Leipzig Heart Center ECG Database

39 recordings from 39 subjects, made during electrophysiological studies at the Leipzig Heart Center — and unusually for this catalogue, every record carries the 12-lead surface ECG and the intracardiac electrograms from the catheters, sampled together at 977 Hz.

The cohort is the point. Two groups, deliberately chosen for conditions the big adult datasets barely contain:

• 29 children (x001–x0029, ages 5–18) referred for ablation of supraventricular tachycardia — 13 with AV nodal reentrant tachycardia, 16 with AV reentrant tachycardia including 5 with Wolff-Parkinson-White and one with permanent junctional reciprocating tachycardia. The children’s records also give the accessory-pathway location.
• 10 adults (x100–x109, ages 21–64) with repaired Tetralogy of Fallot, 8 of them with ventricular tachycardia.

Two cardiologists annotated every beat — one annotating, the second checking and correcting — using LightWAVE. The result is 118,214 annotations over 18.5 hours, with beat class, tachycardia mechanism, rhythm segment and signal quality all recorded. Signals are the raw CardioLab export, filtered 0.05–100 Hz at acquisition and not touched since.

Licence is ODC-By 1.0 and access is open, so ECGBench publishes the fold CSVs to the Hub and they download automatically.

The README describes two layouts: children with the 12-lead ECG plus 5 coronary-sinus, 1 right-ventricular-apex and 1 ablation channel (19), and adults with the ECG plus RVA and ABL, with the coronary sinus catheter “in some studies”. The files hold six layouts and four different channel counts. Verified against all 39 headers:

Three consequences:

1. Only channels 0–11 are stable. The 12-lead ECG is the same channel in the same position in all 39 records. Channel index 12 is ABL12 in 28 records, RVA12 in 4 and ART in one.
2. ABL_uni and ART are documented nowhere in the release — not in the README, not in dataset_info.csv. ART is most likely an arterial pressure trace, but the release does not say, so ECGBench does not guess.
3. Four children’s records carry no ablation channel at all, against the README’s “every recording of the children includes … an ablation catheter”.

So config.lead_names declares the 12 surface ECG leads and nothing else, deliberately: there is no dataset-wide intracardiac order for ECGDataset(leads=…) to resolve against. To reach an intracardiac channel, look its name up in that record’s own header:

from ecgbench.labels.leipzig_heart_center_ecg import channel_index

names = labels["channel_names"]            # 'I|II|…|ABL12|RVA12|CS12|…'
channel_index(names, "RVA12")              # 13 in most records, 18 in x100
channel_index(names, "CS12")               # None in the 14-channel records

And pass leads= if you want a homogeneous batch — without it a batch mixes 14-, 18-, 19- and 20-channel tensors.

Records run from 77.7 s (x0027) to 2 h 30 m 32 s (x003), median 699 s. Two things follow.

validation.expected_samples is deliberately empty — a truncation check would fail almost every record — and window=(start, length) is how you batch at all. 60 s (58,620 samples at 977 Hz) is the longest window that fits every record; anything longer raises WindowOutOfRangeError on x0027.

Recording length also correlates with case complexity, so a naive per-beat pooling is dominated by a handful of long children’s studies: x001, x003, x004 and x006 alone contribute 61,806 of the 118,214 annotations.

The shipped diagnosis has seven classes over 39 records, three of them singletons. Ten folds cannot hold that, so the label loader derives a diagnosis_family — the leading token, giving AVRT (16), AVNRT (13) and TOF (10) — and folds are stratified on that. The result is even: every one of the ten folds holds 3–4 records with all three families represented.

stratify_class is for fold construction only. It throws away whether an AVRT had a manifest accessory pathway and whether an adult’s VT was sustained, which is the very thing the adult cohort is interesting for. Train on diagnosis, and better still on the beat-level tachy_* columns: a child referred for AVNRT still contributes AFIB and VT beats, and those are recorded.

Because the families map cleanly onto the cohorts — all TOF are adults — the fold stratification also spreads children and adults evenly without being asked to.

