MIT-BIH Atrial Fibrillation Database

The reference dataset for AF detection, and for onset/offset work in particular. 25 long-term two-lead Holter recordings of subjects with atrial fibrillation, digitised at 250 Hz, each running 10 h 13.7 min, with 623 manually reviewed rhythm episodes marking where AF starts and stops over 254.7 hours of recording.

Every subject here has AF, so the label that matters is burden, not diagnosis. AF burden — the fraction of annotated time in atrial fibrillation or flutter — ranges from 0.24% (05091, 43 s of AF in ten hours) to 100% (07162 and 07859, a single episode each covering the whole record). That spread is the dataset: a detector can be evaluated against records where AF is a needle in a haystack and records where it never stops. There is no non-AF control group, so nothing here measures specificity in a general population.

Two of the 25 records ship no ECG. 00735 and 03665 are in the release’s own RECORDS file and carry real rhythm annotations, but their signal files were never published — notes.txt says “Signals unavailable”. ECGBench keeps them, so original has all 25 and matches the published count, and marks them invalid so clean has the 23 that can be read. See the validation section.

The two channels are not named leads. The headers call them ECG1 and ECG2 and state no electrode placement anywhere in the release. Unlike its sibling MIT-BIH Arrhythmia Database, which documents MLII/V1, this one gives you two channel positions and no anatomy. Do not assume they are MLII and V1.

Records are 10 hours long — 9,205,760 samples, about 74 MB of float32 per record. Batching needs a window=(start, length), which is read at load time rather than cropped afterwards. Length is not uniform: 06453 stops at 8,325,000 samples, so a window must fit inside that to work on every record.

Every figure on this page was recomputed from the 25 headers and the .atr and .qrs annotation files, after verifying all 155 shipped files against the release’s own SHA256SUMS.txt — all 155 match.

• 25 records, 23 with signals. PhysioNet’s description says 25 and that is what the RECORDS file lists; the two without a .dat are named in notes.txt. ECGBench reports both numbers rather than picking one.
• 254.73 hours annotated, of which 234.28 hours have signal behind them. The gap is the two annotation-only records.
• 1,221,559 beat annotations, from the .qrs files. These are unaudited automatic detections, not reviewed ground truth: every symbol in every .qrs file is N, because the detector emits N for everything it finds. Manually corrected beats (.qrsc) exist for two records — 05091, which notes.txt names, and 07859, whose file was added in 2014, long after that file was written, which is why it is not listed there.
• dominant_rhythm is not the stratification label and neither is af_class. Dominant rhythm by duration is N for 14 records, AFIB for 10 and J for 1 (03665, which spends 52% of its time in junctional rhythm). Folds are stratified on a binary cut at 20% AF burden — 14 records above, 11 below — because StratifiedKFold needs at least as many members per class as there are folds, and with 25 records over 10 folds neither the 3-class af_class (3 sustained) nor dominant_rhythm (1 J) can be spread across them. Train on af_burden, dominant_rhythm or the rhythm_secs_* columns; stratify_class is for fold construction only.

The rhythm table measures time, not markers. A rhythm annotation opens an episode that runs to the next one, so episode count and duration say different things — 04043 has 82 AFIB episodes covering 21.5% of its record, while 07162 has one covering 100% of its.

There are no demographics at all. This release’s headers carry no comments: no age, no sex, no medications, no clinical description, and no subject or tape identifier. That is why patient_id_column is null — one record per subject is the most that can be asserted, so folds are stratified but ungrouped. Mean heart rate, derived here from the .qrs RR intervals, runs 61.0–103.2 bpm across the 25.

Beat annotation ends at the nominal ten hours — sample ~9,000,000 — while the .dat files hold 9,205,760 samples. So the last 823 seconds of each full-length record carries signal that nobody annotated (1,692 s for 04048, whose detector output stops earlier still, and 100 s for the short record 06453). A window= reaching into that tail returns waveform with no reference behind it, which is fine for unsupervised work and wrong for evaluation.

unannotated_tail_secs in the labels reports it per record. The single exception is 07859’s 2014 .qrsc, which does cover its whole record.

