ECGBench

Reproducible ECG benchmark datasets with standardised splits, validation, and Croissant metadata.

ECGBench provides a curated catalogue of 64 publicly available ECG datasets, a config-driven pipeline for generating validated fold splits, and a unified PyTorch Dataset class for loading any supported dataset.

Installation

Base (config, catalogue, validation, splitting)

pip install ecgbench

With PyTorch support

pip install ecgbench[torch]

With HDF5 datasets

sph stores one HDF5 array per record, which needs h5py. It is the only dataset that does, so the dependency is its own extra:

pip install ecgbench[hdf5]

With everything

pip install ecgbench[all]

From source (development)

git clone https://github.com/vlbthambawita/ECGBench.git
cd ECGBench
uv pip install -e ".[dev]"

Quick Start

from ecgbench import ECGDataset, ecg_collate_fn
from torch.utils.data import DataLoader

# Load PTB-XL training data (downloads fold CSVs from HuggingFace Hub)
train_ds = ECGDataset("ptbxl", split="train", data_path="/path/to/ptb-xl/1.0.3/")
loader = DataLoader(train_ds, batch_size=32, collate_fn=ecg_collate_fn)

for batch in loader:
    signals = batch["signal"]   # (B, 12, 5000) float32 tensor
    ecg_ids = batch["record_id"]
    break

Dataset Catalogue

Query the curated index of 64 ECG datasets:

import ecgbench

# List all datasets
datasets = ecgbench.list_datasets()
print(f"{len(datasets)} datasets available")

# Search by name, origin, format, or paper
ecgbench.search("PTB-XL")

# Filter by category and access type
ecgbench.search(category="12-Lead (PhysioNet)", access="Open")

# Look up a single dataset
ecgbench.get_dataset("MIMIC-IV-ECG")

# List categories
ecgbench.categories()

# Get as pandas DataFrame
df = ecgbench.to_dataframe()

Loading ECG Data

Standard train/val/test splits

from ecgbench import ECGDataset, ecg_collate_fn
from torch.utils.data import DataLoader

train_ds = ECGDataset("ptbxl", split="train", data_path="/data/ptb-xl/1.0.3/")
val_ds = ECGDataset("ptbxl", split="val", data_path="/data/ptb-xl/1.0.3/")
test_ds = ECGDataset("ptbxl", split="test", data_path="/data/ptb-xl/1.0.3/")

loader = DataLoader(train_ds, batch_size=32, collate_fn=ecg_collate_fn)

Selecting specific folds

fold_numbers picks individual folds out of a split. Folds are 1-indexed.

ECGDataset("ptbxl", split="train", fold_numbers=[3], data_path="...")       # one fold
ECGDataset("ptbxl", split="train", fold_numbers=[1, 2, 5], data_path="...") # several

Each fold belongs to exactly one split — 1-8 under train/, 9 under val/, 10 under test/ — so split="train", fold_numbers=[9] is an error. To select folds regardless of that layout, for custom cross-validation, pass split=None:

# Hold out fold 7 as test and fold 10 as val, train on the other eight.
test  = ECGDataset("ptbxl", split=None, fold_numbers=[7],  data_path="...")
val   = ECGDataset("ptbxl", split=None, fold_numbers=[10], data_path="...")
train = ECGDataset("ptbxl", split=None,
                  fold_numbers=[n for n in range(1, 11) if n not in (7, 10)],
                  data_path="...")

split=None requires fold_numbers, and each returned sample's ["split"] reports the record's own default split rather than one name for the whole set. Unlike stitching per-split datasets together with ConcatDataset, this returns a single ECGDataset, so .metadata_df and .labels_df still describe the whole selection.

