European ST-T Database (EDB)

The reference benchmark for ST-segment and T-wave change analysis, and the only dataset in this catalogue whose ground truth is episodes rather than labels. 90 two-hour ambulatory recordings from 79 subjects with diagnosed or suspected myocardial ischaemia, assembled between 1985 and 1990 under a European Society of Cardiology concerted action across centres in seven countries, coordinated from Pisa. Two cardiologists annotated every record beat by beat and marked the onset, extremum and end of each interval of significant ST or T change, working on the two signals independently, with a third resolving their disagreements. 180 hours, 802,909 annotations, 368 ST episodes and 401 T episodes.

Lead layout varies more here than in any other dataset in this catalogue, and no lead is present in every record. All 90 records store two leads — and they use fifteen different orderings of eleven different lead pairs. V5 reaches 51 records, MLIII 47, V4 34, and D3 appears exactly once. MLIII/V4 and V4/MLIII are both present, 15 records each, so signal[0] is a limb lead in one half of those 30 records and a chest lead in the other with nothing in the metadata to say which. Lead placement was never standardised across the contributing centres. This is the single most important thing on this page: config.lead_names is ["V5", "MLI"] and describes only 19 of the 90 records, so it is a name for the modal layout and not something to index against.

The deviations are measured against each subject’s own reference waveform, not against an absolute isoelectric line. ST deviation is taken 80 ms after the J point (60 ms above 120 bpm) and compared with a reference complex from that record’s own first 30 seconds; an episode requires 0.1 mV of ST deviation (0.2 mV for T) sustained for at least 30 s. So peak_st_deviation_uv = 600 means 600 µV above where the ST segment sat in the first 30 seconds of that record, and subjects with prior infarction carry fixed elevation or depression underneath that these annotations do not describe. A fixed ST-level threshold cannot reproduce these labels. The reference complexes themselves were printed on transparent plastic rulers which, the release notes, no longer exist.

The signals carry a large uncorrected DC offset. Gain was calibrated against the original analog calibration signals to a uniform 200 ADC units per millivolt; offset was not. 116 of the 180 signals sit more than 1 mV off zero and 58 more than 3 mV, up to +9.05 mV, and 21 records never cross 0 mV at all — e0114 lives entirely between +5.635 and +9.785 mV. Peak-to-peak is a normal 4.02 mV median. Baseline removal is a prerequisite here, not a refinement, and it is consistent with the annotation semantics above: this database never claims an absolute ST level.

No subject identifier ships. edb.txt states the 90 records come from 79 subjects and nothing in the files says which. ECGBench reconstructs it; see “Subject identity is reconstructed” below before relying on patient_id.

Every figure on this page was recomputed from the 90 headers and .atr files, after verifying the shipped data against the release’s own SHA256SUMS.txt — all 275 listed files match. There is no changelog in the release directory.

The T-episode count matches exactly, which is what validates the counting convention. An “episode” is counted once, at its onset annotation. Getting 401 requires all three of: truncating aux_note at its NUL terminator, excluding the 166 ++/-- extreme-T threshold crossings, and excluding the 21 lower-case axis-shift spans. Get any one wrong and the total is 394, 567 or 411 instead. That the same convention gives 368 ST episodes against the paper’s 367 is therefore a real one-episode difference between the shipped annotations and the 1992 paper, not a counting error — and with no changelog in the release there is nothing to attribute it to. The same goes for the 43 extra annotations: they are not at the record boundaries (no annotation in the release falls at sample 0 or at or past sample 1,800,000) and they are not a single annotation type. Treat the recomputed figures as authoritative for v1.0.0 and the published ones as describing whatever the authors held in 1992.

The women’s age range differs and the reason is visible. Seven of the eight female subjects have an age: 53, 55, 56, 62, 64, 70 and 71. e0418’s is -. The paper says “8 women aged 55 to 71”, which excludes the 53-year-old (e0818). The men’s figures — 70 subjects, 30 to 84 — reproduce exactly.

One derived figure is a convention, and it is the largest one. e0409 shows ischaemic_fraction 0.930, the highest in the release, and that number exists because its two ST depressions have onsets (8.44 min and 17.15 min) and extrema (103.51 min and 99.06 min) but no end annotations. The episodes are genuinely very long — the annotators marked their peaks more than 80 minutes after the onsets — but where they end was never recorded, and ECGBench closes an unterminated episode at the end of the record. 10 of the 12 unterminated episodes open in the final four minutes and the convention barely matters for them; for e0409 it sets the figure. n_unterminated_episodes flags it.

st_episode_secs can exceed the recording length, and that is correct. The two signals are annotated independently, so concurrent ST depression in both channels counts twice: e0607 reaches 131.5 min of ST episode in a 120-minute record. st_secs_any_signal is the bounded union — 29.2 h across the release against 40.2 h summed — and ischaemic_fraction is that over the 7,200 s record. Use the union for any fraction.

config.record_lead_layouts declares all fifteen, which makes ECGDataset(leads=[...]) open each record’s own header and resolve the requested names against it. mitdb is the only other config here that needs this; edb needs it far more, because mitdb’s declared layout covers 40 of 48 records and this one’s covers 19 of 90.

