MIT-BIH Normal Sinus Rhythm Database

The reference normal cohort. 18 full-day two-lead Holter recordings of subjects referred to the Arrhythmia Laboratory at Boston’s Beth Israel Hospital who, on review, were found to have had no significant arrhythmias — 5 men aged 26 to 45 and 13 women aged 20 to 50. 437.5 hours of signal at 128 Hz, and 1,729,629 reference beats of which 127 are not normal.

There is no diagnostic label here, and that is the point. The release ships no rhythm annotations at all, so cohort_label is normal_sinus_rhythm for every record. Use this database as a control group, a negative class, a pretraining corpus, or the baseline for normal heart-rate variability — not as a classification task. Folds are stratified on the subject’s sex instead, which is the one axis PhysioNet documents about this cohort.

Beat annotation stops long before the signal does, and nothing says so. It covers 79.5% (19090) to 95.7% (16539) of each record. The remaining 52.9 hours — 12.1% of the release — carry waveform with no reference behind it, one to five hours at the end of every record. This is the largest trap in the dataset and it has its own section below.

Records are a full day: 10,659,840 to 11,960,320 samples, 85–96 MB of float32 each. Batching needs a window=(start, length), which is read at load time rather than cropped afterwards. Length is not uniform — 23.13 h to 25.96 h — so a window sized for one record need not fit another.

The two channels are not named leads. The headers call them ECG1 and ECG2 and the release states no electrode placement anywhere. Its sibling MIT-BIH Arrhythmia Database, from the same laboratory, does document MLII/V1; this one gives you two channel positions and no anatomy. Do not carry the mitdb naming across.

Every figure on this page was recomputed from the 18 headers and .atr files, after verifying the shipped data against the release’s own SHA256SUMS.txt — all 74 dataset files match. (That file also lists 22 .hea-, .hea-- and .atr- entries, which are superseded backup copies PhysioNet keeps beside the current revisions. They are absent from this copy, and the record list comes from the shipped RECORDS file, so they could not enter the partition either way.)

The published cohort description reproduces exactly, which is unusual enough in this catalogue to be worth saying: PhysioNet states “5 men, aged 26 to 45, and 13 women, aged 20 to 50”, and the header comments give precisely that — men at 26, 32, 34, 42, 45 and women at 20, 20, 26, 28, 32, 32, 34, 35, 38, 38, 41, 45, 50. There is no changelog and no discrepancy to explain.

Ectopy exists, but only just. 127 of 1,729,629 beats are not normal — 7.3 per 100,000. Broken down: 91 supraventricular premature (S), 26 ventricular premature (V), 8 fusion (F) and 2 nodal premature (J). Three records (16795, 18184, 19140) have none at all; the worst two, 16265 and 16773, have 27 each. That is what “no significant arrhythmias” means operationally. There is no usable ectopy class here — n_ectopic_beats is exposed because “how clean is clean” is a real question, not because it can be trained on.

cohort_label is a constant, and it is not the stratification label. It records PhysioNet’s assertion about the cohort so that a user combining this database with an arrhythmia one has a record-level class to join on. Nothing in the files derives it. Folds are stratified on sex (13 F / 5 M) — see the fold section below. Train on mean_hr_bpm, sdnn_ms, rmssd_ms, the beat_* counts or the signal-quality seconds; never on stratify_class.

The HRV figures are descriptive, not a result. mean_hr_bpm, sdnn_ms and rmssd_ms are computed over RR intervals in [0.3 s, 2.0 s] from the reference beats. That filter is load-bearing: without it the multi-hour unannotated gaps below enter as single enormous “RR intervals” and every figure is meaningless. 894 intervals are rejected across the release. These are whole-record summaries over ~24 h of mixed activity and sleep, not the segmented, artefact-corrected analysis an HRV study would run.

This is the thing to know before choosing a window. Beat annotation covers 79.5% to 95.7% of each record and then simply stops; the .dat file keeps going for another 1.1 to 5.0 hours, and neither the header nor the annotation file says anything about it.

Five records (19088, 19090, 19093, 19140, 19830) additionally open with an unannotated head of 23–34 seconds. Every other record’s first beat is within a second of the start.

A window reaching into either region returns waveform with nothing to score it against. That is fine for self-supervised or unsupervised work and wrong for evaluating a beat detector. annotated_secs, unannotated_head_secs, unannotated_tail_secs and annotated_fraction in the labels report it per record, so a supervised window can be kept inside the annotated span.

The shipped ANNOTATORS file promises “reference beat and signal quality annotations”, and the second half is easy to miss. The ~ annotations mark quality transitions, and their WFDB subtype is a bitmask over the two channels — 0 clean, 1 ECG1 noisy, 2 ECG2 noisy, 3 both. Each transition opens an interval running to the next one, so ECGBench exposes it as seconds per state per record, not as a marker count: a one-second glitch and a three-hour noisy stretch are one marker each.