With 39 records, prefer cross-validation to the default 8/1/1 layout, which leaves 3 records in test and 4 in val. Pass split=None with fold_numbers=[…] to select folds directly.

Recomputed from the files. All 126 files match the release’s own SHA256SUMS.txt, so everything below is the data as published rather than a damaged local copy.

The 113,924 beats reproduce exactly, and so does every published percentage. The abstract’s breakdown — 33% normal sinus, 12% sinus with pre-excitation, 11% cRBBB, 15% AVRT, 8% AVNRT, 12% paced (7% atrial, 5% ventricular), 7% premature — recomputes as 32.6%, 12.1%, 10.6%, 15.4%, 7.5%, 12.1% (7.1%/5.0%) and 6.7%. Nothing to correct.

But 113,924 is not the number of annotations in the files; 118,214 is. The published figure counts exactly the beat classes the README tabulates. It omits 1,824 Q (unclassifiable) beats, 228 ~ signal-quality change markers — which are not beats, and which the ANNOTATORS file does mention — and 2,238 + rhythm markers. So the release is right about what it counts; a naive len(annotation.symbol) is 3.8% higher. ECGBench exposes n_beats (113,924), n_unclassifiable, n_quality_marks, n_rhythm_changes and n_annotations separately for that reason.

The total recording time does not reconcile. The abstract reports 1,075.85 minutes. The headers give 1,107.02 minutes (66,421 s, 18.45 h), and the shipped ecg_duration values agree with them to 1,107.07 minutes. The 31.2-minute gap is almost exactly the length of x005 (30.9 min): the total excluding x005 is 1,076.09 minutes, still 14 s off. x005 is also the one record whose ecg_duration disagrees with its header — 0:30:58.974 against 1,856.169 s, a 2.8 s difference, where the other 38 agree to within 0.01 s. So the published total appears to omit x005, but we cannot confirm that from the release. The recomputed 18.45 hours is what the entry above reports.

The min and max do match: 0:01:17 to 2:30:31, recomputed as 77.66 s and 9,031.75 s.

One age is malformed. x007 ships age = .14.3, which no float parser accepts. The loader strips the single leading . — giving 14.3, in range for this cohort — logs a warning, and keeps the verbatim string in age_raw so the repair is visible rather than assumed.

dataset_info.csv has one row wrong. Its PVC row reads PVC,PVC,V,Premature ventricular beats, i.e. symbol PVC with aux string V, while the README and the data both use symbol V with no aux string. No annotation anywhere carries the symbol PVC.

The children’s CSV column is spelled ap_loacation. ECGBench exposes it as ap_location; the typo is the source’s. It is empty for the 13 children with AVNRT, which has no accessory pathway — that is correct, not missing data.

The signals are clipped at the recording amplifier’s rail, and the rail differs per channel and per record: ±10.24 mV on most channels, ±20.48 mV on x007, ±51.20 mV on x002’s RVA12, ±49.83 mV on x0023’s ABL_uni. Six records reach the rail on the surface ECG itself — x001, x002, x003, x004, x006 and x109 — with x001 clipped on 11 of its 12 ECG channels.

amplitude_range_mv is therefore set to ±52.0, which is not a physiologic bound and is not comparable with the other datasets’ ranges. These records mix surface ECG with intracardiac electrograms, and check_amplitude_outlier applies one range to every channel: at ±10 mV, 24 of the 39 records fail — not because anything is wrong with them, but because an intracardiac electrogram is not a surface-ECG amplitude. ±52 sits just above the widest rail measured, so it excludes nothing while still catching a record that is broken outright.

Use the cardiologist’s own signal-quality annotations instead. The 228 ~ markers, exposed as n_quality_marks, appear in 10 records — including all six of the clipped ones — which makes them the better quality signal here.

Validation is otherwise clean: 0 NaN samples, 0 absent channels, 0 corrupt headers. Exactly one record is excluded from clean/: x0027, whose ABL12 channel is a constant 0.076 mV for all 77.7 s — a disconnected ablation catheter, caught by flat_line. So clean/ holds 38 records and original/ 39.