Sudden Cardiac Death Holter Database

23 complete Holter recordings of patients who died — or nearly died — during the recording. Every subject sustained a ventricular tachyarrhythmia and most had an actual cardiac arrest. 446.6 hours of two-channel signal at 250 Hz, collected mainly in the 1980s in Boston-area hospitals by Scott Greenwald at MIT and published by PhysioNet in 2004 as the first sudden cardiac death database. Every other long-term database in this catalogue describes a rhythm the subject lives with; this one describes the one they did not survive.

Use version="original", not the default clean. 20 of the 23 records contain NaN samples, so they fail nan_values and clean/ holds 3 records with empty val and test. The NaN is not corruption — it is WFDB’s invalid-sample marker for brief analog-tape dropouts — but it is real and it lands in your tensor. This has its own section below and it is the first thing to know.

The terminal event is in a header comment, not an annotation. #vfon: HH:MM:SS gives the onset elapsed from the record start, in 20 of the 23 headers. There is not one [ (VFON) annotation anywhere in the release. Onset lands from 6.0% (record 37) to 98.9% (record 35) of the way through, leaving between 976 s and 85,007 s of signal after it — so no single window= captures the event across records. Window relative to vf_onset_secs.

There are two annotators and they cover different records. .ari is unaudited detector output for all 23 records (1,888,495 beats); .atr is the audited reference for only 12 (849,831 beats), which PhysioNet itself calls “an incomplete set of audited annotation files”. Every beat column is prefixed ari_ or atr_, because an unprefixed n_beats would mean one thing for half the release and another for the rest.

The (AFIB markers are not an atrial fibrillation label. They disagree with the release’s own published rhythm column in both directions — they miss one of the four AF subjects entirely and flag six sinus subjects at 22–36%. Section below.

No two records are the same length, 3,540,000 to 22,627,500 samples — the longest is 6.4x the shortest, where chfdb’s spread is 1.2%. Any window must fit inside 14,160 s.

The two channels are not named leads. Both signal lines of every current header end in the bare description ECG, as in ltafdb; the release states no electrode placement anywhere. ECG1/ECG2 are channel positions.

Every figure on this page was recomputed from the 23 headers and both annotation layers, after verifying the shipped data against the release’s own SHA256SUMS.txt — all 109 files match. That total includes 23 superseded .hea- backup headers, which predate the 2008 revision that renamed both signal descriptions to ECG. The record list comes from the shipped RECORDS file, so none of them can enter the partition.

The published cohort description reproduces exactly. PhysioNet states “18 patients with underlying sinus rhythm (4 with intermittent pacing), 1 who was continuously paced, and 4 with atrial fibrillation”, and the landing page’s clinical table gives precisely that: 14 plain sinus plus records 32, 43, 49 and 51 intermittently paced makes 18; record 40 is the continuously paced one; 35, 36, 37 and 50 are the atrial fibrillation subjects. That agreement is the only external check the clinical table admits, and clinical_frame() raises rather than quietly rebalancing folds if a later edit breaks it.

The landing page’s duration for record 49 is stale. Its table lists 25:01:40 (90,100 s), but 49.hea declares 22,380,957 samples = 89,523.8 s (24:52:03.8) — 576 s less. That header was revised 2017-05-02, where every other header dates from 2008-04-05, and the .dat file size (67,142,872 bytes at 1.5 bytes per sample pair) confirms the shipped count. The file is right and the page was not updated; the superseded 49.hea- still declares the old 22,525,000. Every other record’s listed duration matches its header exactly, so this is the one discrepancy in the release.

Nothing clinical is in the data files. Sex, age, history, medication and underlying rhythm exist only in the “Clinical information” table on the landing page, transcribed into ecgbench/labels/sddb.py as CLINICAL_TABLE. This is weaker provenance than any sibling — chfdb has a header comment, mitdb a shipped directory — and it cannot be recomputed from the data. PhysioNet is explicit about why so much is missing: “Because of the retrospective nature of this collection… Patient information is limited, and sometimes completely unavailable, including data regarding drug regimens and drug dosages.” Age is absent for 4 subjects, sex for 2, and history and medication read “Unknown” for 12 and 13 respectively. Record 37 is 89 — a real recorded value at the boundary of the usual age-ceiling convention, not a censored one.

Six subjects were on digoxin and five on quinidine, agents that alter rate and repolarisation. The medication column is free text and populated for only six records, so treat it as a hint rather than a variable.

ari beats and atr beats are different quantities, not two estimates of one. Where both exist they disagree substantially — record 30 is 131,323 against 127,418 — because one is a detector and the other a corrected reference over a shorter span. Never sum or average them.

ari AF % is detector output, not atrial fibrillation. It has its own section below; the column is on this table because it describes the annotation file, and it is named ari_* for the same reason.

