MIT-BIH Arrhythmia Database

The reference arrhythmia benchmark. 48 half-hour excerpts of two-channel ambulatory ECG, digitised at 360 Hz from Holter tapes recorded at the Beth Israel Hospital Arrhythmia Laboratory between 1975 and 1979, carrying 109,494 beat annotations across fifteen types plus rhythm episodes, signal-quality changes and artefact markers. Each record was annotated independently by two or more cardiologists and the disagreements resolved.

The database is two deliberate halves, and ECGBench stratifies on that. Records 100–124 (23 of them) were chosen at random from a pool of over 4000 recordings, to be representative of routine clinical material. Records 200–234 (25) were selected from the same pool for rare but clinically important phenomena. With only 48 records over 10 folds, a fold drawn without regard to that split can easily be all one half — the difference between a test set with ventricular flutter in it and one without.

The two leads are not the same two in every record. This is the one dataset in the catalogue where records store the same number of leads under different names, so nothing about a signal’s shape reveals which lead is which. See “Lead layout” below — it changes how you must call leads=.

Records are long — 650,000 samples, about 5 MB each. Take a fixed window=(start, length), which is read at load time rather than cropped afterwards, or use batch_size=1.

Folds are grouped by analog tape. 48 records come from 47 subjects: records 201 and 202 were cut from the same tape (1960), which the shipped directory states and the tape number in the header confirms. One subject out of 47 — but it is the only one there is, and ungrouped those two land in different folds most of the time.

In most records the upper signal is a modified limb lead II (MLII), taken from chest electrodes, and the lower a modified V1. Not always. Counted from all 48 headers:

Records 102 and 104 have no MLII at all: surgical dressings made a modified II impossible, so V5 was used for the upper signal. Record 114 has the predominant pair the wrong way round, which the source documents as something that “happens occasionally in clinical practice” and which arrhythmia detectors should cope with.

Every one of the 48 records stores exactly 2 leads, so alternate_lead_names — which maps a lead count to a layout — cannot express any of this. The config declares record_lead_layouts instead, and ECGDataset then resolves the requested lead names against each record’s own header:

ds = ECGDataset("mitdb", split="train", data_path=..., leads=["MLII"])
ds[0]["signal"]    # record 100: MLII, read from position 0
# record 114: MLII, read from position 1 — an index would have returned V5
# record 102: raises ValueError; it stores V5/V2 and has no MLII to return

Without this, signal[0] is a limb-type lead in 46 records and a chest lead in 2, and leads=["MLII"] returns V5 for three of them with no error.

Every figure on this page was recomputed from the 48 headers and 48 .atr files, after verifying all 226 signal and annotation files against the release’s own SHA256SUMS.txt. Most of them reconcile with the published record exactly, which is unusual in this catalogue and worth stating:

• 109,494 beats is the figure the database has published since 1980, and it is what the fifteen beat symbols sum to here. The .atr files hold 112,647 annotations in total; the other 3,153 are markers that are not beats — 1,291 rhythm changes, 616 signal-quality changes, 472 ventricular-flutter waves, 437 comment annotations, 193 non-conducted P waves, 132 isolated artefacts and 12 flutter-episode delimiters. ECGBench counts them in their own columns and keeps them out of n_beats; adding any of them is how a wrong total gets quoted.
• 47 subjects, 25 men aged 32–89 and 22 women aged 23–89, exactly as the shipped directory states. Age is recorded as -1 on records 103 and 219, which ECGBench returns as NaN rather than as a number — the mean age of 63.6 is therefore over 45 known subjects, not 47.
• The subject identity had to be derived. No header names a subject; the third field of the first comment line is the analog tape, and it is distinct for every record except 201 and 202, which share 1960 — exactly the pair the directory names. The fourth field is the Del Mar Avionics recorder: grouping by it reproduces the directory’s recorder table to the record, including record 208 being the one whose recorder was never traced, and the one whose field reads N/A. Both are exposed (patient_id, recorder); the second is a real confounder, since one recorder accounts for 13 of the 48 records.

The rhythm table measures time, not markers. A rhythm annotation opens an episode that runs until the next one, so a record with two AFIB markers and two sinus markers may be 99% sinus. The 1,444.3 minutes annotated is 9 seconds short of the 1,444.5 minutes the 48 records contain, because in a few records the first rhythm annotation comes a beat or two after the start.

Beat classes are not a record-level label, and they are wildly imbalanced. Almost every record carries several beat types, so there is no single beat class per record; and several classes live in one or two records (all 472 ventricular-flutter waves are in record 207, 427 of the 428 missed-beat markers are in record 231, and the 7,028 paced beats come from 4 records). The stratify_class used for folds is neither — it is the random/selected halves, and it is for fold construction only.

amplitude_range_mv is [-5.11, 5.11] here rather than ECGBench’s usual [-10, 10], and it is set from the hardware rather than from physiology. The recordings are format 212 at 11 bits with a gain of 200 adu/mV and adc_zero 1024, so the full scale is exactly −5.120 to +5.115 mV — no sample in this database can lie outside it, and any range wider than that is a check that provably never fires.

At the rail it does something real. Three records reach it, i.e. clip:

The next-highest peak anywhere in the release is 4.94 mV (record 200), so the threshold separates the saturating records from every other one cleanly. The clipping is brief — well under a second in total per record — so for beat detection these records are perfectly usable; take the original version if you want them, where they are present and flagged.

There are no NaN samples and no flat or all-zero leads anywhere in this release, and all 48 records are exactly 650,000 samples, so the truncation check is enabled rather than skipped.

The download contains three things ECGBench deliberately ignores, all checked against the files rather than assumed:

• x_mitdb/ (23 records). Each is the first 600 s of a record already in the release, produced by xform with the baseline removed — correlation 1.0 against its parent over all 216,000 samples. Including them would put the same recording into the partition twice. ECGBench takes its record list from the shipped RECORDS file, which names the 48 and none of these.
• 102-0.atr. 2,192 annotations with the same symbol counts as 102.atr, and no header of its own.
• 108.at_. A superseded copy of 108.atr.

The shipped ANNOTATORS file lists atr and nothing else, which is what ECGBench reads.

Note also that the two channels are not synchronous: the directory records skew of up to 40 ms between them, part fixed per recorder and part variable from tape wobble. Anything comparing the channels sample-for-sample has to allow for it.