precordials: header comment, and ECGBench exposes it as reconstructed_precordials (a list) and has_reconstructed_precordials.

Exclude them when evaluating precordial reconstruction, which is this dataset’s headline use. On those channels V = UV − WCT holds exactly by construction, so a method that estimates V from UV and WCT is being scored against its own output.

from ecgbench import load_labels

labels = load_labels("wctecgdb", "/path/to/wctecgdb/1.0.1/")
measured = labels[~labels["has_reconstructed_precordials"]]
len(labels), len(measured)          # (540, 532)

All 1,082 shipped files were verified against the release’s own SHA256SUMS.txt before any figure here was computed — all OK. Record and patient counts match the landing page exactly at 540 and 92, as do the 27 female patients, the mean age of 65.23 (SD 12.12), the 10 patients with unreported diagnoses and the 8 reconstructed segments. There is nothing to reconcile.

Every figure above is recomputed from the 540 .hea files, which are the only place any label exists — the release ships no metadata table (just the .dat/.hea pairs under patient001/…patient092/, plus RECORDS, SHA256SUMS.txt and LICENSE.txt). ECGBench parses the #Age:, #Sex:, #Diagnosis report: and #Reconstruct Precordials: comment lines; see ecgbench/labels/wctecgdb.py.

The two count columns are the same information weighted differently, and the difference is large. Age, sex and diagnosis are constant within a patient (verified across all 540 headers), while segment counts run 1–31 (mean 5.9, median 4). Five patients contribute 132 records — 24.4% of the dataset — and 18 patients contribute one each. Cardiomyopathy or heart failure is 5.4% of patients but 12.6% of records. Any per-record rate you quote is a statement about segment counts as much as about patients, so weight by patient or group by patient_id.

The diagnosis is an admission label, not a waveform label. It records why the patient came in, not what these ten seconds show: a segment in the “Other tachyarrhythmia” group need not contain a tachycardia. There is no beat, rhythm or interval annotation anywhere in this release.

The group column is a reduction, and the reduction is a judgement call. The headers hold 43 distinct free-text strings over 92 patients (40 after correcting the misspellings), 28 of them held by a single patient — so the strings cannot be stratified on directly. diagnosis_group reduces them to the 8 groups above via an explicit, auditable map in ecgbench/labels/wctecgdb.py, which raises rather than guessing on a string it does not know. Six patients’ strings name two conditions (e.g. Non ST segment elevation myocardial infarction (NSTEMI)- rapid Atrial fibrillation), and the map picked one. Train on diagnosis, not diagnosis_group.

Three transcription traps live in these header lines, all handled and regression-tested:

• The headers are not UTF-8. The dash in “ST segment elevation” is byte 0xA0, a Windows-1252 non-breaking space, so a strict UTF-8 read raises and errors="replace" silently yields ST�segment�elevation. ECGBench decodes cp1252 and folds the NBSPs; diagnosis_raw keeps the original bytes.
• Some strings are misspelt or inconsistently cased: Atypica chest pain, Type 2 Myocaridal infarctoin, Congestive Cardic failure (CCF), and sinus bradycardia alongside Sinus bradycardia. Left alone, each variant becomes its own class. diagnosis is corrected, diagnosis_raw is verbatim.
• not reported is a value, not a blank. 10 patients / 38 records. Exposed as diagnosis_reported = False, because “nothing recorded” and “column missing” are different facts.

Every record passes validation, but “valid” is not “unclipped”. 128,802 samples across the release sit at exactly ±9.2250 mV, the acquisition rail: 140 of the 540 records have at least one channel pinned there, 61 of them for a plateau of 10 ms or longer, the longest 227.5 ms in patient069/seg07 UV1-Raw.

It is concentrated where you would expect — the unreferenced channels, before DC removal: UV2-Raw (83 records), UV6-Raw (46), UV1-Raw (44), UV3-Raw (39), UV5-Raw (32), UV4-Raw (24), LA-Raw (20), RA-Raw (13), and single figures for the rest. This is saturation of the raw unipolar and limb-potential channels, i.e. a property of the recording setup, not download damage — all 1,082 files match the release’s SHA256SUMS.txt.

amplitude_range_mv is therefore ±20, wide enough not to reject the rail as an outlier. If clipped samples matter for your method, test for the rail yourself:

import numpy as np, wfdb
sig = wfdb.rdrecord("patient069/seg07").p_signal          # (8001, 37)
clipped = np.abs(np.abs(sig) - 9.2250) < 1e-3
clipped.sum(), clipped.any(axis=0).nonzero()[0]           # (1486, array([3, 12]))  -- V1-Raw, UV1-Raw

Only 4 records exceed ±10 mV, and three of them are exactly the synthesised channels of patients 008 and 031 (V1-Raw down to −16.51 mV in patient008/seg01): V = UV − WCT is a difference of two channels, so it can overshoot a rail that neither input can pass.

Fold boundaries follow patients, not records: StratifiedGroupKFold on patient_id, stratified on diagnosis_group. Without grouping, a patient with 31 near-identical ten-second windows of one admission would land on both sides of every split — and since the diagnosis is constant within a patient, the label would leak outright. Verified on the shipped release: no patient spans two folds.

Two consequences of grouping 92 patients into 10 folds:

• Fold sizes vary — 7–11 patients and 49–58 records per fold rather than a uniform 54, because segment counts run 1–31. The default 8/1/1 mapping gives 439 train / 49 val / 52 test.
• Stratification is approximate, and quantifiably so: three groups hold 3–6 patients, fewer than the 10 folds. Bradyarrhythmia or conduction block (3 patients) reaches only 3 folds, and the test fold contains no bradyarrhythmia and no cardiomyopathy record at all. Other tachyarrhythmia lands 13 records in test against 18 in train, because one high-segment patient carries it. StratifiedGroupKFold does not emit the “least populated class” warning StratifiedKFold does, so the silence is not evidence of balance — check the per-fold table if a group matters to you.

For a stricter cross-validation harness, rotate the held-out fold rather than using the default 8/1/1 mapping:

from ecgbench import ECGDataset

# Hold out fold 3, train on the rest. With split=None each sample's
# ["split"] reports that record's own default split.
train = ECGDataset("wctecgdb", split=None,
                   fold_numbers=[1, 2, 4, 5, 6, 7, 8, 9, 10],
                   data_path="/path/to/wctecgdb/1.0.1/")
held  = ECGDataset("wctecgdb", split=None, fold_numbers=[3],
                   data_path="/path/to/wctecgdb/1.0.1/")