Brugada-HUCA

363 twelve-lead resting ECGs, one per individual, from people investigated for Brugada syndrome at Hospital Universitario Central de Asturias in Oviedo, Spain. Brugada syndrome is a rare inherited arrhythmia disorder marked by coved-type ST-segment elevation in the right precordial leads V1–V3, often with a right bundle branch block pattern, and it carries a risk of sudden cardiac death.

This is the smallest and cleanest dataset in ECGBench, and the only one sampled at 100 Hz alone — PTB-XL offers 100 Hz as an alternative to 500 Hz, but here it is the native and only rate. Every record is a uniform 12 leads × 1200 samples (12.0 s), the lead order is standard, and all 363 records pass every quality check: no NaN samples, no flat leads, and no sample beyond ±3.81 mV.

It is a screening cohort, not a case-control study. Everyone here was investigated for suspected Brugada syndrome, so the brugada = 0 class means “investigated and not diagnosed” rather than a general-population control. A classifier trained on this data estimates diagnosis among referred individuals, which is a different quantity from population prevalence, and the 19% positive rate reflects referral practice at one tertiary hospital.

Because each subject contributes exactly one record, there is no within-patient leakage to guard against, so folds are stratified but not grouped.

All figures recomputed from the shipped metadata.csv, which was verified against the release’s own SHA256SUMS.txt along with RECORDS, README.md, LICENSE.txt, the data dictionary and a sample of signal files. The record and subject counts match the release description exactly at 363 and 363, so there is nothing to reconcile — unusual in this catalogue.

The three labels are independent columns, not one taxonomy, and the release documents basal_pattern as independent of the diagnosis. They are far from orthogonal in practice: 19 of the 46 subjects with a pathological baseline are also confirmed Brugada cases, and 4 of the 7 atypical cases have one. Do not treat a pathological baseline as a proxy for the diagnosis.

The stratification label is brugada, used verbatim. There is no derivation and therefore nothing that can drift from what labels=True returns — unlike the reduced labels ECGBench derives for PTB-XL, Challenge 2021 or MIMIC-IV-ECG.

The rare class is deliberately not pooled. brugada = 2 has 7 records, fewer than the 10 folds, so it cannot appear in every fold and scikit-learn emits a warning to that effect. Pooling it would be clinically wrong: “other/atypical” is neither healthy nor confirmed. In the released folds those 7 records are spread one per fold across 7 of the 10 folds, and the class shares per default split are 79.0/19.2/1.7% in train against 80.6/16.7/2.8% in test.

Three quirks belong to the shipped release, not to any one download — each was confirmed against SHA256SUMS.txt:

• RECORDS lists 364 lines for 363 records: files/596382/596382 appears twice. ECGBench enumerates subjects from metadata.csv instead, which is authoritative and carries the labels.
• A macOS files/.DS_Store ships inside the release and is itself checksummed, so any code globbing files/* must filter to directories.
• metadata_dictionary.csv documents a diagnosis variable that metadata.csv does not contain. The four real columns are patient_id, basal_pattern, sudden_death and brugada.

No age, sex or ancestry is published, so no demographic balance check is possible.