The HRV figures are descriptive, not a result. mean_hr_bpm, sdnn_ms and rmssd_ms come from .ari RR intervals in [0.3 s, 2.0 s] — 17,946 intervals rejected release-wide. They span the moment the subject’s heart stopped working. Averaging heart rate across a terminal ventricular arrhythmia is not an HRV measurement of anything; segment around vf_onset_secs if you want physiology.

cohort_label is a constant and is not the stratification label. All 23 subjects sustained the arrhythmia, so there is no negative class here. Folds are stratified on the underlying rhythm — see the fold section. Train on vf_onset_fraction, the aami_* counts or the HRV columns; never on stratify_class.

This is the one dataset in the catalogue whose clean version is unusable, and the reason is worth understanding rather than working around.

WFDB’s invalid-sample marker in format 212 is digital −2048, and wfdb returns those samples as NaN. There are 201,708 of them across the release, in every record except 31, 33 and 46:

These are short scattered dropouts from 1980s analog Holter tape, not gaps — which is exactly the problem. check_nan_values has no threshold, so all 20 records fail it, and because the three unaffected records all land in train folds:

So clean/val/fold_9.csv and clean/test/fold_10.csv are header-only, and because ECGDataset defaults to version="clean" it fails in one of two ways — neither of which names the real cause:

ECGDataset("sddb", split="val", labels=True)
# ValueError: No record in split 'val' matched a label row.   <- blames the join

ds = ECGDataset("sddb", split="val")     # constructs fine
len(ds)                                  # 0                 <- says nothing
ds[0]                                    # IndexError

The real cause is that the split is empty. Pass version="original".

Why the check was kept anyway. Dropping nan_values from this config would make clean/ equal original/ and leave quality_issues empty for every record — so a user would get NaN in their tensors, and a NaN loss, with no warning from ECGBench at all. A degenerate clean/ fails loudly and is documented on this page; silent NaN does not. The trade-off is real rather than free, and it is recorded in the config and pinned by a test so nobody “fixes” it by silencing it.

Handling it. original/ carries the per-record count in quality_issues; ecgbench.labels.sddb.scan_invalid_samples(data_path) recomputes per-channel counts, fractions and longest runs from any copy. Then either torch.nan_to_num(signal), mask the loss, or drop affected windows — a 10 s window usually contains none, but a whole record will.

The database is named for an event that has no annotation. There is not one [ (VFON) or ] (VFOFF) marker in the release. What there is, in 20 of the 23 headers, is a comment:

30 2 250 22099250 12:00:00
30.dat 212 800 12 0 51 -24065 0 ECG
30.dat 212 800 12 0 145 21051 0 ECG
#Produced by xform_new from record 30, beginning at 26:35.000
#vfon: 07:54:33

#vfon: is elapsed from the start of the record, not a time of day. Reading it as a clock time is wrong in both directions: record 30 starts at 12:00:00 and its onset is 07:54:33, and record 35’s onset is 24:34:56, which is not a clock time at all. vf_onset_secs parses it as elapsed seconds.

Three records have no onset, and the landing page says why: 40 “(paced, no VF)”, 42 “(no VF)”, 49 “(paced, no VF)”. has_vf_onset distinguishes them from a parse failure.

No single window= can capture the event. Onset spans 6.0% to 98.9% of the record, so secs_after_vf_onset runs from 976 s (record 31) to 85,007 s (record 37) — in some records there is barely a quarter-hour of signal after the terminal event, in others most of a day before it. A fixed window from the record start captures it for a couple of records and misses it for the rest. Window per record instead:

from ecgbench import ECGDataset
from ecgbench.labels import load_labels

labels = load_labels("sddb", "/path/to/sddb/1.0.0/")
onset = int(labels.loc["30", "vf_onset_secs"] * 250)

ds = ECGDataset("sddb", split="train", version="original",
                data_path="/path/to/sddb/1.0.0/",
                window=(onset - 1250, 2500))   # 5 s either side

Because window= is pushed into the reader, that decodes 2,500 samples rather than 22 million.

The shipped ANNOTATORS file names both, and the distinction is not cosmetic:

ari     unaudited beat annotations
atr     reference beat annotations

Every other MIT-BIH-family release in this catalogue has one annotator, so the natural move is to reach for .atr and treat it as the reference. That silently drops 11 records. has_audited_annotation is the column to filter on; PhysioNet describes the audited set as incomplete and invites contributions to finish it.