Labels

Fold CSVs are identification-only by design — record ID, patient ID, signal paths, fold, split. Ground truth stays with the source dataset, so labels=True needs a local copy of it:

ds = ECGDataset("ptbxl", split="train", data_path="/data/ptb-xl/1.0.3/", labels=True)

ds[0]["labels"]["superclasses"]   # ['MI', 'STTC']  — multi-label
ds[0]["labels"]["report"]         # the cardiologist's text
ds.labels_df                      # the whole split's labels, aligned to metadata_df

Or without a Dataset at all, for class weights and filtering:

from ecgbench import load_labels

labels = load_labels("chapman_shaoxing", data_path="/data/chapman-figshare/")
labels["Rhythm"].value_counts()

Each dataset exposes its own fields — SCP codes plus diagnostic super/subclasses for PTB-XL, SNOMED-CT codes for ecg_arrhythmia, rhythm/beat annotations and eleven automated measurements for chapman_shaoxing, free-text machine reports plus nine interval/axis measurements for mimic_iv_ecg, reference beat counts for incartdb, protocol phase and balloon-occlusion timings for staffiii, AHA/ACC/HRS statements with their modifiers for sph, the ablation-confirmed arrhythmia origin for ningbo_iva. A dataset that genuinely has none (mimic_iv_ecg_demo) raises LabelsUnavailableError naming where labels could come from, rather than returning empty columns.

Leads and units

Select and reorder leads by name, and choose the output unit:

ds = ECGDataset("mimic_iv_ecg_demo", split="train", data_path="...",
                leads=["I", "II", "aVL", "V5"], units="uV")

ds[0]["signal"].shape   # (4, 5000)
ds.lead_names           # ('I', 'II', 'aVL', 'V5')
ds.units                # 'uV'

Names, not indices, because lead order is not consistent across datasets:

One dataset has no physical units at all. echonext ships waveforms its publisher median-filtered, percentile-clipped and standardised with an unreleased mean and SD, so no scale factor recovers millivolts. Its config declares signal_units: zscore, and units= refuses rather than silently multiplying dimensionless numbers by 1000:

ds = ECGDataset("echonext", split="test", data_path="...", metadata_source="local")
ds.units                       # 'zscore' -- reported honestly, not 'mV'
ds[0]["signal"].min()          # -6.829

ECGDataset("echonext", units="uV", ...)
# UnitConversionError: This dataset's samples are stored as 'zscore', not a
# physical unit, so they cannot be converted to 'uV'. ...

Every other dataset declares signal_units: mV (the default) and is unaffected. amplitude_outlier validation is skipped for non-mV sources, since its thresholds are millivolts.

One dataset's millivolt scale is an estimate rather than a declared value. ningbo_iva ships bare integers, and neither its paper nor figshare states a gain — the paper's own figures plot the raw counts. Its signal_unit_scale of 6.1035e-05 (1 mV = 2¹⁴ counts) was measured by comparing median lead-II R-peak amplitude, sex for sex, against sph, whose samples are millivolts by declaration; the two sexes bracket it at 14,029 and 17,111 counts/mV. Waveform shape is exact; absolute calibration is good to roughly ±20%. Divide the millivolt values by 6.1035e-05 to recover the shipped integers if you would rather calibrate them yourself.

signal[4] is aVL in most of them and aVF in both MIMIC datasets, so slicing by index across datasets silently crosses two leads. Matching is case-insensitive — leads=["aVL"] works on the lowercase datasets too — an unknown lead lists what is available, and a duplicate is rejected.

