IKEM Dataset (Institute for Clinical and Experimental Medicine, Prague)

98,130 ECGs from 30,290 patients recorded in routine care at IKEM in Prague, a national tertiary referral centre for cardiology, transplantation and diabetology. Acquisition runs 2004-2022 but is overwhelmingly 2017 (57,179 records) and 2018 (33,355).

Four facts about this release differ from what it says about itself, and each was measured from the files rather than taken from the description. They are the reason this page is long.

1. Eight leads, not twelve — and the order is the most unusual in the catalogue. Only the independent leads are stored, as V1, V2, V3, V4, V5, V6, II, I: precordial first, and II before I. So signal[0] is V1, not lead I. III, aVR, aVL and aVF are exact linear combinations of II and I and are simply absent; ECGBench returns the 8 stored leads and does not synthesise the others. Deriving them is exact arithmetic — the example script shows it.

2. It is 8.192 s, not 10. The description says “sampled at a rate of 500 Hz for a duration of 10 seconds each”, which would be 5,000 samples. The arrays hold 4,096. The rate really is 500 Hz — see below for how that was settled — so 4,096 / 500 = 8.192 s.

3. The samples are microvolts, not the 4.88 µV counts stated. Applying the declared granularity makes the median per-record peak 9.5 mV, roughly five times physiologic.

4. No diagnoses ship. exams.csv carries age, sex, weight, height and two cart-measured rates, and nothing else. The paper trains on IKEM diagnostic labels that are not part of this release. This dataset looks like a 98,130-record classification corpus and is not one: what it supports is age and sex estimation, rate regression and self-supervised pretraining.

Patients repeat heavily. 19,078 of the 30,290 patients contributed more than one recording and one contributed 96 — 86,918 records (88.6%) belong to a multi-record patient. ECGBench’s folds are grouped on patient_id, a 40-character SHA-1 surrogate. Unusually, there is nothing else hiding: every one of the 98,130 waveform SHA-1s the release ships is distinct, so unlike CODE-15% there are no byte-identical duplicate recordings behind different patient ids.

The release ships a LICENSE file that is the verbatim text of CC BY-NC-ND 4.0. The NoDerivatives term makes republishing a derived fold table — 98,130 exam_ids and 30,290 patient hashes — on a public, ungated, commercially operated repository legally unclear, and publication is effectively irreversible. So ECGBench withholds it. This is the second such dataset after MIMIC-IV-ECG and the first withheld by licence rather than by a data use agreement.

(Zenodo’s own metadata field says only other-at — “Other, Attribution” — which is both less specific and less restrictive than the licence text travelling with the data. The licence text governs. If the depositors confirm that a fold table is acceptable, this decision should be revisited.)

The split is distributed as a recipe instead. Fold assignment is a pure function of exams.csv and a fixed seed (random_state=42), so anyone can regenerate the identical partition:

1. ecgbench splits --dataset ikem --data-path /path/to/IKEM_dataset_v1.0.0/
2. python -c "from ecgbench import verify_splits; verify_splits('ikem', 'output/ikem')"
3. cp -r output/ikem/clean output/ikem/original /path/to/IKEM_dataset_v1.0.0/

Step 2 compares your run against the reference manifest shipped in the package (ecgbench/data/manifests/ikem.json — 728 bytes of seed, checksums, counts and one fold digest, no identifiers) and raises on mismatch, naming the differing input file. Step 3 is what lets metadata_source="local" find the fold CSVs: in local mode data_path serves as both the signal root and the splits root.

ECGDataset("ikem") with the default metadata_source="hf" raises SplitsNotPublishedError quoting these steps, rather than a bare 404.

All five published files match the md5s Zenodo lists and the shipped zenodo_md5sums.txt (verified with md5sum -c). The headline figures — 98,130 records, 30,290 patients — agree with the release, and the three HDF5 parts hold exactly 48,264 + 48,683 + 1,183 = 98,130 rows whose exam_id sets match exams.csv exactly, with no padding row (CODE-15% has one per part).

-1 is a missing-value sentinel in every numeric column, and the missingness is severe rather than incidental: weight 89.6%, height 89.3%, age 9.0%, atrial rate 0.4%, sex 0.4%, ventricular rate 0.006%. Because the sentinel is a value rather than a blank, notna() on the raw CSV reports all 98,130 records complete and every summary statistic comes out wrong — read literally, the cohort’s mean weight is about −76 kg. ECGBench converts every -1 to NaN, exposes is_male as a nullable boolean so its 376 unknowns survive, and adds has_age/has_weight/ has_height. That is lossless: the source has no genuine blanks and no genuine -1 measurement.

Values that survive the filter still deserve suspicion — 546 records give age 0, 21 give 100 or more, and one gives a ventricular rate of 0. Those are left as they are, since they are not sentinels and guessing which are real is not a loader’s job.

acquisition_date is MM-DD-YYYY, which sorts wrongly as a string: taking string min/max suggests a 2018-2021 range where the true range is 2004-03-18 to 2022-07-26. Day-first parsing would silently swap month and day for the first 12 days of every month.

48 records are shorter than they look. The arrays are rectangular at 4,096 samples, but the parts carry a real_lengths dataset and 48 records hold only 2,500 real samples (5.0 s), zero-padded into the rest. That column is exposed as real_length_samples.

The stratification label is a measurement, not a diagnosis. With no diagnoses to stratify on, folds use the cart’s ventricular rate banded into BRADY (12,387), NORMAL (76,426) and TACHY (9,317). A rate of 75 bpm is as compatible with atrial fibrillation as with sinus rhythm. Never train on it.