Brno University of Technology ECG Quality Database (BUT QDB)

The only dataset in this catalogue whose ground truth is a label per sample. Every other release answers “what is wrong with this patient”; BUT QDB answers “can this stretch of signal be analysed at all”, and it answers it at 1 kHz over 99.4 hours of free-living wearable ECG.

18 single-lead recordings, each longer than 24 hours, from 15 people wearing a Bittium Faros 180 on the chest between August 2018 and October 2019 while carrying out ordinary daily activities. 478.7 hours of ECG at 1,000 Hz, with 3-axis accelerometry at 100 Hz alongside every recording. There is no cardiac diagnosis anywhere in the release — this is a quality-assessment benchmark.

Three ECG experts graded the signal independently, and the release ships all three opinions plus their consensus. Three classes: class 1, every waveform (P, QRS, T) measurable; class 2, QRS reliably detectable but nothing finer; class 3, QRS not reliably detectable and the signal unusable for anything.

Only 20.8% of the recorded time is graded, and 88.6% of the graded time is three of the 18 records. That is the first thing to plan around, and the second is where the graded segments are — see the next section.

clean and original are the same 18 records, deliberately. Excluding recordings for being noisy would destroy exactly what this database exists to measure: the noisiest recording, 105001, is the single most valuable one.

Three recordings — 100001, 105001, 111001 — are graded end to end. The other 15 get two 20-minute segments each, plus five extra segments the authors picked for being noisy (four of 20 minutes, one of 2 minutes).

Those two standard segments sit at exactly the same offsets in all 15 records, which the release does not mention and which was recomputed from the annotation files:

So the obvious first thing to try returns unlabelled signal:

from ecgbench.labels.butqdb import quality_vector
quality_vector(path, "100002", start=0,         length=10_000)  # all 0 = ungraded
quality_vector(path, "100002", start=28_800_000, length=10_000)  # classes 1, 2

annotated_blocks(path, record_id) returns the bounds per record, and the five extra segments are elsewhere: 113001 at 36,119,999; 114001 at 11,214,750 (the 2-minute one, which is 120,001 samples, not 120,000) and 11,674,750; 124001 at 33,699,999 and 65,099,999.

Over the 99.4 graded hours, by the expert consensus:

Two reasons not to read 15.2% as a property of 24-hour wearable ECG. Five of the graded segments were “subjectively selected” to raise the proportion of poor signal. And 87.0% of all class-3 time is one record: 105001 alone contributes 13.15 of the 15.12 unusable hours.

Only six records carry an appreciable class-3 burden at all — 122001 50.0%, 105001 34.0%, 114001 22.6%, 124001 14.3%, 111001 4.4%, 113001 2.8% — and the other twelve sit between 0.00% and 0.38%.

All three agree on 248,410,246 of the 357,799,001 graded samples — 69.4%. Pairwise agreement averaged over the 18 records is 0.82 (experts 1–2), 0.81 (1–3) and 0.89 (2–3). Expert 1 is systematically stricter, calling class 2 where the other two call class 1: on 121001 expert 1 gives 51.8% class 1 against expert 2’s 90.4%.

Anything reported as accuracy against “the” label is reporting agreement with a consensus that three humans reached this loosely. Per-record and per-expert figures are in the labels (expert_1_class1_fraction … expert_unanimous_fraction, mean_expert_agreement), so the ceiling is visible rather than implicit. Records range from 53.1% unanimous (124001) to 97.6% (125001).

The fourth column triple is the consensus, and it behaves as a majority vote. The release states the annotation file holds “3 columns × 3 annotators + consensus” but not how the consensus was formed. Measured over every graded sample: a majority of the three exists at 99.863% of them, the consensus equals that majority at all but 3,103 of those samples (99.99913%), and it differs from every expert at only 378 samples in the whole release. The residuals sit at interval boundaries — what one expects from a consensus drawn segment by segment. consensus_matches_majority records it per record, so a re-release formed by a different rule would show up rather than being assumed away.

Every figure on this page was recomputed from the shipped files, after verifying the local copy against the release’s own SHA256SUMS.txt — all 95 listed files match.

The release’s own summary mixes two denominators in one sentence. “Broadly balanced in terms of gender (10 female records and 8 male records) and age (21 to 83 years, mean 41 years, median 37 years)” — the gender figures are per record and the age figures are per subject. Per record the cohort is much younger: mean 37.7, median 28, because the two repeat subjects are 28 and 21 and contribute five of the 18 records. Neither figure is wrong; quoting them together is.

The age distribution is bimodal, not broad. Eight subjects are 21–28 and seven are 37–83, with nobody between 28 and 37, so any age effect estimated here is estimated across a gap. Three subjects smoke; height 153–184 cm, weight 43–85 kg, BMI 17.4–29.4.