Three datasets are not 12-lead at all. STAFF III stores only 9, and in the opposite order to everything else: V1-V6 first, then I, II, III. aVR, aVL and aVF are exact linear combinations of I and II and were never stored, so the montage is a standard 12-lead one clinically while signal[0] is V1 rather than lead I — the single most likely way to misread this dataset. PTBDB stores 15 signals, the conventional twelve plus the three Frank vectorcardiography leads; leads= is how you take the standard twelve out of it. wctecgdb stores 37: I/II/III, V1-V6, the three limb electrode potentials LA/RA/LL and the six true unipolar chest leads UV1-UV6, each present both raw and after DC removal plus a 0.05-150 Hz band-pass, plus the Wilson Central Terminal itself. Index 0 is raw lead I and index 18 is filtered lead I — the same signal in two preprocessing states — so leads= by name is the only safe way to read it, and aVR/aVL/aVF have to be derived from I and II. Its records are also variable length (32 s to 120 s), so batching needs a fixed window= — see examples/load_ptbdb.py.

leipzig_heart_center_ecg goes further: it is the one dataset where the channel count is not constant. Every record holds the 12-lead surface ECG plus the intracardiac electrograms from whichever catheters were in place, giving 14, 18, 19 or 20 channels in six distinct layouts — and only channels 0-11 are the same channel in the same position in every record (index 12 is ABL12, RVA12 or ART depending on the record). Its lead_names therefore declares the ECG and nothing else, deliberately, so leads= resolves to the right physical lead everywhere. To reach an intracardiac channel, look it up by name in that record's own header:

from ecgbench.labels.leipzig_heart_center_ecg import channel_index

channel_index(labels["channel_names"], "RVA12")   # 13 in most records, 18 in x100
channel_index(labels["channel_names"], "CS12")    # None where that catheter is absent

Pass leads= if you want a homogeneous batch; without it a batch mixes 14-, 18-, 19- and 20-channel tensors. See examples/load_leipzig_heart_center_ecg.py.

incartdb is the one dataset whose primary labels are reference beat annotations rather than record-level diagnoses: 175,907 manually corrected beats over ten types, exposed as per-record counts (beat_N, beat_V, …, pvc_fraction) alongside the patient diagnosis and free-text per-record findings. Its records are 1800 s (~44 MB each), so batching needs a window=. It is also the clearest case for patient-grouped folds — 3,166 of its 3,174 RBBB beats come from a single patient — see examples/load_incartdb.py.

staffiii is the one dataset whose label is a position in a procedure rather than a diagnosis. Each of its 104 patients was recorded before, during and after an elective coronary angioplasty, so recording_type (BR/BC/BI/PC/PR) marks which recordings were taken while a balloon was occluding a coronary artery — 152 inflations, 28-595 s each, with sample-accurate inflation, deflation and contrast-injection times from the shipped .event files. That makes it the reference set for transient ischaemia, with each patient as their own control. Two traps: its 9 leads start with V1, and record length correlates strongly with the label (inflation records have a median of 518 s against 300 s elsewhere), so window to a fixed length before training. See examples/load_staffiii.py.

brugada_huca is the smallest and cleanest dataset here — 363 records, one per subject, all 363 passing validation — and the only one sampled at 100 Hz alone (PTB-XL offers 100 Hz as an alternative to 500). Its labels are bare integers with no string form in the CSV: brugada is 0 healthy / 1 confirmed / 2 other-atypical, and ecgbench.splitting.strategies.brugada_huca.BRUGADA_CLASSES carries the meanings. Treat it as a screening cohort: class 0 means "investigated and not diagnosed", not a general-population control. See examples/load_brugada_huca.py.

mimic_iv_ecg is the largest dataset here — 800,035 records from 161,352 patients (~96.5 GB) — and the one where fold_numbers= matters most: a single fold is a tenth of it. Two facts to know before using its labels. They are free-text machine reports (up to 18 lines per study, joined into report_text), not codes, and primary_report is only the first line, which is sometimes a data-quality warning rather than a rhythm. And its numeric measurements encode "not measurable" as integer sentinels — 29999, 32767, 65535 — which ECGBench converts to NaN; read the CSV yourself and a mean P-wave axis comes out meaningless. See examples/load_mimic_iv_ecg.py. Its 659-record open demo is a separate config, mimic_iv_ecg_demo, which has no labels at all.

challenge2021 and challenge2020 are the datasets where sampling rate varies per record (257/500/1000 Hz), because each concatenates several source cohorts. Rate is therefore a label to filter on, not a sampling_rate= argument, and record length spans 5 s to 1800 s so batching needs a window= too. challenge2021 contains PTB-XL, PTBDB, INCART, CPSC-2018, Chapman-Shaoxing and Ningbo; challenge2020 contains the first four. Their source label says which cohort each record came from, and evaluating on any of those after training is testing on training data. See examples/load_challenge2021.py and examples/load_challenge2020.py.

