St Petersburg INCART 12-Lead Arrhythmia Database

75 half-hour recordings extracted from 32 Holter records, collected at the St Petersburg Institute of Cardiological Technics from patients being tested for coronary artery disease. Each record is 12 standard leads, 257 Hz, exactly 1800 s. The dataset’s primary product is not a record-level diagnosis but 175,907 reference beat annotations across ten beat types, placed by an algorithm and then corrected manually.

Patient grouping is the thing that matters here, and ECGBench enforces it. Only 2 of the 32 patients contributed a single record; 18 contributed two, 11 contributed three and one contributed four. The concrete danger is easy to show: 3,166 of the 3,174 right-bundle-branch-block beats come from one patient (patient08, records I16 and I17, where essentially every beat is RBBB). Split those two records across train and test and an RBBB classifier scores well by memorising one person. Folds here are grouped by patient, and no patient spans a fold.

Records are long — 1800 s, about 44 MB each in memory. They cannot be batched as they are; take a fixed window=(start, length), or use batch_size=1. Because window= is read at load time rather than cropped afterwards, it also cuts per-record load time from ~106 ms to ~8 ms.

Gain varies between records (240 to 1063 adu/mV, uniform across the 12 leads within a record). wfdb divides by the per-record gain, so signals reach ECGBench in millivolts and nothing downstream needs to care.

Lead names are uppercase AVR/AVL/AVF, the PTB-XL spelling. Lead matching is case-insensitive, so leads=["aVL"] still works.

Every figure on this page was recomputed from the 75 headers and 75 .atr files. Two disagreements with the shipped README, both worth knowing:

• Demographics. The README states “17 men and 15 women, aged 18-80; mean age: 58”. The headers give 18 men and 14 women, ages 18–80, mean 56.2 (computed per patient, not per record — records are unevenly distributed across patients, so a per-record mean is a different number).
• Gain range. The README says gains vary “from 250 to 1100”; the headers actually run 240 to 1063.

The beat total does reconcile: the README’s “over 175,000 beat annotations” against a recomputed 175,907. The .atr files hold 175,919 annotations in total — the extra 12 are + rhythm-change markers, which are not beats and are exposed separately as n_rhythm_changes.

The README’s diagnosis table cannot be reproduced from the <diagnoses> field alone, and its categories sum past 32 patients. It mixes two sources. WPW (2 patients), atrial fibrillation, AV nodal block and bundle branch block appear only in the per-record free-text findings, never in <diagnoses>, which is why the table above looks so different. Use record_features alongside diagnosis; neither alone is the full label.

14 of the 32 patients (34 of 75 records) carry no diagnosis at all — the <diagnoses> token is absent from those headers entirely, rather than present and empty. ECGBench labels them UNKNOWN.

The stratification label is not a clinical grouping. With 32 patients over 10 folds, and pooling any class holding fewer than 10 patients (patients, not records, because folds are grouped by patient), stratify_class collapses to two values: UNKNOWN (34 records) and OTHER (41). It exists so the folds are defined. Train on diagnosis, record_features and the beat counts.

Two shipped text files, files-patients-diagnoses.txt and record-descriptions.txt, restate the header content. Both were checked against all 75 headers and agree on every record — no patient-id, diagnosis or findings disagreement — so ECGBench reads the headers only.

amplitude_range_mv is [-30, 30] here rather than ECGBench’s usual [-10, 10], and that is a measured decision, not a convenience. In 30-minute ambulatory Holter data, amplitudes past 10 mV are ordinary: 29 of the 75 records contain at least one sample beyond ±10 mV, and one (I31) runs 12–20 mV across all twelve leads for the full half hour with entirely normal per-lead medians. At ±10 the check would discard 39% of a 75-record dataset.

At ±30 it does the job it is there for — catching four records with a lead pinned at the 16-bit rail for most or all of the recording, which missing_leads cannot see because those leads are railed rather than zero:

I69 is excluded too, on a ±43.24 mV excursion with no railed lead. No single threshold separates it from the four above, because its peak sits higher than their ±31.5 mV rail — the rail depends on each record’s own gain.

There are no NaN samples and no all-zero leads anywhere in this release.

74 of these 75 records also sit inside the PhysioNet/CinC Challenge 2021 release (renamed I0001–I0075, with I36 absent), and inside the 2020 one. Prefer this original. The challenge re-quantised the signals to a gain of 1000 adu/mV, which cannot represent this dataset’s amplitude range — records here reach ±107 mV — so the challenge copy carries roughly 467,000 NaN samples across those 74 records, where the original has none. The challenge copies also drop the .atr beat annotations and the patient numbers, which is what makes grouped folds possible here.