Long-Term AF Database (LTAFDB)

1,960.6 hours of annotated two-lead Holter from 84 subjects with atrial fibrillation — about eight times the recorded time of the MIT-BIH Atrial Fibrillation Database, and the largest annotated AF recording set in this catalogue. Records run 6.1 to 26.4 hours, typically 24–25, digitised at 128 Hz.

The annotations are what make it worth having. The .atr files hold 8,995,973 typed beat annotations and 53,704 rhythm episodes in nine codes, produced by MEDICALgorithmics’ PocketECG algorithm and then manually verified by their ECG technicians. They were contributed to PhysioNet in 2012, four years after the signals. AFDB, by contrast, ships 623 rhythm episodes and unaudited beats. So AFDB is the AF-detection benchmark and this is what you use when you need reference beats at scale, or AF burden measured over a whole day rather than ten hours.

AF burden is the label, not a diagnosis. 83 of the 84 subjects have annotated AF, and burden runs the full range from 0% (record 30, the only AF-free record and also the shortest at 6.1 h) to 100%. The distribution is strongly bimodal — 33 records are in AF essentially throughout and 18 spend under 5% of the recording there — which is why af_class splits three ways rather than two.

There is no atrial flutter code. AFDB annotates N/AFIB/AFL/J; this release annotates nine codes and AFL is not among them, while ventricular bigeminy, trigeminy, atrial bigeminy, sinus bradycardia, SVT, VT and idioventricular rhythm are. af_burden here is AFIB alone, so the two databases’ identically named columns are computed over different code sets.

The two channels are not named leads — and unlike AFDB they are not even numbered. Every header calls both channels ECG, the same string twice, and states no electrode placement anywhere in the release. Two identically named channels cannot be told apart by name, so ECGBench declares the positional names ECG1/ECG2 and says so here. Do not read them as MLII/V1 by analogy with the MIT-BIH Arrhythmia Database.

Records are a full day long, so batching needs window=(start, length), which is read at load time rather than cropped afterwards. Length is very much not uniform: 55 distinct record lengths, from 2,826,240 samples (record 30) to 12,142,080 (record 70), so a window must fit inside the shortest.

Every figure on this page was recomputed from the 84 headers and the .atr and .qrs annotation files, after verifying all 339 shipped files against the release’s own SHA256SUMS.txt — all 339 match.

They were then checked against the release’s own summary tables (tables.shtml), cell by cell: all 336 beat counts (84 records × N/A/V/Q) and all 756 rhythm cells (84 × nine codes, episode count and duration) agree. Two differences are worth recording:

• Record 20’s published AFIB duration is 1.1 s short. tables.shtml gives 24:19:08 (87,548 s) where the shipped header yields 87,549.1. The landing page thanks Mariano Llamedo Soria “for reporting an error in the original version of 20.hea, and for providing a correction incorporated in the current version” — the table was generated against the pre-correction header, which was 144 samples shorter. The current header is the authority, so ECGBench reports 87,549.1.
• Record 30’s comment-annotation count. Its .atr ends with the WFDB file terminator at sample 4,198,064, while the record holds only 2,826,240 — 2.98 hours past the end of the data. That marker’s position is not a claim about the signal; ECGBench excludes out-of-range annotations from every measurement, which is what makes record 30’s count agree with the published 1 rather than reading 2.

Two more things the numbers do not say on their own:

• af_burden’s denominator is annotated rhythm time, not record time. The first rhythm annotation sits a little way into each record — 47.5 s into record 20 — and nothing classifies what precedes it. Across the release that lead-in is 25.7 of the 1,960.6 recorded hours, which is the gap between the 1,934.92 h in the rhythm table and the 1,960.6 h of signal. rhythm_annotated_secs states the denominator per record.
• 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 — record 74 has 1,044 AFIB episodes covering 23.5 h, while record 12 has one covering 24.1 h.

There are no demographics at all. This release’s headers carry no comment lines: 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 .atr RR intervals, runs 40.4–129.8 bpm across the 84.

The shipped ANNOTATORS file lists two annotators, and they are not interchangeable:

• .atr — reference beat and rhythm annotations. Beats are typed (N/A/V/Q), rhythm changes carry one of nine codes, and " comment markers flag missed beats (651) and pauses (5,224). Algorithm output, manually verified. Everything on this page is derived from these.
• .qrs — unaudited sqrs detections. 8,611,567 of them, every one labelled N whatever it actually is, plus 2,549 artifact markers. Useful as a detector baseline; never add them to the .atr counts.

The .qrs files also carry 81 hand-placed T markers recording spontaneous ends of AF episodes lasting a minute or more, inserted by Steven Swiryn and George Moody. The AF Termination Challenge Database’s 80 one-minute excerpts were cut around them. They exist only in records 00–75 — all 24 records that carry any are in that range, and the 100- and 200-series carry none — so a model trained on terminations sees roughly half the database.