The two challenge years are the same recordings. All 43,101 challenge2020 records are in challenge2021, bit-identical — verified against both releases' published SHA256SUMS.txt. They are separate configs because the label encodings differ: 2020 scored 27 classes and 2021 scored 30, and 631 of the 2020 headers list a SNOMED code twice inside their own #Dx field (ecgbench.labels.challenge2020 deduplicates them, which is what makes the shipped data reproduce the official code table). Never train on one year and evaluate on the other.

cpsc_2018 is the CPSC-2018 public training set as a dataset in its own right — 6,877 records, one rate (500 Hz) but 6 s to 144 s record length, so window= is mandatory and must fit the 6 s minimum. Its nine classes are multi-label (476 records carry two or three), and its primary diagnosis is gone: the WFDB copy everyone uses sorted each #Dx list by class index, so CPSC's original First/Second/Third labelling is unrecoverable and stratify_dx is a folds-only reduction. All 6,877 records are byte-identical to the cpsc_2018 cohort of both challenge years, under the same A#### names — so this is the fourth way into the same recordings. See examples/load_cpsc_2018.py.

sph is the largest single-source dataset here — 25,770 records from 24,666 patients at one Chinese hospital — and the only one stored as HDF5 (pip install ecgbench[hdf5]), one (12, N) float16 array per record, already in millivolts. Its labels are AHA/ACC/HRS standardised statements rather than a bespoke vocabulary: 44 primary statements in 11 categories, each optionally qualified by one of 15 modifiers, so a record reads 60+310;147. 14.45% of records carry more than one statement and there is no primary diagnosis, so stratify_code is a folds-only rarest-code reduction. 1,066 patients contributed 2-5 records, so folds are grouped on patient_id — the grouping was verified on the output, not assumed. Length runs 10-56 s, and the metadata's N column gives it exactly per record, so nothing has to open a signal file to learn a length. See examples/load_sph.py.

ningbo_iva is the only dataset here whose label is invasive ground truth: 334 12-lead ECGs recorded during catheter ablation, each labelled with the outflow tract (RVOT 257 / LVOT 77) the ablation proved the arrhythmia came from, so the task is to predict the origin from the surface ECG before the procedure. Three things about it are unlike everything else: the lead order is alphabetical (signal[0] is aVF), the sampling rate is 2000 Hz — the highest in the catalogue, from an EP-lab system rather than a diagnostic cart — and the samples carry no declared unit, so the millivolt scale is an ECGBench estimate (see "Leads and units"). Length runs 2.9-59.3 s in 317 distinct values over 334 records. See examples/load_ningbo_iva.py.

norwegian_athlete_ecg is the smallest dataset here — 28 records, one per elite Norwegian endurance athlete — and the only one whose amplitudes are not calibrated. Every lead of every record was independently min-max normalised to the full int16 range (all 336 lead-records bottom out at exactly -32767), so with the headers' nominal 50000/mV gain each lead spans exactly ±0.6553 mV. Absolute and inter-lead voltages are therefore gone — no LVH or ST-elevation-in-mm criteria — and no signal_unit_scale or units= can undo a per-lead normalisation. Morphology and timing survive. This is undocumented upstream and was established from the files. A knock-on effect: missing_leads and flat_line cannot fire on it, because a dead lead would be rescaled to full amplitude like any other.