Across the release 98.61% of the recorded time is annotated clean — 1.56 h of ECG1-only noise, 2.68 h of ECG2-only and 1.84 h of both. But the average hides the spread: 16272 is 9.60% noisy and 16786 is 0.23%, a factor of 40. The | isolated-artifact marker varies even more, spanning three orders of magnitude (52 in 16273, 30,782 in 16773), so a per-record metric is not comparable across records without controlling for it.

The span before a record’s first ~ is counted as clean. That is checked rather than assumed: in all 18 records the first ~ is a transition into noise (subtype 1, 2 or 3), never a return to clean, so nothing before it was ever marked otherwise.

All 18 records pass every check, so original and clean hold the same 18 rows. There are no NaN samples, no flat or all-zero leads and no unreadable header anywhere. Two checks are worth explaining because they cannot fire:

• truncated_signal is disabled, by leaving expected_samples empty. Records run 10,659,840 (17052) to 11,960,320 samples (16483, 18177) and every one is a complete full-day recording, so any single threshold would drop sound records as truncated. Omitting the rate disables the check rather than making it fire — the same escape hatch ptbdb, afdb and ltafdb use.
• amplitude_range_mv is [-10.24, 10.235], the 12-bit rail computed from the hardware (adc_zero 0 at wfdb’s fallback gain of 200 puts every possible sample in [−2048, 2047] adu). The extreme sample anywhere in the release is ±5.115 mV, half of that, so the check cannot fire on NSRDB 1.0.0. 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.

One thing no threshold in that range would catch, so it is recorded here: record 16272 clips. It saturates flat at ±1023 adu (±5.115 mV) for 1,246 samples on ECG1 and 2 on ECG2; 18184 does for 3 samples on ECG2. 1023 is the 11-bit rail — half the span the header’s 12-bit adc_res declares — so the recorder’s usable range was narrower than the header implies. It is 0.01% of one record, and 16272 is also the noisiest record in the release at 9.60%.

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.

This is the same situation as AFDB, with one simplification: PhysioNet’s description of AFDB states a ±10 mV range that implies a different gain (204.8 adu/mV) and has to be reconciled. NSRDB’s description states no millivolt range at all, so the 200 that wfdb applies is the only calibration anything in or around this release supports. Waveform shape is unaffected either way; absolute calibration rests on wfdb’s default.

Two consequences of the arithmetic, stated rather than left to be discovered:

• The default split leaves one record in val and one in test. ECGBench’s convention is folds 1–8 → train, 9 → val, 10 → test, and 18 records over 10 folds gives eight folds of two and two folds of one. So train holds 16 records, val holds 1 and test holds 1. For anything that needs a real evaluation set, use cross-validation: split=None with fold_numbers=[...] selects by fold from folds.csv and ignores the default layout.
• Half the folds contain no man. The five men land in five different folds (16265, 16483, 16773, 19090, 19093 in folds 2, 5, 3, 1 and 4), so folds 9 and 10 — val and test — are both female. Five records cannot be spread over ten folds; stratification keeps them apart rather than letting them clump, which is the most it can do here.

Why sex, and not something clinical? There is nothing clinical to use. cohort_label is one value for all 18 records, and StratifiedKFold requires at least one class holding n_folds members. Sex gives 13/5 and clears it. A median cut on age gives 10/8 and clears it by nothing. Anything ectopy-based fails outright — three records have no ectopic beats at all and the rest differ by single-digit counts out of ~100,000 beats. sklearn warns that the smallest class has 5 members; that warning is expected and correct.

Folds are ungrouped. The header comment holds age and sex and nothing else — no tape number, no recorder, no subject code — and PhysioNet describes 18 recordings from 18 subjects, so one record per subject is the most that can be asserted.

NSRDB, the MIT-BIH Arrhythmia Database and the MIT-BIH Atrial Fibrillation Database all come from the Beth Israel Hospital Arrhythmia Laboratory’s Holter collection, and none of them 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 — the same method used for LTAFDB.

Against controls that make a null result mean something — a positive control re-finding each NSRDB record in its own pool at 100%, and a negative control of each record against the pool of the other 17 known-distinct subjects at a maximum of 0.0000% — the result is:

• 0 of 18 NSRDB records share any sequence with the MIT-BIH Arrhythmia Database pool (highest 0.0000%);
• 0 of 18 share any with the MIT-BIH AFDB pool (highest 0.0000%);
• 0 of 18 share any with the Long-Term AF Database pool in substance: five records register a non-zero hit, but the highest is 0.0176%, and repeating the check at a 30-interval signature drops it to one record at 0.0062%. A genuinely shared recording stays near 100% as the signature lengthens; these decay, which is what chance collisions against an 8.8-million-signature pool do.

No related: edge is declared on those grounds. 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.

Note also that PhysioNet’s MIT-BIH Normal Sinus Rhythm RR Interval Database (nsr2db, 54 subjects) is a different database, not a derived layer over this one, and is not in this catalogue.