T means something else entirely in an .atr rhythm code: ventricular trigeminy. Same letter, different annotator, unrelated meanings.

Both annotators stop at the same place in each record, so this is a property of the recording rather than of one annotator. The median record’s beat annotations stop 4.9 s from the end — but:

unannotated_tail_secs in the labels reports it per record. A window= reaching into that tail returns waveform with no reference behind it, which is fine for unsupervised work and wrong for evaluation.

Rhythm durations nonetheless run to the end of the signal, because the last episode has no annotation after it to close it and PhysioNet’s own tables close it at the record end. ECGBench follows that convention — which is what makes its figures reproduce those tables — so a record like 117 attributes its final 8 unannotated hours to whatever rhythm was running at 15.9 h.

Unlike AFDB, whose every header declares a gain of 0 (WFDB’s “uncalibrated”) and leans on wfdb’s 200 adu/mV fallback, LTAFDB’s headers carry 50 distinct measured gains, sometimes different for the two channels of one record (record 100 is 88.968 and 131.062). wfdb applies each header’s own gain, so samples arrive in genuine millivolts and signal_unit_scale is 1.0. There is no calibration argument to have here.

With one exception, which ECGBench flags and does not correct. Record 62’s ECG1 declares 1123.6 adu/mV, where the other 167 signal lines run 75.0188 to 222.222 — a 5.5× outlier with nothing in between. Its raw swing is entirely normal (2,756 adu peak-to-peak, against a release median of 1,220 and a maximum of 2,808), so that gain turns the largest raw excursion in the release into its smallest calibrated one: 2.45 mV peak-to-peak against a release median of 6.65.

ECGBench reports what the header declares, because a silently corrected record would disagree with every other tool reading the same file. But anything comparing absolute amplitudes across records should exclude record 62’s ECG1 or rescale it. Its ECG2 (202.429) is unaffected, and adc_gains in the labels carries the pair for every record.

Nothing fails validation: all 84 records read, none has a NaN sample, and no lead is flat or all-zero. Two checks are worth explaining because they cannot fire on this release:

• truncated_signal is disabled, by leaving expected_samples empty. There are 55 distinct record lengths spanning 2,826,240 to 12,142,080 samples, so no single threshold can distinguish a truncated record from a short one, and any threshold would drop sound records. Every .dat file holds exactly the number of samples its header declares — checked for all 168 channels.
• amplitude_range_mv is [-27.3, 27.3], the 12-bit rail computed from the hardware. PhysioNet describes 12 bits over a 20 mV range, so a sample is confined to ±2048 adu — and the raw files bear that out, with every sample of all 168 channels inside [−1497, 1398]. In millivolts that rail moves with the header gain, and the loosest gain (75.0188 adu/mV) puts it at ±27.3 mV; a single range has to accommodate the loosest record or it fires on a sound one. The observed extreme anywhere in the release is −10.599 mV (record 100, ECG1) to +11.583 mV (record 34, ECG1), so nothing comes close. What the check guards is a mis-scaled copy — microvolts, or a re-release with the gains dropped — which would exceed it by orders of magnitude on the first record.

00 is not 0. Seven of the 84 record names begin with a zero — 00, 01, 03, 05, 06, 07, 08 — and pandas reads a column of digits as int64, which strips them. The id then no longer names a record, the label join misses, and data_path / "0" is not a file, so every record fails corrupt_header for a reason nothing in the traceback mentions.

This is the second ECGBench dataset to need zero_padded_identifiers: true, after AFDB. If you read the published fold CSVs yourself, pass dtype={"record_name": str, "signal_path": str}.

LTAFDB and the MIT-BIH Atrial Fibrillation Database are both long-term two-lead Holter recordings of AF subjects, and neither ships a subject identifier that would join, so the question was settled from the annotation files rather than assumed. RR intervals in seconds are commensurable across sampling rates (128 Hz here, 250 Hz for AFDB, 360 Hz for MITDB), so the check compares sequences of 20 consecutive RR intervals quantised to 8 ms, on two half-bin-shifted grids.

Against controls that make a null result mean something — a positive control re-finding each LTAFDB record in itself at 100%, and a negative control of each record against the pool of the other 83 known-distinct subjects at a median of 0% and a maximum of 0.15% — the result is:

• 0 of 84 LTAFDB records share any sequence with the MIT-BIH AFDB pool (highest 0.04%);
• 0 of 84 share any with the MIT-BIH Arrhythmia Database pool (highest 0.000%).

No related: edge is declared on those grounds. The AF Termination Challenge Database is a genuine derivative — its 80 one-minute excerpts are cut from records 00–75 of this release — but it is not in this catalogue, so there is nothing to link it to yet.

One limitation is worth stating rather than glossing: the RR signature survives refinement of annotations but not re-detection, so a shared recording annotated by two genuinely different detectors could evade it. Subject-level overlap cannot be checked at all, because none of these releases ships a subject identifier.