It is also the only dataset carrying two independent interpretations per record, as WFDB header comments: the GE Marquette SL12 algorithm's and a cardiologist's, exposed as separate sl12_* and cardiologist_* label fields. SL12 is the system under test, not the ground truth — it reads 13 of 28 records as borderline or abnormal where the cardiologist reads normal, and raises a critical ACUTE MI/STEMI alert on 4 athletes, three of whom the cardiologist calls a plain "Normal ECG". Human labels are degenerate (26 of 28 "Normal ECG", no abnormal class at all), so folds are stratified on cardiologist_primary_rhythm instead, and with 2-3 records per fold you should rotate folds via split=None, fold_numbers=[...] rather than use the default 24/2/2 mapping. See examples/load_norwegian_athlete_ecg.py.

mhd_effect_ecg_mri is the one dataset where the distortion is the point: 53 ECGs recorded inside 1T, 3T and 7T MRI scanners, where the magnetohydrodynamic effect (blood ions moving through the static B0 field) superimposes a voltage that buries the P wave, ST segment and T wave. Amplitudes reach −31 mV, far past the recorders' nominal ±6 mV and ±2.4 mV input ranges, so amplitude_range_mv is ±35 — a conventional ±10 would exclude 16 of 53 records for being exactly what they are meant to be. 10 records are reference ECGs taken outside the bore for the same subjects (−0.88…+3.09 mV over the same window), standing in for the in-bore ground truth that cannot be measured. There is no diagnosis to predict: all subjects were healthy and the 14,950 manual QRS marks carry no beat classification, so the label is the acquisition condition and the task is signal separation.

It is also the one dataset whose patient ID had to be derived. Filename subject numbers are scoped per scanner — ECGMRI1T01 and ECGMRI3T01 are different people — and three slots belong to subjects recorded in more than one scanner, so grouping on the number would split one person across folds. subject_key is the sex/age/weight/height tuple instead, collapsing 29 slots into 26 people; folds are grouped on it and no subject spans a fold. Records mix 12-lead and 3-lead layouts (only I, II, III are present in every one) and run 24 s to 12 min, so batching needs both leads= and window=(0, 25000). Note the shipped release has 53 records where the README, PhysioNet page and CinC paper all say 43. See examples/load_mhd_effect_ecg_mri.py.

wctecgdb is the one dataset that measures the reference instead of assuming it. Conventional ECG treats the Wilson Central Terminal — the point V1–V6 are measured against — as 0 V; this release brings the three limb electrodes out individually so the WCT can be recorded, and its authors report amplitudes reaching 241% of lead II. Each of the 540 ten-second segments therefore holds 37 channels at 800 Hz (8001 samples, 10.00125 s): I/II/III, V1–V6, the limb electrode potentials LA/RA/LL and the true unipolar chest leads UV1–UV6, each present both raw and filtered (DC removal plus 0.05–150 Hz), then WCT. Index 0 is raw lead I and index 18 is filtered lead I, so leads= by name is the only safe way in, and aVR/aVL/aVF do not exist here at all. amplitude_range_mv is ±20 because the raw unreferenced channels carry several mV of DC offset — and 140 of the 540 records have a channel clipped at the ±9.2250 mV acquisition rail, which validation passes deliberately rather than treating saturation as damage.

Its 540 segments come from 92 patients, 1–31 each — five patients are 24% of the dataset — so folds are grouped on patient_id and any per-record rate is weighted by segment count. The only label is a patient-level free-text admission diagnosis (43 distinct strings, 10 patients with none, Windows-1252 bytes and four misspellings), which says why the patient was admitted, not what the ten seconds show; the 8-way diagnosis_group reduction exists to stratify folds, not to train on. Eight records carry precordial channels synthesised as V = UV − WCT — flagged per record, and to be excluded when evaluating precordial reconstruction. See examples/load_wctecgdb.py.

