MedalCare-XL (Synthetic 12-Lead ECGs from Simulations)

16,842 twelve-lead ECGs that no heart ever produced. Every record is the output of a multi-scale electrophysiological simulation: atrial and ventricular models were run independently over anatomical meshes, composed into 10 s rhythms at 500 Hz, and projected onto torso models to give a surface 12-lead trace. Eight conditions ship, one of them subdivided eight ways.

That makes it the odd one out in this catalogue, and worth being precise about why. The label is exact by construction — it is the setting the simulator was given, so there is no reader disagreement, no comorbidity, no borderline case and no diagnostic uncertainty. A model that separates these classes perfectly has learned to separate simulator configurations, which is a weaker claim than separating patients. Used as intended — pre-training, augmentation, controlled ablations where you need to vary one physiological parameter and hold the rest fixed — the exactness is the point. Used as a benchmark on its own, it will overstate.

Three practical things to know before loading it:

The CSVs are transposed relative to every other CSV dataset here. Each file is 12 rows × 5000 columns with no header row — one row per lead, samples running along the row. chapman_shaoxing and ningbo_iva are the other way round, samples in rows under a header naming the leads. ECGBench gives this its own signal_format, csv_lead_rows, because reading one layout with the other’s reader returns a plausibly-shaped array of the wrong thing rather than raising.

Each record ships three times. <n>_raw.csv is the noise-free simulator output, <n>_noise.csv adds superimposed noise, and <n>_filtered.csv applies the release’s 0.5–150 Hz order-3 Butterworth pair. These are one record in three renderings, not three records. ECGBench wires up filtered, the closest of the three to a recorded ECG; the other two are one string substitution away and the label loader exposes both.

The split is the authors’ own, not a generated one. Their train/validation/test directories become folds 1, 2 and 3, so --n-folds has no effect here and there are three fold CSVs rather than ten. See “The split guarantee has one hole” below for the one place it does not do what the README says.

The title says 16,900 and what ships is 16,842. Both the paper’s title and the Zenodo title read “16,900 … electrocardiograms”. The v1.3 deposit holds 16,842, and there is no changelog explaining the difference.

The recomputed figure was derived three independent ways, which all agree:

• 16,848 *_raw.csv files under WP2_largeDataset_Noise/, minus the 6 in mi/examples/, which are figure illustrations sitting outside the <split>/run_<model>/ layout and are byte-distinct from every record in mi/*/test/run_S62/;
• 16,842 *_AtrialParameters.txt files;
• 16,842 *_VentricularParameters.txt files.

The shortfall is concentrated in iab, the only class that does not reach a round total (1,244 against 1,300 for its peers). 13 of the 186 run directories have gaps in their file numbering — iab/train/run_S66 holds 107 files numbered up to 000130, and iab/test/run_S62 is missing numbers 17, 71, 72 and 75 out of 130. ECGBench enumerates records from the files that exist rather than from a range, so a gap costs a record rather than producing a broken path.

Class counts are single-label and sum to the record total, unlike the multi-label datasets in this catalogue. Every record carries exactly one condition and none is unlabelled. The 15-class pathology_subclass above is also what the config stratifies on, so the split table and the stratification label are the same quantity here — which is not true of most pages in this catalogue.

Myocardial infarction is 46.3% of the release and each remaining condition is 7.4–7.7%. That is an authoring choice about how finely to subdivide MI, not a prevalence.

The local copy was checksum-verified before any figure above was computed. MedalCare-XL.zip has md5 96497fcbc5c443bec1280f2033836776, matching Zenodo’s published value, and the extracted tree is an exact bidirectional match with the archive listing — 84,416 files, 0 missing, 0 extra.

The release README states that ECGs “calculated with the same anatomical model but different electrophysiological parameters are only present in one of the test, validation and training datasets but never in multiple.”

That holds within each condition and fails across them. Ventricular model S64 is the test-side model for sinus, AV block, LBBB, RBBB and every MI subclass — and simultaneously a train-side model for the three atrial conditions (fam, iab, lae). S67 does the same for validation against train. Eleven of the thirteen models sit in exactly one fold; those two straddle.

This is not a name collision. Checked at the parameter level: sinus/test/run_S64 and fam/train/run_S64 agree on all 84 ventricular parameters that are constant within a run directory — same ionic model, same conductivities, same action-potential durations, same fibre angles. It is one ventricular model appearing on both sides of the split.

