| Format | 15-lead (12 + 3 Frank) · 32-120 s · 1,000 Hz |
|---|---|
| Patients | 290 |
| Records | 549 |
| Leads | 15 |
| License | ODC-By 1.0 |
| Origin | Univ. Clinic Benjamin Franklin — Germany — Berlin |
549 records from 290 subjects, recorded at the Benjamin Franklin University Clinic in Berlin. This is the most unusual dataset in the catalogue on three counts, all verified against the files.
15 signals, not 12. The conventional leads plus the three Frank
vectorcardiography leads vx, vy, vz. Leads 1-12 live in the .dat
file and 13-15 in a companion .xyz; wfdb reads both as one record. Names
are lowercase in every header. ECGDataset(leads=[...]) takes the standard
twelve by name.
Records are variable length — 11 distinct lengths between 32 s and
120 s, at 1000 Hz. They cannot be batched as they are: torch cannot stack
tensors of differing width. Take a fixed window=(start, length), or use
batch_size=1.
For the same reason expected_samples is deliberately empty in the config,
which disables the truncation check rather than failing every short record.
No metadata file ships at all. Every clinical field is in the per-record
.hea comment block — 47 keys covering the admission diagnosis, infarction
localisation, a full haemodynamics panel, coronary stenosis findings and the
therapy history. ECGBench parses them into a metadata CSV on first run.
113 of the 290 patients contributed more than one recording (up to seven), so folds are grouped by patient. No patient spans a fold.
| Diagnosis | Records | Patients |
|---|---|---|
| Myocardial infarction | 368 | 148 |
| Healthy control | 80 | 52 |
| (none recorded) | 27 | 22 |
| Cardiomyopathy | 17 | 15 |
| Bundle branch block | 17 | 15 |
| Dysrhythmia | 16 | 14 |
| Hypertrophy | 7 | 7 |
| Valvular heart disease | 6 | 6 |
| Myocarditis | 4 | 4 |
| Stable angina | 2 | 2 |
| Heart failure (NYHA 2/3/4) | 1 each | 1 each |
| Palpitation | 1 | 1 |
| Unstable angina | 1 | 1 |
Recomputed from the # Reason for admission: field of all 549 headers.
Single-label — one diagnosis per record, and no patient has conflicting
diagnoses across their recordings, so the record and patient columns are
both meaningful and neither sums to a multi-label total.
Two disagreements with the README, worth knowing:
CONTROLS
file exactly, across 52 patient directories.27 records carry no diagnosis (n/a in the header). ECGBench labels
those UNKNOWN and, because seven classes have fewer than 10 records,
pools those into OTHER for stratification. primary_diagnosis is that
pooled view — train on diagnosis, which is verbatim.
| Version | Records | Note |
|---|---|---|
| original | 549 | all records, with is_valid + quality_issues |
| clean | 546 | 99.5% pass rate |
| excluded | 3 | 2 amplitude outliers past ±15 mV, 1 with flat leads |
ecgbench splits --dataset ptbdb --data-path /path/to/ptbdb/1.0.0/from ecgbench import ECGDataset
# Variable-length records: take a fixed window to allow batching, and the
# standard twelve leads by name out of the fifteen stored. window= is read
# at load time, so the other 22-110 s are never decoded.
STANDARD_12 = ["i", "ii", "iii", "avr", "avl", "avf",
"v1", "v2", "v3", "v4", "v5", "v6"]
ds = ECGDataset(
"ptbdb",
split="train",
data_path="/path/to/ptbdb/1.0.0/",
leads=STANDARD_12,
window=(0, 10_000), # first 10 s at 1000 Hz
labels=True,
)
ds[0]["signal"].shape # (12, 10000)
ds[0]["labels"]["diagnosis"] # 'Myocardial infarction'
ds[0]["labels"]["age"] # 77.0
# Without leads= you get all 15 signals including vx, vy, vz; without the
# window, a DataLoader raises as soon as a batch mixes two record lengths.