ZZU pECG (Zhengzhou University Pediatric ECG Database)

14,190 ECGs from 11,643 hospitalised children at the First Affiliated Hospital of Zhengzhou University, recorded 2018-2024 at 500 Hz. Ages run from 1 day to 15.0 years (median 8.6), which makes this the only genuinely paediatric 12-lead resource in the catalogue.

Paediatric ECGs are not small adult ECGs. Right-axis dominance and T-wave inversion in the right precordial leads are normal in an infant and abnormal in an adolescent, so age is not an optional covariate here and an adult-trained model should not be assumed to transfer. Age ships in days, and 546 records are under a year old — rounding to whole years collapses the entire neonatal and infant range.

Not every record has 12 leads, and this is the trap on this page. 12,334 records store the standard 12; the other 1,856 store only 9, dropping V2, V4 and V6. The 9-lead layout is not a prefix of the 12-lead one — stored position 7 is V2 in one and V3 in the other — so index-based lead selection silently crosses those two leads for 13% of the release. ECGBench declares both layouts, so ECGDataset(leads=["V2"]) raises on a reduced record instead of handing back V3, and n_leads is exposed for filtering. This is the only dataset in the catalogue that needs that mechanism.

Dropping the 9-lead records is not a neutral filter. They carry a studied cardiovascular diagnosis at more than twice the rate of the 12-lead records (933 of 1,856, 50.3%, against 2,783 of 12,334, 22.6%), so excluding them preferentially discards sick children.

Labels are unusually rich for an open release. Every record carries one or more ECG findings coded in both the AHA and Chinese vocabularies with human-readable descriptions, a list of ICD-10 discharge diagnoses, and three per-lead signal-quality indices (pSQI, basSQI, bSQI) computed by the depositors.

Record length varies by a factor of 24, from 5 s to 120 s across 67 distinct lengths. So expected_samples is deliberately empty in the config, and any fixed window= must fit the 5 s shortest record — which is also what makes a DataLoader batch possible at all.

All five published files match the md5s figshare lists — AttributesDictionary.csv ffbbb7ebd8ad4425b3a859739eb65eb1, ECGCode.csv 7612c5af0e01e052eb85792c5e362f12, DiseaseCode.csv 0bff597c23397c0d05a937fe09807ad4, and both halves of the split zip.

The structural figures agree with the data descriptor exactly: 14,190 records, 11,643 children, 12,334 twelve-lead, 1,856 nine-lead, 500 Hz, 5-120 s.

The diagnosis count does not. The descriptor states that 3,516 records were “diagnosed with cardiovascular diseases”; recomputing gives 3,716 records over 2,597 patients. The derivation is stated exactly so the difference is checkable: split the ICD-10 code cell on ;, strip the surrounding quotes, and count a record if any element is one of the 19 real ICD-10 codes in DiseaseCode.csv. Two details matter — the codes carry study prefixes ((FO) Q21.1, (OSD) Q21.1, (F) I40.0) which are part of the key and are matched literally, and that file’s 20th row is the placeholder See attribute dictionary file rather than a code, so it is ignored. The 200-record gap is unexplained by the release.

The two code columns are parallel and complementary, and neither is sufficient alone. AHA_code and CHN_code list the same findings in the same order, and each gives its own vocabulary’s code where one exists and the plain-English description where it does not: ECGCode.csv has no AHA code for 14 of its 105 findings (Osborn wave, left ventricular high voltage, prolonged QTc, abnormal Q wave, …) and no CHN code for 29. So 6,473 of the 26,797 AHA_code entries are prose, and reading the column as a code vocabulary invents 15 phantom “codes”. ECGBench normalises both and exposes ecg_findings as the canonical description regardless of which vocabulary names the finding.

Modifiers are glued onto codes, differently in each vocabulary. AHA writes L145+Modifier362 and bare qualifiers like Suggests208; CHN writes L121+Depression, F55+Frequent, D21+Occasional, plus composites such as J(111+112+113). aha_base_codes and chn_base_codes strip the modifier — and note the composite must not be split at its internal +, or the base comes out as the truncated J(111+112.

ICD-10 codes are admission diagnoses, not readings of the tracing. A record can carry a diagnosis whose ECG signature is absent from that particular recording. The ECG-finding columns are the tracing-level labels; the disease groups are the patient-level ones. They are different quantities and the table above is the latter.

Amplitudes have a hard rail at about 26.6 mV that 11.8% of records touch, with sustained rather than spike excursions (a median 2.1% of samples above 10 mV, up to 41%). These are genuine saturation, and the release’s own baseline quality index independently agrees: railed records have a median basSQI of 0.961 against 0.983 for the rest (r = −0.36 against peak amplitude). The validation range is ±20 mV rather than ±10 because exclusion rates barely move across that span (±10: 18.8%, ±15: 16.4%, ±20: 14.8%, ±25: 11.8%) — almost every offender is railed rather than merely large — and genuine paediatric voltages are high: “left ventricular high voltage” is the second commonest finding in the release at 2,789 records.