Because no lead is in every record, every name-based selection here raises for some records. leads=["V5"] loads 51 and raises ValueError: Record 'e0604' stores ['V2', 'MLIII'] … Lead 'V5' is not in 'edb' for the other 39. That is the intended behaviour — the alternative is silently returning a different physical lead — but it means leads= alone does not make this dataset batchable. You need a lead and a record filter, and the widest single choice available is V5 at 57% of the release.

Selecting positionally is wrong for at least 71 of the 90 records, and the failure is silent. It is worth restating that MLIII/V4 and V4/MLIII are both present in equal numbers: those 30 records are indistinguishable by shape, by lead count, and by anything in the exported fold CSVs.

The per-record layout is in the labels as lead_names, pipe-separated, so it can be read without opening a header.

An episode is three annotations and the WFDB symbol does not distinguish them — s for ST and T for T change, onset or extremum or end alike. The aux_note text is the only source: (ST0+ opens an episode of ST elevation in signal 0, AST0+600 marks its extremum at +600 µV, and ST0+) closes it. Three things about that text cost real accuracy:

1. aux_note carries bytes past its NUL terminator. Seven T-episode onsets read '(T0+\x00\x13', '(T0-\x00N', '(T1-\x00\x1a' and so on. Neither .strip() nor .rstrip('\x00') removes the trailing garbage, so those seven fall into categories of their own and the T total comes out at 394 instead of

1. Truncate at the first NUL, which is what the format means.

2. ++ and -- are not episodes. Inside a T episode whose deviation exceeds 400 µV, extra annotations mark each crossing of that threshold. There are 166 of them; counting them as episodes inflates the T total by 41%. They are exposed as n_extreme_t_markers.

3. Lower-case episode text is recognised artefact, and case-folding merges it with the findings. In six records — e0161, e0509, e0601, e0611, e0613, e0615 — a positional change shifts the axis and produces something that looks exactly like an ST or T change. The annotators marked those 21 spans on " comment annotations spelled (st0+, at1-800, st0-) in lower case precisely so they can be told apart from the upper-case findings. They are n_axis_shift_episodes and are excluded from every episode count.

A fourth, in the documentation rather than the data: the ~ signal-quality subtype table in the shipped annotations.shtml disagrees with the files. It tabulates nine values and gives un as 0x12, cu as 0x20 and nu as 0x21; the release contains 0x13, 0x22 and 0x23 and none of those three. The table is internally inconsistent — its own uc is 0x11, which already sets signal 0’s noisy bit — so ECGBench reads the subtype as the bitmask it is (bit 0/1 noisy for signal 0/1, bit 4/5 unreadable, unreadable also setting noisy). That accounts for all 8,918 quality annotations with nothing left over.

Because it is the quantity this database exists to measure, and nothing else here works. The header clinical text is subject background — every subject was selected for suspected ischaemia, so it is not a case/control axis. dominant_rhythm is sinus in all 90 records. st_t_class has two classes of two records. What genuinely differs between records is how much ST change each one holds, and it differs enormously: e0604 holds 20 episodes and four records hold none.

The band edges are fixed counts, not quantiles — 1, 3 and 6 episodes, giving 4 / 32 / 29 / 25 records. Quantile edges would balance slightly better and would move every record’s label the next time an annotation is revised; fixed edges reproduce against a re-release, which matters more for a partition that gets published.

Every one of the ten folds carries all three populated bands — each holds 3–4 1-2, 2–3 3-5 and 2–3 6+ records. The none band holds 4 records, fewer than the 10 folds, so it cannot appear in every fold; in this partition all four land in train, which means val and test contain no ST-free record. It is kept as its own band regardless, because a record with no ST change at all is the negative control an ST detector is scored against and folding those four into 1-2 would make them invisible. With 90 records over 10 folds, train holds 74, val 8 and test 8.

st_burden_band is a fold-construction label and a coarsening at that. Train on the episode counts, ischaemic_fraction, peak_st_deviation_uv, or — for episode detection, which is what this database was built to evaluate — on the .atr files directly. Never on stratify_class.

The release ships no subject identifier. edb.txt says the 90 records come from 79 subjects; nothing in the files says which are which. That matters, because ungrouped folds would put the same person in train and test — e0118, e0119, e0121 and e0122 are one 51-year-old man recorded four times.

ECGBench therefore reconstructs it from the header: records agreeing on age, sex, recorder model, medications and the set of clinical findings are taken to be one subject. That gives 80 groups. The remaining merge is e0206 and e0210, which agree on everything except age — 55 against 53, two recordings of the same 3-vessel-disease man on the same Oxford Medilog MR-20 in the same V5/MLI placement, two years apart. Merging them gives 79.

