MIT-BIH Supraventricular Arrhythmia Database

78 half-hour recordings assembled to supply what the MIT-BIH Arrhythmia Database lacks. That database holds 109,494 beats but only 2,781 supraventricular ones — 2.54%. This one holds 12,198 supraventricular ectopic beats in 184,583, or 6.61%, over 39 hours at 128 Hz. It exists for exactly one reason and the selection shows.

The beat symbol for a supraventricular beat is S here and A in mitdb. All 12,188 supraventricular beats in this release are annotated S; mitdb annotates its 2,546 as A and uses S exactly twice. Concatenating the two databases on the raw symbol trains a model on two disjoint vocabularies for one phenomenon. This is the single most important thing on this page. The label loader exposes aami_N/S/V/F/Q — the AAMI EC57 five-class reduction, under which A, a, J and S all collapse to S — and that is the column to combine on.

The headers carry no comment lines at all. No age, no sex, no subject identifier, no medications, no clinical description — not withheld, simply never released. mitdb has all of these in its header comments and nsrdb has age and sex; this database has none, and PhysioNet does not even state how many subjects its 78 recordings represent. Folds are therefore stratified but ungrouped, and nothing here can be linked to any other database.

There is effectively no rhythm annotation layer. The entire release contains one + annotation — (N at sample 112,143 of record 852 — against 1,291 in mitdb. There is no dominant rhythm to be had, so the record-level label is derived from the beat counts rather than released.

The two channels are not named leads. The headers call them ECG1 and ECG2 and the release states no electrode placement anywhere. Do not carry mitdb’s MLII/V1 naming across — this catalogue’s own entry claimed “MLII + V1 · 360 Hz” before the config was written, and both halves of that were wrong.

Every figure on this page was recomputed from the 78 .atr files and headers, after verifying the shipped data against the release’s own SHA256SUMS.txt — all 416 listed files match, including the 78 .hea- and 24 .atr- files. Those are PhysioNet’s superseded revisions kept beside the current ones (the 24 revised .atr files are dated 2010 against the originals’ 1992), not extra records; the record list comes from the shipped RECORDS file, so they cannot enter the partition.

There is no published table to disagree with. PhysioNet’s description of this database is a single sentence — “78 half-hour ECG recordings chosen to supplement the examples of supraventricular arrhythmias in the MIT-BIH Arrhythmia Database” — and states no beat counts, no subject count, no sampling rate and no lead names. Unusually for this catalogue, then, the only figures that exist are the recomputed ones. The record count matches; everything else on this page is new.

This catalogue’s previous entry was wrong, in two ways worth naming. It described the format as “2-lead (MLII + V1) · 30 min · 360 Hz”. The recordings are 128 Hz, not 360, and the headers name the channels ECG1 and ECG2 with no electrode placement stated anywhere in the release. Both errors are the kind that come from assuming a sibling database’s properties carry across; they did not.

The database is not uniformly supraventricular. sveb_fraction runs from 0.000 to 0.575. Five records — 802, 803, 804, 805, 893 — contain no supraventricular ectopy whatsoever, and 811 contains a single beat. At the other end, record 865 is the only record in this catalogue where ectopic beats outnumber normal ones: 1,818 supraventricular and 235 ventricular against 1,102 normal, in 3,162 beats.

Ventricular ectopy is here too, and is not incidental. 9,943 beats, 5.39% of the release, in 67 of the 78 records — and three records are more ventricular than supraventricular by a wide margin (860 at 30.1% VEB, 879 at 26.3%, 892 at 20.3%). A model trained here on “ectopic vs not” learns both classes whether or not that was intended.

The rare classes cannot support evaluation. 23 fusion beats in 6 records, 9 nodal premature beats in one record (878), one aberrated atrial premature beat (821) and one bundle branch block beat (868). They are exposed for completeness, not as trainable classes.

Rate and variability 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 — 407 intervals are rejected across the release. Mean heart rate spans 47.8 bpm (811) to 137.3 bpm (848), and beats per record 1,436 to 4,259, the same two records at either end. These are whole-record summaries over a rhythm that is by construction ectopic, so on a record like 865 they describe the recording and not the subject’s sinus node.

Because there is nothing else. No demographics, no subject identifiers, no diagnoses, and — with one + annotation in the whole release — no rhythm labels. What differs between records, and what the database was assembled to vary, is how much supraventricular ectopy each one holds.

The band edges are fixed fractions, not quantiles. 1%, 3% and 10% of beats give 21 / 20 / 23 / 14 records, so every band clears the 10 members StratifiedKFold needs and no fold can end up holding only zero-ectopy records. Quantile edges would balance the bands slightly better and would move every record’s label the next time a record is added or an annotation revised; fixed edges reproduce against a re-release, which matters more for a partition that gets published. The values are conventional clinical granularity for ectopic burden rather than anything this release states.

