Norwegian Endurance Athlete ECG Database

28 twelve-lead resting ECGs, one per elite Norwegian endurance athlete — 24 rowers, 2 kayakers and 2 cyclists, 19 men and 9 women aged 20–43, training roughly 800 hours a year — recorded at the University of Oslo in February and March 2020 on a GE MAC VUE 360.

This is the smallest dataset in ECGBench, and size is not what it is for. Every record carries two independent interpretations, appended to its WFDB header as comment lines: one from the GE Marquette SL12 algorithm (version 23, v243) and one from a cardiologist trained in athlete ECG interpretation using international criteria. The dataset exists to show how badly general-purpose automated interpretation over-calls findings in hearts remodelled by endurance training, and it does: SL12 reads 13 of the 28 records as borderline or abnormal where the cardiologist reads them as normal, and raises a critical ACUTE MI/STEMI alert on 4 athletes, three of whom the cardiologist signs off as a plain “Normal ECG”.

So SL12 is the system under test here, not the ground truth. Training a model against the sl12_* fields inverts the point of the release.

No echocardiography was performed, so the authors note it is not confirmed whether these athletes had athletic remodelling of the heart. Each athlete contributes exactly one record, so there is no within-patient leakage and folds are stratified but not grouped.

Every lead of every record was independently min–max normalised to the full signed 16-bit range. All 336 lead-records have their minimum at exactly -32767 and their maximum within 8 counts of +32767, so with the headers’ nominal 50000/mV gain every lead spans exactly ±0.6553 mV.

This is undocumented upstream — PhysioNet states only “16-bit, 50000/mV” — and was established from the files. Reproduce it in one line:

import glob, numpy as np, wfdb
print({wfdb.rdrecord(f[:-4], physical=False).d_signal.min() for f in glob.glob("*.hea")})
# {-32767}   -- every lead of every record, not a coincidence

Two consequences, both load-bearing:

• Amplitude is meaningless in absolute terms and incomparable between leads. Voltage criteria — LVH thresholds, ST elevation in millimetres — cannot be computed from this release, which is a real loss given that athlete ECG criteria are largely voltage criteria. Morphology and timing are unaffected.
• There is no scale factor that fixes it. The normalisation was per-lead, so the information needed to invert it is not in the release. signal_unit_scale is 1.0 and units= cannot help; ECGBench hands back what the headers declare and says so rather than implying calibration.

A knock-on effect: the missing_leads and flat_line checks cannot fire on this dataset. A disconnected or railed lead would be rescaled to full amplitude like any other, so the two checks that exist to catch dead leads are blind here. amplitude_range_mv is therefore set to a deliberately tight ±1.0 — not a physiological range, but one that still passes all 28 records while flagging any future record that is not normalised this way.

All 58 shipped files were verified against the release’s own SHA256SUMS.txt before any figure here was computed — all OK. The record and subject counts match the release description exactly at 28 and 28, so there is nothing to reconcile.

Every figure above is recomputed from the 28 .hea files, which are the only place the labels exist — the release ships no metadata table at all (just the .dat/.hea pairs, RECORDS, SHA256SUMS.txt and LICENSE.txt). ECGBench parses the #SL12: and #C: comment lines; see ecgbench/labels/norwegian_athlete_ecg.py.

The findings table is multi-label and does not sum to 28. The 28 cardiologist readings carry 43 findings between them (1–4 each); the 28 SL12 readings carry 55 across 27 distinct statements. Every record carries at least one finding, so none falls into no class.

The findings table is case-folded; the label loader is not. The source spells the same finding two ways — normal sinus rhythm once against Normal sinus rhythm 16 times, and first degree AV block once against First degree AV block twice. The table above folds case (hence 17 and 3); cardiologist_findings preserves each string verbatim, because silently rewriting a clinician’s text is worse than reporting it. Fold case yourself if you are counting.

The stratification label is a different quantity from this table. Folds are stratified on cardiologist_primary_rhythm — the rhythm each cardiologist reading opens with, one per record, 16 / 7 / 5. The cardiologist’s verdict would be degenerate for that purpose: 26 of 28 are “Normal ECG”. Use cardiologist_findings for training targets, not the stratification label.

There is no supervised task with a positive class on human ground truth. Not one record is read as abnormal by the cardiologist. Anyone wanting a normal/abnormal target has only the 2 borderline records, or the SL12 output — which is the thing being evaluated, not a label.

No per-record demographics ship. Age, sex and sport are published only as cohort aggregates on the landing page, so they cannot be joined to a record and ECGBench does not expose them.

Three parsing traps live in these header lines, all handled and all regression-tested:

• A comma does not reliably end a statement. Three GE statements contain commas of their own — ST elevation, consider early repolarization, pericarditis, or injury alone splits into four fragments on a naive comma split, turning 2 records into 7 spurious findings.
• Capitalisation cannot be used to detect the continuations, because the cardiologist writes genuine findings in lowercase (Sinus bradycardia, normal sinus rhythm, First degree AV block). ECGBench masks an explicit list of known comma-carrying statements instead of guessing.
• One verdict is misspelt: ath_010 ends Abnormal EKG, not “ECG”. The normalised field folds it in; the raw string keeps it.

With 28 records over 10 folds, folds 1–8 hold 3 records each and folds 9 and 10 hold 2 — so the default mapping gives 24 train / 2 val / 2 test. Both val and test end up holding only Normal sinus rhythm records, and with 2-record folds and a 16/7/5 rhythm split no assignment avoids that. scikit-learn also warns that the least populated class has 5 members against n_splits=10; that is expected and does not raise.

Treat this as a cross-validation harness rather than a train/val/test split: pass split=None with fold_numbers and rotate which folds you hold out. The rhythm is spread as evenly as 28 records permit across folds 1–8.

from ecgbench import ECGDataset

# Custom CV: hold out fold 3, train on the rest. With split=None each
# sample's ["split"] reports that record's own default split.
train = ECGDataset("norwegian_athlete_ecg", split=None,
                   fold_numbers=[1, 2, 4, 5, 6, 7, 8, 9, 10],
                   data_path="/path/to/norwegian-athlete-ecg/1.0.0/")
held  = ECGDataset("norwegian_athlete_ecg", split=None, fold_numbers=[3],
                   data_path="/path/to/norwegian-athlete-ecg/1.0.0/")
len(train), len(held)               # (25, 3)