No records are shared, though, so this is not duplicate-row leakage. Two independent checks:

• No two records in those directories have the same full ventricular parameter vector — 0 matches over 100 × 130 pairs.
• A waveform signature over a 600 ms QRS-centred window across all 12 leads puts the nearest cross-directory pair at distance 1.13, against a within-directory median of 3.78. The signature detects exact duplicates at distance 0.0 (verified on a self-match), so the null result is evidence of absence rather than of an insensitive measure.

What it means in practice: if you train a classifier that spans the atrial conditions (fam/iab/lae) and the ventricular ones (everything else), the same ventricular anatomy appears in both your training and your test set. Within either arm alone, the guarantee holds as stated. model_id is exported in every fold CSV precisely so this is checkable — and regroupable — by hand rather than being something you have to take on trust.

All 32 exclusions fail amplitude_outlier, and 30 of the 32 fail on V1 (the other two on V2 and V3). The magnitudes are not marginal: 10.1 to 879.6 mV, median 25.7 mV, where a random 400-record sample of the release peaks at a median of 2.09 mV with a 99th percentile of 5.30 mV. There is a wide empty gap between the bulk of the release and the failures, so the [-10, 10] mV threshold is not slicing into a continuum.

Two things rule out a configuration mistake on ECGBench’s side. First, the blowups are already present in the raw simulator output, before the noise and filtering stages ever ran, so they are a forward-solution artefact rather than something the release’s processing introduced. Second, if the unit scale were wrong, every record would fail, not 0.19% of them.

V1 is the precordial electrode closest to the ventricular surface, which is where a torso mesh node sitting too near the source would produce exactly this. The exclusions span six of the eight conditions (24 of the 32 are MI records), so it is not a property of one class.

They are not deleted — version="original" returns all 16,842 with is_valid and quality_issues marking them, and output/medalcare_xl/validation_report.json names each one and the lead and range that failed.

Every record ships two plain-text parameter files alongside it: <n>_AtrialParameters.txt (~21 keys) and <n>_VentricularParameters.txt (~105 keys), holding the ionic model, tissue conductivities, regional conduction velocities, action-potential durations, fibre angles, stimulus sites, ischaemic-region geometry, the atrial and torso meshes and all ten electrode positions. For a simulated dataset these are the ground truth — the pathology label is a summary of them.

ecgbench.labels.medalcare_xl.load_simulation_parameters returns them as a frame indexed by record id, with atrial. and ventricular. prefixes because the two files reuse key names (im.name, G.torso) and an unprefixed concat would silently drop one of each. It is opt-in: labels=True does not read them, because that would be 33,684 file opens on every dataset construction. Pass record_ids= to restrict it to a split.

The key set is ragged, and that is information rather than damage:

So the union is 119 ventricular and 21 atrial keys, of which 70 and 17 appear in every condition. A NaN means the condition has no such parameter, not that a value is missing.

One thing ECGBench deliberately does not read: the per-run-directory siginfo.csv. For fam, iab and lae its info2 column holds a foreign simulation id rather than the record number, and in 13 of the 186 run directories it has more rows than there are records — so joining it means guessing at row order, on exactly the tables where the guess is least checkable. The parameter files carry the same anatomy (geo.atria, geo.torso) keyed by record number, so nothing is lost by ignoring it.

Not generated — adopted. The release’s own train/, validation/ and test/ directories become folds 1, 2 and 3, giving 12,022 / 2,434 / 2,386 (71.4% / 14.5% / 14.2%, matching the README’s stated ~70/15/15). Because the splits are predefined, --n-folds has no effect and there are three fold CSVs, not ten.

patient_id_column is set to model_id — the ventricular simulation model, S62–S74. There are no patients here, and rather than leave the field null this points at the grouping unit that plays the same role: the thing the authors defined the split around, and the one whose reuse is worth being able to see. Read “The split guarantee has one hole” above before assuming it is fold-disjoint; two of the thirteen models are not.

Fold membership is identical between original/ and clean/ — clean/ is a row subset of 16,810, not a re-split.

With one train fold and one test fold there is no ten-fold rotation to run here. If you want cross-validation, regroup by model_id, which every fold CSV carries.