The result is that every one of the ten folds carries all four bands — each fold holds 1–2 high, 2 low, 2–3 minimal and 2–3 moderate records. With 78 records over 10 folds, train holds 64, val 7 and test 7.

sveb_burden is a fold-construction label, and a coarsening of a continuous quantity at that. Train on aami_S, aami_V, aami_N, the beat_* counts, or sveb_fraction itself — never on stratify_class.

Folds are ungrouped, because the release ships no subject identifier of any kind. mitdb can group its records by the analog tape number in the header comment; this database has no header comments. It follows that nothing rules out one person having contributed several of the 78 recordings, and no split here can protect against that.

The shipped ANNOTATORS file promises “reference beat and signal quality annotations”. 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 opens an interval running to the next, so ECGBench exposes it as seconds per state per record, not as a marker count.

Across the release 97.41% of the recorded time is annotated clean — 0.37 h of ECG1-only noise, 0.45 h of ECG2-only and 0.19 h of both, out of 39.00 h. The average hides a wide spread: 35 of the 78 records carry no quality annotation at all, while record 868 is 22.4% noisy, 887 is 19.7% and 854 is 19.2%. The | isolated-artifact marker varies independently — 2,211 across the release, 508 of them in record 865 alone, none in 32 records — so a per-record metric is not comparable across records without controlling for both.

The leading span is not always asserted, and here that assumption is not free. In 39 of the 43 records carrying a ~, the first one is a transition into noise, so everything before it is clean by implication. In the other four it is a transition into clean: 803, 855, 857 and 885. Nothing ever asserted what those spans were. For 855, 857 and 885 they are 1.8 s, 6.6 s and 3.3 s and the point is academic. For record 803 it is 1,555.1 s — 86% of the record. ECGBench counts these spans as clean, which is what WFDB itself does, and reports them in quality_head_unasserted_secs so the assumption is visible rather than buried.

Worth stating because the sibling database traps people here. In MIT-BIH Normal Sinus Rhythm, beat annotation stops one to five hours before the signal does, silently. In this database it does not: the first beat falls between 0.01 s and 1.27 s and the last between 1798.87 s and 1800.00 s in every one of the 78 records, so annotated_fraction never drops below 0.9988.

Any window inside any record therefore has reference annotation behind it. annotated_secs, unannotated_head_secs, unannotated_tail_secs and annotated_fraction are reported anyway, so a re-release that changes this cannot do so silently.

All 78 records pass every check, so original and clean hold the same 78 rows. There are no NaN samples, no flat or all-zero leads and no unreadable header anywhere. Records are exactly 230,400 samples each, uniformly, so unlike afdb and nsrdb the truncated_signal check is switched on rather than disabled.

amplitude_range_mv is [-10.24, 10.235], and it is deliberately not the declared resolution. Every signal line declares 10 bits with adc_zero 0, which would put every sample in [−512, 511] adu = ±2.56 mV. The data does not respect that: samples run to −1125 and +1022 adu, i.e. −5.625 to +5.11 mV, so the declared resolution is nominal and a threshold set from it would fail most of the release. What actually bounds the samples is format 212’s 12 bits — [−2048, 2047] adu at 200 adu/mV.

So the check cannot fire on SVDB 1.0.0, which is the intended state. What it guards is a mis-scaled copy — microvolts, or a re-release declaring a real gain on the uncalibrated records — which would exceed it on the first record.

37 of the 78 records declare a gain of 0 on both signal lines, WFDB’s code for “uncalibrated”; the other 41 declare 200. wfdb substitutes its default of 200 adu/mV wherever it reads 0, so rdrecord reports millivolts for all 78 and adc_gain comes back as 200.0 for every one of them. ECGBench’s signal_unit_scale is therefore 1.0 and nothing is rescaled.

Waveform shape is unaffected either way; absolute calibration of those 37 records rests on wfdb’s default rather than on anything the release states. header_declares_uncalibrated in the labels records which is which, so an analysis that cares about absolute amplitude can exclude them.

This database, the MIT-BIH Arrhythmia Database and the MIT-BIH Normal Sinus Rhythm Database come from the same laboratory and era, 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, 360 Hz for mitdb), so the check compares sequences of 20 consecutive RR intervals quantised to 10 ms.

Against controls that make a null result mean something — a positive control re-finding each SVDB record in its own pool at 100%, and a negative control between different SVDB records topping out at 27 shared sequences — the result is:

• 0 shared sequences between any SVDB record and any MIT-BIH Arrhythmia Database record;
• at most 11 between any SVDB record and any NSRDB record, well under the 27 the negative control shows two genuinely distinct records can reach by chance.

No related: edge is declared on those grounds, matching the other MIT-BIH entries in this catalogue. The limitation is the same one stated there: an 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 — this release ships nothing that identifies a subject.