The two symbol vocabularies are disjoint where it matters, which makes the AAMI EC57 reduction mandatory rather than convenient:

So beat_V alone is not ventricular ectopy in either file: in .ari it misses the 58,820 R-on-T beats, which outnumber plain V three to one, and in .atr it misses nothing ventricular but sits beside 54,725 B and 23,123 / that a naive “normal beats” count would drop. Use ari_aami_V / atr_aami_V, which are the only cross-annotator-comparable counts here.

Two .ari layers are not beats at all. Exactly 50 ? (LEARN) annotations open every one of the 23 records, inside the first 30–65 s — the detector’s start-up phase. And 3,577 s (STCH) markers delimit unaudited ST-segment episodes, with aux notes like (ST0+ and ST0+) for channel and direction, in 22 of 23 records (record 46 has none; record 47 has 363 episodes). Together that is 4,727 annotations a scanner assuming every symbol is a QRS would fold into n_beats. They are counted in ari_n_learning and ari_n_st_episodes instead — and episodes are counted from openings, because record 44’s two episodes are never closed.

The .ari files carry 1,019 + rhythm markers spelling (AFIB and (N, in 22 of the 23 records. They look exactly like afdb’s or ltafdb’s reference rhythm annotations. They are not: they are unaudited detector output, and held against the landing page’s own “Underlying Cardiac Rhythm” column they are wrong in both directions.

Three of the four published AF subjects come out at 72.9–99.4%, which is why the column is tempting. Record 37 comes out at 0.9%, and six sinus subjects at 22–36%. A model trained on this column learns the detector, not the arrhythmia.

Use underlying_rhythm (the page’s verbatim text) or rhythm_class (sinus / afib / paced) instead. The ari_afib_* columns are exposed because they describe the annotation file — and carry the ari_ prefix so it is never ambiguous which they are.

Note also that record 51 carries no + marker at all, so its 0.0% means “never marked”, not “no atrial fibrillation”. ari_has_rhythm_annotation distinguishes the two.

Where the last audited beat actually falls, for the 12 records that have a .atr:

Records 30, 32, 35 and 51 all stop within 1.4 s of 86,400 s — a hard 24-hour cutoff applied to recordings that run to 25.1 h. Record 51 is additionally unannotated for its first 1,078.7 s.

Nothing errors if you window past the cutoff: you get signal with no reference behind it, and an evaluation that scores a detector against silence. Check atr_unannotated_tail_secs and atr_unannotated_head_secs.

The .ari files are the mirror image — they start 29.7–65.2 s in, because of the 50 LEARN annotations, and reach within 0.2–452.7 s of the end.

The 12 .atr files carry 83 ~ quality transitions and 16,403 | isolated-artifact markers. WFDB defines a ~ subtype as a bitmask over the signals — 0 clean, 1 first channel noisy, 2 second, 3 both — and svdb and nsrdb decode exactly that.

This release only ever writes 0 or 51, strictly alternating 51, 0, 51,

1. So 51 plainly means “noisy” and 0 “clean”, but 51 is not a valid two-channel mask (both channels noisy would be 3), and ECGBench does not reinterpret it as one: it reports atr_noisy_secs and atr_clean_secs with no per-channel split, plus atr_quality_subtypes so a re-release adopting the documented encoding is visible rather than silently misread.

The annotated noise is slight in any case: 579.5 s — 0.161 h — across the whole release, at most 0.95% of a record (41), and 3 of the 12 annotated records carry no ~ at all (36, 46, 49).

Record 35 is the one head-unasserted case. Its single ~ sits 12,600.3 s in and has subtype 0 — a transition into clean — so the span before it was never asserted to be anything. It is counted as clean, which is what WFDB does, and reported as atr_quality_head_unasserted_secs so the assumption is visible. svdb makes the same choice for the same reason.

Of the six checks, exactly one fires. No lead is all-NaN or all-zero, no channel has near-zero variance across a whole record, wfdb reads all 23 headers, and no sample leaves the amplitude bound.

truncated_signal is disabled, by leaving expected_samples empty — and this is the least marginal such case in the catalogue. All 23 lengths differ, from 3,540,000 to 22,627,500 samples, so the longest record is 6.4x the shortest (chfdb varies by 232 s, nsrdb by three hours). Every one is a complete recording, so any single threshold would drop most of the release. Omitting the rate disables the check rather than making it fire.

amplitude_range_mv is [-10.235, 10.235], and it is the looser of two rails because the gain is per record. adc_zero is 0 and no channel declares a baseline, so a sample is confined to [−2047, 2047] adu once −2048 is excluded as the invalid marker:

A single range has to accommodate the loosest record or it fires on a sound one. It is not idle: records 39 and 47 reach both ends, and the observed extremes over all 3.2 billion samples are exactly −10.235 and +10.235 mV. What it guards is a mis-scaled copy — microvolts, or a re-release with the gains dropped — which would exceed it by orders of magnitude at once.