ecgrdvq, ecgdmmld and ecgcipa are the three datasets here with no diagnosis at all — sibling releases from one FDA programme, in order SCR-002, SCR-003 and SCR-004, and the set to read together because almost every convention they share, they share inverted.

ecgcipa is 5,749 ten-second 12-lead ECGs at 1 kHz (10,000 samples — the largest 12-lead tensor in the catalogue) from 60 healthy volunteers in an FDA Phase I trial, and what varies is the drug: ranolazine, verapamil, lopinavir+ritonavir, chloroquine, placebo or a dofetilide/diltiazem crossover. The labels are drug, time from dose, plasma concentration and nine interval measurements (QT, QTcF, J-Tpeak, J-Tpeakc, …), so treatment is the stratification label and everything else is continuous. Samples are microvolts (signal_unit_scale: 0.001), and units="uV" returns the source scale.

Three things to know before using it. Records come in near-duplicate triplicates — three segments per subject per timepoint, so 5,749 records are closer to 1,917 observations; patient grouping keeps each triplicate intact. Every record ships twice, as the raw segment and as a derived 16-channel median beat (+VCGMAG/X/Y/Z) whose .atr fiducials are what the published intervals were measured from — the median beats deliberately get no fold of their own. And the study's own endpoints cannot be attached to a waveform: change from baseline lives only on adeg.csv's triplicate-average rows, which carry no record ID. See examples/load_ecgcipa.py.

ecgdmmld is the same shape and inverts three of those details. 4,211 ten-second 12-lead ECGs at 1 kHz from 22 healthy volunteers in a complete 5-period crossover — every subject took dofetilide alone, dofetilide with mexiletine, dofetilide with lidocaine, moxifloxacin with diltiazem, and placebo. Samples are millivolts (signal_unit_scale: 1.0, not ecgcipa's 0.001), the limb leads are spelled AVR/AVL/AVF rather than aVR/aVL/aVF, and the 1 kHz is up-sampled from a 500 Hz acquisition. The study's endpoint is attachable here — is_baseline flags each period's pre-dose triplicate, so load_baseline_deltas() returns change from baseline per record, the thing ecgcipa cannot give you.

Its own trap is the label. treatment names the period's randomised regimen, not the drug on board: the agents were staged hours apart, so only 57% of the dofetilide-arm records contain dofetilide and a "Mexiletine + Dofetilide" record at 2 h is a mexiletine-only ECG. Stratify on it, train on the six plasma_* columns. Because the crossover is complete, every fold gets all five arms automatically and no split can separate them — and with 2–3 subjects per fold, a per-fold metric describes two or three people. See examples/load_ecgdmmld.py.

ecgrdvq is the earliest of the three (SCR-002) and the one whose label you can actually trust. 5,232 ten-second 12-lead ECGs at 1 kHz from 22 healthy volunteers in a 5-period crossover of single agents — ranolazine, dofetilide, verapamil, quinidine and placebo, one per period — so treatment names the drug rather than a staged combination, and 93–94% of each active arm's records carry a measured concentration of exactly that drug. It shares ecgdmmld's millivolts, its uppercase AVR/AVL/AVF and its 500 Hz → 1 kHz up-sampling, and it computes change from baseline the same way. Reconstructed placebo-corrected from the shipped files, it recovers its own finding: all four drugs prolong QTcF by +17 to +95 ms, while J-Tpeak separates them — +37 and +24 ms for the predominant-hERG blockers (dofetilide, quinidine) against +6 and −8 for the multichannel ones (ranolazine, verapamil).

Four things differ from its siblings. Triplicates are exact (all 1,744 groups hold 3, so 5,232 records are ~1,744 observations). The pharmacokinetic table is long, not wide — plasma_analyte names the one agent measured — and dofetilide is pg/mL while the other three are ng/mL, so use the derived plasma_concentration_ng_ml across arms; dose carries the same split (500 µg vs 120–1500 mg). Its median beats are variable length (968–1,876 samples, against ecgdmmld's fixed 1,200) and 9 are missing entirely, which is why 9 records have no PR/QRS/QT/J-Tpeak. And secondary T peaks are real here — 42 records populate tpeak_tpeakp_ms, where ecgdmmld's copy of that column is empty in every row. Two PR values are stored as a 32-bit arithmetic wrap and are repaired, flagged by pr_ms_repaired. See examples/load_ecgrdvq.py.