That the reconstruction lands on the published subject count is the check that it is not merely plausible, and it reproduces the published demographics too: 70 men aged 30–84, against edb.txt’s “70 men aged 30 to 84”, 8 women, and exactly one subject whose information is missing. The result is 72 subjects with one record, five with two, and two with four.

Two things to know about the method. The findings must be compared as a set: e0126 lists exactly the same five clinical lines as e0123–e0125 but with “Aortic valvular regurgitation” and “1-vessel disease (LAD)” swapped, so an order-sensitive comparison drops it from its own subject’s group. And the reconstruction is conservative in one direction only — two genuinely different subjects sharing age, sex, recorder, medication and findings would be merged, which costs a little fold flexibility and no correctness, while the error that matters, splitting one subject across folds, is bounded by the count agreement.

An attempt to confirm the grouping from the signals was inconclusive, and is not the basis for it. Normalised median QRST complexes compared across the same lead pair do not separate subjects in this cohort: a different-sex control pair (e0203 F / e0206 M) scored 0.978 against a within-group minimum of 0.758, and the control median over 468 same-layout pairs was 0.823. Ischaemic patients recorded on the same two leads simply look alike at that resolution. The claim here rests on the header agreement and the published-count match, not on the waveforms.

Signal quality is annotated per channel by ~ transitions, each opening an interval that runs to the next. Across the 180 h per channel, signal 0 is 93.21% clean (11.52 h noisy, 0.70 h unreadable) and signal 1 is 94.35% clean (9.11 h noisy, 1.07 h unreadable). The averages hide a wide spread: e0148 is 52.2% noisy or unreadable across its two channels, e0808 39.4% and e0817 33.1%, while e0417 carries no ~ annotation at all and is clean end to end. usable_fraction — the share of the record in which at least one channel is readable — never drops below 0.982.

No record has a ~ at sample 0. The first falls at a median of 9.3 minutes, so the leading span was never asserted to be anything; ECGBench counts it clean, which is what WFDB does, and reports it as quality_head_unasserted_secs so the assumption is visible rather than buried.

Rhythm, by contrast, is annotated from the start of every record: all 90 open with a + between sample 2 and 264 — (N in 88 of them and (SBR in e0303 and e0611 — so rhythm_secs_* covers the whole recording bar under a second. Sinus dominates every record, which is why dominant_rhythm is not the label column, but the non-sinus spans are real: 20.7 min of atrial fibrillation in one record, 17.2 min of sinus bradycardia in three, 6.5 min of ventricular tachycardia in nine, plus bigeminy, trigeminy, supraventricular tachyarrhythmia, sino-atrial block and one third-degree block.

Also marked: 54 patient-event button presses in 12 records (note_BUTTON — the subject pressing a button, typically at symptom onset, which is a label of a kind), one tape slippage (note_TS), and 78 isolated QRS-like artefacts in 15 records.

Ten Holter recorder models were used, unevenly: ICR 7200 for 37 records, Oxford Medilog MR-20 and Del Mar Avionics 445B for 14 each, Oxford Medilog 4-24 for 12, and six models for 1–3 records each. Recorder-specific artefact is a real confounder in a collection of this era, so recorder_type is exposed to control for it.

All 90 records pass every check, so original and clean hold the same 90 rows. There are no NaN samples, no flat or all-zero leads and no unreadable header. Records are exactly 1,800,000 samples each, uniformly, so unlike afdb, nsrdb and ptbdb the truncated_signal check is switched on rather than disabled — and any window= inside 7,200 s fits all 90 records.

amplitude_range_mv is [-10.24, 10.235], and here that is the only defensible value. Format 212 is 12-bit two’s complement, so with adc_zero 0 and a gain of 200 adu/mV full scale is exactly [−2048, 2047] adu = −10.240 to +10.235 mV. A tighter physiologic window would be meaningless for this release, because the DC offset is uncorrected and large: 21 records never cross 0 mV and one sits entirely above +5.6 mV, so absolute amplitude here measures the offset rather than the ECG.

Measured across all 90 records, samples run −8.195 to +10.045 mV, so no record saturates and the check cannot fire on EDB 1.0.0 — which is the intended state. What it guards is a mis-scaled copy: a read in microvolts, or a gain applied twice, would be 200× out on the first record.

The QT Database overlap is real, verified, and declared — 33 of its 105 records are 15-minute excerpts of 33 of these recordings, confirmed sample-for-sample against the provenance line in QTDB’s own headers. See the related-datasets note above; the practical consequence is that this database and QTDB’s sele0* records cannot be used as train and test for one another.

The Long-Term ST Database is not declared, because it cannot be checked. LTSTDB comes from the same ST-analysis research community and shares contributors, and its 86 records are 21–24 h rather than 2 h, so they are certainly different recordings. Whether any of its 80 subjects also appear here is unanswerable from the files: its record identifiers (s20011, s30661) share no scheme with these, and neither release publishes a subject identifier that could join. Nothing is asserted either way.

The MIT-BIH ST Change Database is a separate collection — 28 records from the Beth Israel Hospital, and QTDB’s sel3* records come from it rather than from here.