Every signal line in every header declares a gain of 0, which is WFDB’s code for “uncalibrated”. wfdb therefore falls back to its default of 200 adu/mV and reports the samples as millivolts, so ECGBench’s signal_unit_scale is 1.0 and nothing is rescaled.

PhysioNet’s description says 12-bit over a ±10 mV range, which implies a true gain of 4096/20 = 204.8 adu/mV — 2.4% finer than the 200 that wfdb applies. ECGBench keeps the 200 for two reasons: it is what the files themselves lead every tool to, so published work on this database is on that scale and a silent 2.4% divergence would make ECGBench the odd one out; and 204.8 comes from one line of prose, not from anything in the data. No record reaches the 12-bit rail — the extreme sample anywhere in the release is 9.065 mV (08455, ECG1) against a full scale of 10.24 — so the files cannot arbitrate between the two figures.

Waveform shape is unaffected either way. Absolute calibration carries a nominal 2.4% uncertainty, which matters for amplitude thresholds and not for rhythm.

The only two records that fail validation are the two with no signal files. Their headers read 00735 0 250 0 — zero signals, zero samples — and wfdb.rdrecord rejects that with sampto must be greater than sampfrom, which is what lands in quality_issues. It means “this header declares an empty record”, not “this file is damaged”. Both records’ labels are real and available through load_labels; only their waveforms do not exist.

One consequence is worth stating plainly: the original version is not iterable for this dataset. There is no signal to return for those two records, so ds[i] raises WindowOutOfRangeError (“has 0 samples”) and any DataLoader over original fails on the batch containing them. Take original to see what was excluded and why; take clean for anything that reads waveforms. Every other ECGBench dataset’s original version holds records that are flagged but readable — this one does not.

Nothing else fails, and two checks are worth explaining because they cannot:

• truncated_signal is disabled, by leaving expected_samples empty. 06453 holds 8,325,000 samples against the other 22 records’ 9,205,760, and notes.txt explains why — “Recording ends after about 9 hours, 15 minutes”. A 9,205,760 threshold would drop a sound record as truncated.
• amplitude_range_mv is [-10.24, 10.235], the 12-bit rail computed from the hardware (adc_zero 0 at a gain of 200 puts every possible sample in [−2048, 2047] adu). Unlike the MIT-BIH Arrhythmia Database, nothing in this release reaches it and no record clips, so the check cannot fire on AFDB 1.0.0. That is the intended state: any tighter threshold would just exclude the highest-amplitude records for having high amplitude. What it guards is a mis-scaled copy — microvolts, or a re-release with a declared gain — which would exceed it by orders of magnitude on the first record.

There are no NaN samples and no flat or all-zero leads anywhere in the 23 readable records.

00735 is not 735. This is the first dataset in ECGBench whose record ids do not survive a naive CSV round-trip, and adding it is what put DatasetConfig.identifier_dtypes() in front of every metadata and fold-CSV read. Read as numbers, the ids lose their leading zeros, the label join misses, and data_path / "735" is not a file — so every record fails corrupt_header for a reason nothing in the traceback mentions. If you read the published fold CSVs yourself, pass dtype={"record_name": str, "signal_path": str}.

• old/ (25 .atr files). The pre-2001 revisions of every rhythm annotation file, under the same names as the current ones. ECGBench reads annotations from the dataset root only, and takes its record list from the shipped RECORDS file, so these cannot leak in.
• .hea-, .atr-, .qrs- files in the root. Superseded copies, ignored for the same reason. 07859’s .qrs was itself revised in 2014, and 07859.qrs- is what it replaced.

The shipped ANNOTATORS file lists three annotators — atr (reference rhythm), qrs (unaudited beats) and qrsc (corrected beats) — and ECGBench exposes all three.

No recordings are shared with the MIT-BIH Arrhythmia Database, despite the shared institution. Checked from the annotation files rather than assumed: neither release carries a subject identifier that would join, so the check compared RR-interval sequences, which are commensurable across the two sampling rates. At a signature of 20 consecutive RR intervals quantised to 8 ms, 0 of MITDB’s 48 records share a single sequence with any AFDB record — against a positive control that re-finds an AFDB record in the AFDB pool at 100% while cross-record leakage between known-distinct AFDB records stays at 0.005%. Subject-level overlap cannot be checked at all, because AFDB ships no subject identifiers.