Gain and baseline differ per record — 0.99998 to 1.996 ADC units per µV, baseline −18,289 to +11,462 — so each record has a different physical span: 100001 covers [−10.30, +22.53] mV, 124001 [−32.77, +15.78], 104001 [−32.77, +32.77]. All 18 records attain both of their 16-bit rails, so the configured amplitude_range_mv has to be their union, [-32.769, 32.769], and the check cannot exclude anything.

The bounds are attained, so the float32 rounding trap applies: what the loader actually produces is −32.76765823364258 mV and +32.768115997314453 mV, so a bound of 32.768 would have excluded the very record it was computed from.

Saturation is much milder than in toliet-thigh-based-ecg-toilet-seat: 66,166 samples across the release sit at a rail, the worst record being 114001 at 0.015% of its samples and 100001 at two samples out of 87 million. No check in CHECK_REGISTRY measures it, so clipped_fraction, min_mv and max_mv are in the labels.

WFDB’s invalid-sample marker does not occur. Format 16 reserves −32768, which wfdb.rdrecord turns into NaN — and nan_values fails a record on a single NaN. There is not one anywhere in the release, so nan_values passes on all 18 records rather than passing by luck; n_invalid_samples reports it so a re-release that introduces the marker is caught.

No record has a NaN sample, an unreadable header, a flat lead or a truncated signal, and amplitude_outlier cannot fire (above). So clean is all 18 records and original is the same 18 with is_valid and quality_issues attached.

That is not a validation gap — it is the only defensible setting for this dataset. ECGBench’s checks exist to exclude recordings that cannot be analysed; here, which stretches cannot be analysed is the ground truth three experts were paid to produce. Use the annotations, not is_valid:

usable = ds.labels_df[ds.labels_df["consensus_class3_fraction"] < 0.05]

expected_samples is deliberately empty in the config: every record is a different length, 86,420,000 to 139,147,000 samples, and each is a complete recording that ran until the battery or the subject stopped.

Folds are built with StratifiedGroupKFold, grouped on subject_id and stratified on stratify_class — whether more than 1% of the record’s graded time is class 3.

Why class-3 burden and not sex. What the folds have to balance is unusable signal, because rejecting unusable signal is the task. Only six records carry an appreciable class-3 burden, so at most six of the ten folds can hold one — and this axis achieves exactly six, the arithmetic maximum. Measured over the shipped files at random_state=42:

Sex is what mit-bih-normal-sinus-rhythm-database and bidmc-congestive-heart-failure-database use, and it is not better here: 15 subjects over 10 folds put one or two records in a fold, so the female fraction per fold is 0, 0.5 or 1 whatever it is stratified on. The threshold is not tuned — twelve records sit at 0.00%–0.38% class 3 and six at 2.8%–50.0%, so any cut in between gives the same partition. Every cross of two axes raises, because StratifiedGroupKFold needs at least one class holding n_folds records and no cross of 18 records has one.

No subject spans a fold: 103’s three recordings are all in fold 1 and 100’s two are both in fold 2. The three fully-graded records land in three different folds (2, 5, 10) without annotation depth being in the stratification at all.

Use the folds, not the default split. With folds 1–8 → train, 9 → val, 10 → test:

One record is not an evaluation set. That is the arithmetic of 18 recordings, not a defect in the partition — for a real evaluation, cross-validate: split=None with fold_numbers=[...] selects by fold from folds.csv and ignores the default layout.

Every recording has a companion WFDB record <id>_ACC: three channels (ACCx, ACCy, ACCz) at 100 Hz in milli-g. The release’s own usage notes propose motion as an input to quality assessment, either alongside the ECG or on its own.

ECGBench gives it no records and does not declare 100 Hz as a sampling rate — it is not an ECG, and declaring it would make ECGBench offer to load accelerometry as though it were an ECG lead. acc_path in the labels points at each one, and wfdb.rdrecord reads them directly.

No related: edge is declared. This is an original 2018–2019 Brno recording campaign on a Bittium Faros 180 under free-living conditions, and no other release in the catalogue shares a subject, a recording or an institution with it — Brno University of Technology appears nowhere else here.

The nearest neighbours are by purpose rather than content. mit-bih-normal-sinus-rhythm-database and the other MIT-BIH Holters carry ~ signal-quality annotations, but those mark noise transitions in a two-lead clinical Holter recording, not a graded three-class scale from three independent experts on a consumer wearable. Nothing else in the catalogue offers per-sample quality labels at all.

The overlap that matters is inside this release: two subjects contributed more than one 24-hour recording (100 twice, 103 three times), which is why folds group on subject_id.