BIDMC Congestive Heart Failure Database

The severe-heart-failure cohort. 15 twenty-hour two-lead Holter recordings of subjects with NYHA class III–IV congestive heart failure — 11 men aged 22 to 71 and 4 women aged 54 to 63 — made at Boston’s Beth Israel Hospital while they were on conventional therapy, before entering the oral milrinone trial Baim et al. reported in 1986. 298.9 hours of signal at 250 Hz and 1,622,282 detected beats. It is the natural counterpart to the MIT-BIH Normal Sinus Rhythm Database’s control cohort: one class, and every record is in it.

The beat annotations are unaudited machine output, and this is the first thing to know. PhysioNet states it plainly — “Annotation files (with the suffix .ecg) were prepared using an automated detector and have not been corrected manually” — and the shipped ANNOTATORS file agrees. The .ecg extension is itself the tell: every audited MIT-BIH database in this catalogue uses .atr. So unlike mitdb, nsrdb, svdb and edb, nothing here is a cardiologist reference standard. Every beat, ectopy and HRV figure on this page describes what one 1980s detector reported. It has its own section below.

Ventricular ectopy is heavy and enormously uneven — the real signal here. 38,524 of the 1,622,282 beats are ventricular, 2.37%, against NSRDB’s 0.0015%. But the spread across records is three orders of magnitude: chf02 is 20.52% ventricular and chf12 is 0.017%. A per-record metric is not comparable across this database without controlling for it.

r is a ventricular beat, and it outnumbers V in 9 of the 15 records. 10,353 annotations are r, R-on-T premature ventricular contractions, which AAMI EC57 classes as ventricular. A pipeline counting only V therefore undercounts ventricular ectopy across most of this release — use the aami_* columns or veb_fraction. There are no fusion beats at all.

Records are ~20 h: 17,789,952 to 17,998,848 samples, ~142 MB of float32 each. Batching needs a window=(start, length), which is read at load time rather than cropped afterwards. Length varies by only 232 s, so one window fits every record — but it must fit the shortest, chf06.

The two channels are not named leads. The headers call them ECG1 and ECG2 and the release states no electrode placement anywhere. Its siblings from the same hospital — the MIT-BIH Arrhythmia Database — do 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 15 headers and .ecg files, after verifying the shipped data against the release’s own SHA256SUMS.txt — all 64 files match. Unlike NSRDB’s copy, that total includes the superseded backups PhysioNet keeps beside the current revisions: 15 .hea- files predating the 2012 revision that added the ECG1/ECG2 signal descriptions, and 2 .ecg- files for chf02 and chf04, whose annotations were regenerated in 2003. 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 “11 men, aged 22 to 71, and 4 women, aged 54 to 63”, and the header comments give precisely that — men at 22, 48, 51, 53, 54, 61, 61, 63, 63, 71 and one unstated, women at 54, 59, 61, 63. There is no changelog and no discrepancy to explain. One age is genuinely missing: chf06’s header records Age: ?, which ECGBench keeps as NaN rather than 0.

Vent. is the AAMI ventricular class, not the V symbol. It sums V (26,166 release-wide), r (10,353) and E (5). The r beats are R-on-T PVCs and are ventricular under AAMI EC57; because they outnumber plain V in 9 of the 15 records, the V column alone would understate ventricular ectopy for most of the database. The full symbol breakdown is 1,578,151 N, 26,166 V, 10,353 r, 5,314 S, 293 Q and 5 E, plus 258 + rhythm markers that are not beats.

Vent. % is a fraction of that record’s beats; AF % is a fraction of its duration. They are not the same denominator, and the AF column carries a further caveat given its own section below.

The HRV figures are descriptive, not a result. mean_hr_bpm, sdnn_ms and rmssd_ms come from RR intervals in [0.3 s, 2.0 s]; 612 intervals are rejected across the release. Two caveats compound here in a way they do not for NSRDB: the beats are uncorrected machine output, and these are whole-record summaries over ~20 h of activity and sleep in subjects with severe heart failure and heavy ectopy. A real HRV analysis would segment, exclude ectopic couplings, and not use these labels at all.

cohort_label and nyha_class are constants, and neither is the stratification label. They record what the release asserts of the cohort so that a user combining this database with a normal or mixed-arrhythmia one has a record-level class to join on. Folds are stratified on sex (11 M / 4 F) — see the fold section. Train on veb_fraction, af_fraction, mean_hr_bpm, sdnn_ms or the aami_* counts; never on stratify_class.

This is the most important difference between this database and the other MIT-BIH-family releases in the catalogue, and it is easy to miss because the annotations look identical in structure to mitdb’s.

PhysioNet’s wording is unambiguous, and the shipped ANNOTATORS file describes the contents as “unaudited beat annotations from an automated detector”. Two consequences worth stating rather than leaving to be inferred:

• Do not train or evaluate a beat classifier on these labels. A model scored against them is being scored against a 1980s detector, and the conditions under which such detectors mislabel most — heavy ventricular ectopy and atrial fibrillation — are exactly what this cohort has.
• The ectopy rates are not clinical facts. They are useful for describing a recording, for stratifying, and as weak supervision a human would have to confirm. They should not be quoted as this cohort’s arrhythmia burden.

What the release is good for is unaffected: 298.9 hours of severe-CHF waveform, a severity-matched positive cohort, and a self-supervised or unsupervised corpus.