No float32 slack is needed here, unlike chfdb. ECGBench loads signals as float32, and float32(10.235) is 10.234999656677246 while float32(-10.235) is −10.234999656677246 — both round toward zero, so neither can trip a bound set at the exact rail. chfdb’s attained rail of 10.585 rounds the other way and needs the extra thousandth; check the direction before copying either pattern.

One record deserves a note of its own: record 42’s channel 0 sits at the positive rail for 2,337,660 samples, 10.3% of the recording. It is inside the bound by construction, and missing_leads cannot see it because the channel is railed rather than zero, so nothing flags it. That record also has no VF onset.

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

• original splits 19 / 2 / 2. ECGBench’s convention is folds 1–8 → train, 9 → val, 10 → test, and 23 records over 10 folds gives three folds of three and seven of two. So val holds records 36 and 39 and test holds 35 and 45. For anything needing a real evaluation set, use cross-validation: split=None with fold_numbers=[...] selects by fold from folds.csv and ignores the default layout.
• clean splits 3 / 0 / 0, which is the section above.
• Half the val/test records are atrial fibrillation subjects. 36 lands in fold 9 and 35 in fold 10, so two of the four AF subjects in the release are the two non-train records with that rhythm. That is chance, not a choice, and another reason to cross-validate.
• Record 39 is in val, and it is both the shortest-but-one record (5.78 h) and by far the most affected by invalid samples (85,560, 0.7–0.9% per channel). A two-record validation set carrying that is not a robust estimate of anything.

Why the underlying rhythm, and not ectopy burden? Every subject sustained a ventricular tachyarrhythmia, so cohort_label is one value and carries no information a fold could be balanced on. What differs is the rhythm underneath the terminal event, and PhysioNet describes the cohort in exactly those terms:

StratifiedKFold raises only when every class is smaller than n_folds, so the 18-record sinus class carries the split on its own and the singleton paced class is tolerated — sklearn warns that the least populated class has one member, and that warning is expected and correct.

Ectopy burden — the axis svdb and chfdb use — is ruled out by provenance, not arithmetic. 11 of the 23 records have no audited annotation at all, so a burden band would be measuring a 1980s detector in half the release and a cardiologist in the other half, in one column. That is a worse failure than an unbalanced fold.

Folds are ungrouped, and for an unusual reason. Unlike nsrdb, svdb, afdb and chfdb, which ship no subject identifier at all, this release does identify its subjects — it identifies them with the record name. The landing page keys the clinical table “Subject Number” with values 30–52, one record per subject, so patient_id_column is null because a patient column would be a verbatim copy of the index and grouping on it would be a no-op.

PhysioNet states the overlap itself: “Half-hour excerpts of these recordings have been (and remain) available as the MIT-BIH Malignant Ventricular Ectopy Database” (vfdb, also called the Malignant Ventricular Arrhythmia Database on the same page). Anyone combining the two is training and evaluating on the same recordings.

vfdb is not in this catalogue, so there is no related: edge to declare — a related: slug has to resolve to an existing entry. If it is ever added, the edge belongs on one side only, with shares_records: true and a note saying that.

Against the other Boston Holter databases, no shared recording was found. SDDB, mitdb, afdb, nsrdb, svdb, stdb, ltafdb, edb and chfdb have overlapping provenance and none ships a subject identifier that would join, so the question was settled from the annotation files. RR intervals in seconds are commensurable across sampling rates, 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 chfdb, ltafdb and nsrdb.

Against controls that make a null result mean something — a positive control re-finding each SDDB record in its own pool at 100.0000%, and a negative control of each SDDB record against the pool of the other 22 known-distinct subjects peaking at 0.0697% — no database shares a recording:

Every figure sits at or below the 0.0697% chance-collision floor except BIDMC CHF’s 0.1915%. Lengthening the signature settles it: a genuinely shared recording stays near 100% as the signature grows — the positive control does exactly that — while this decays by a factor of 40, which is what chance collisions do. The chfdb page reports the same comparison measured from the other direction (0.1079% → 0.0026%) and reaches the same conclusion.

Two limitations 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 — and this check used SDDB’s unaudited .ari annotations, because they are the only layer covering all 23 records, which makes that marginally more plausible here than elsewhere. And subject-level overlap cannot be checked at all, because none of these releases ships a subject identifier that crosses databases.