258 of the annotations are + rhythm markers carrying (AF and (N. They are very unevenly distributed:

For those 11 records, af_secs == 0 means the rhythm was never assessed, not that there was no atrial fibrillation. Filtering on af_secs == 0 would pick them up as though they were confirmed negatives, which is why the labels expose has_rhythm_annotation — check it before treating a zero as evidence.

chf06 has one further wrinkle. Its first + sits 1,757.0 s into the record and is (N — a return to normal rhythm, which implies the preceding 29 minutes were AF without ever saying so. Counting that span as AF would invent annotation; counting it as normal would assert what the annotator contradicted. It is reported separately as rhythm_head_unasserted_secs, the same choice svdb makes for its signal-quality head.

Note that chf06 is also the single record in the test fold, so the default test set is one predominantly-AF recording. For anything needing a real evaluation set, use cross-validation — see the fold section.

nsrdb, svdb and mitdb all ship ~ signal-quality transitions with a per-channel bitmask, and | isolated-artifact markers. This release contains neither — zero ~ and zero | across all 15 files.

ECGBench exposes n_quality_changes and n_isolated_artifacts as 0 so that a re-release adding a quality layer is visible rather than silent, but deliberately provides no clean_secs/noisy_secs columns. Deriving them from an absent annotation layer would assert that 298.9 hours of tape are clean when nobody assessed them. Judge quality from the waveform.

All 15 records pass every check, so original and clean hold the same 15 rows. There are no NaN samples anywhere in the 540 million read, no flat or all-zero leads and no unreadable header. Two checks need explaining:

• truncated_signal is disabled, by leaving expected_samples empty. Records run 17,789,952 (chf06) to 17,998,848 samples and every one is a complete 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, ltafdb and nsrdb use.
• amplitude_range_mv is [-10.24, 10.586], which is not symmetric. 29 of the 30 channels have adc_zero 0, which at wfdb’s fallback gain of 200 puts them in [−2048, 2047] adu = [−10.24, 10.235] mV. chf15’s ECG2 is the one exception: a baseline of −70 adu shifts its range to [−9.89, 10.585] mV, and that channel actually reaches +2047 adu — for 12 samples, 0.0001% of the record. So the bound has to be the union of the two rails, or the release’s own top record fails a check computed to admit it.

And it cannot be the exact rail, which is worth knowing before setting one of these anywhere else. ECGBench loads signals as float32, and float32 cannot represent 10.585 — the nearest value is 10.585000038146973, which is greater than a float64 bound of 10.585. A bound set to the attained rail therefore excludes the very record it was derived from; that is what the extra thousandth of a millivolt is for. The negative bound needs no such slack, because float32(−10.24) rounds toward zero.

Observed extremes across the whole release: −10.17 mV (chf09 ECG1) to +10.585 mV (chf15 ECG2). Unlike NSRDB, which never exceeds half its declared span, this release uses nearly all of it.

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, NSRDB and LTAFDB — with the difference that here the default is corroborated. PhysioNet’s description states “250 samples per second with 12-bit resolution over a range of ±10 millivolts”, and 12 bits at 200 adu/mV is exactly [−2048, 2047] adu = [−10.24, 10.235] mV. The declared gain and the described range agree, which is not true of AFDB (whose stated range implies 204.8 adu/mV and has to be reconciled) and is simply unstated for NSRDB.

One further property of the recording chain matters for modelling: the signals are band-limited to roughly 0.1–40 Hz, so despite the 250 Hz sampling rate there is no high-frequency content, and they are not comparable in detail with modern digital ECG.

Three 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 15 records over 10 folds gives five folds of two and five of one. So train holds 13 records, val holds 1 (chf04) and test holds 1 (chf06). 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.
• The single test record is the AF-dominant one. chf06 lands in fold 10, so the default test set is 80% atrial fibrillation and is also the record whose age is unstated. That is chance, not a choice, and it is another reason to cross-validate rather than trust one held-out record.
• Most folds contain no woman. The four women land in four different folds (chf02, chf05, chf09, chf11 in folds 4, 3, 2 and 5), so folds 9 and 10 — val and test — are both male. Four 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 the ventricular ectopy burden? Because StratifiedKFold requires at least one class holding n_folds members, and with 15 records over 10 folds that means a 10/5 or more lopsided cut. Ectopy burden is the clinically interesting axis and svdb does stratify on exactly that, in bands — here every meaningful banding fails:

Sex clears the requirement with margin, is documented by PhysioNet, and does not rest on the unaudited detector output every ectopy figure here comes from. sklearn warns that the smallest class has 4 members; that warning is expected and correct.

Folds are ungrouped. The header comment holds age, sex and NYHA class and nothing else — no tape number, no recorder, no subject code, and not even the trial arm the cohort is defined by — and PhysioNet describes 15 recordings from 15 subjects, so one record per subject is the most that can be asserted.

CHFDB, the MIT-BIH Arrhythmia, Atrial Fibrillation, Normal Sinus Rhythm, Supraventricular Arrhythmia and ST Change databases and the Sudden Cardiac Death Holter Database all come out of Boston’s Beth Israel Hospital, 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, 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 and NSRDB.

Against controls that make a null result mean something — a positive control re-finding each CHFDB record in its own pool at 100%, and a negative control of each record against the pool of the other 14 known-distinct subjects peaking at 0.0869% — no database shares a recording:

Every figure sits at or below the 0.0869% chance-collision floor the negative control establishes, except Sudden Cardiac Death’s 0.1079%, which marginally exceeds it against a 1.77-million-signature pool. Repeating the check at a 30-interval signature settles it: a genuinely shared recording stays near 100% as the signature lengthens, and these decay by one to two orders of magnitude, which is what chance collisions 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 — and this release’s annotator is an uncorrected detector, which makes that marginally more plausible here than elsewhere. Subject-level overlap cannot be checked at all, because none of these releases ships a subject identifier.

Note also that PhysioNet’s Congestive Heart Failure RR Interval Database (chf2db, 29 subjects) is a different database, not a derived layer over this one, and is not in this catalogue.