Both are read-time adapters: they shape the returned tensor only. Source files, fold CSVs and validation are untouched — a record excluded for a flat V6 stays excluded even if you never load V6.

ECGDataset parameters

Sample windows

window=(start, length) returns a fixed slice of each record, in samples:

first  = ECGDataset("ptbxl", split="train", data_path="...", window=(0, 2500))
second = ECGDataset("ptbxl", split="train", data_path="...", window=(2500, 2500))
first[0]["signal"].shape    # (12, 2500) -- samples 0-2499
second[0]["signal"].shape   # (12, 2500) -- samples 2500-4999

length=None reads to the end of the record. Prefer window= over a cropping transform for two reasons:

• It is pushed down into the reader, so only those samples are decoded. On long records that is a large difference — incartdb goes from ~106 ms to ~8 ms per record; on 10-second records it changes nothing.
• It is picklable. transform=lambda x: x[:, :2500] fails in a DataLoader(num_workers>0) under the spawn start method, the default on macOS and Windows. window= works under both fork and spawn.

A window that does not fit raises WindowOutOfRangeError, naming the record and its true length. Record length is not constant in every dataset — cpsc_2018 runs 6-144 s, ptbdb 32-120 s, staffiii 94.5-960 s, sph 10-56 s and ningbo_iva 2.9-59.3 s — so a fixed window can fit most records and not all.

window combines freely with fold_numbers, leads and units; it is applied first, then lead selection, then units, then transform.

Derived datasets (annotations for another dataset's records)

Some releases contain no recordings of their own — they annotate, or re-cut, someone else's. There are three: PTB-XL+ (3 feature tables, 2 statement tables, derived median beats and 283,326 fiducial-point files, all keyed by PTB-XL's ecg_id), MIMIC-IV-ECG-Ext-ICD (ICD-10-CM discharge diagnoses for all 800,035 MIMIC-IV-ECG studies, keyed by study_id) and Symile-MIMIC (a multimodal cohort pairing 11,610 of those same MIMIC-IV-ECG studies with a chest X-ray and 50 blood labs).

Those get no config and no splits, deliberately. Their records are the host dataset's, so generating folds would create a second ECGBench partition of the same recordings and let someone train on one and evaluate on the other. They are label providers instead: load the host on its own folds and join.

from ecgbench import ECGDataset
from ecgbench.labels.ptbxl_plus import load_ptbxl_plus

ds = ECGDataset("ptbxl", split="train", data_path="/data/ptb-xl/1.0.3/", labels=True)
plus = load_ptbxl_plus("/data/ptb-xl-plus/1.0.1/", features=("unig",))

joined = plus.reindex(ds.metadata_df["ecg_id"].values)   # 17,376 of 17,376
joined.iloc[0]["ptbxl_scp_codes"]     # [('NORM', 100.0), ('LVOLT', 100.0), ('SR', 100.0)]
joined.iloc[0]["12sl_statements"]     # ['NSR', 'NML']  -- the algorithm's opinion
joined.iloc[0]["unig_QRS_Dur_Global"] # 86.0 ms

You need both downloads, since PTB-XL+ has no waveforms. Feature columns are provider-prefixed because the three providers reuse names. See examples/load_ptbxl_plus.py, and the dataset page for the release's own defects — notably that 12sl_features.csv ships with no key column.

Ext-ICD works the same way, and adds one wrinkle worth knowing: it ships the upstream authors' own 20-fold split alongside the labels, which is independent of ECGBench's 10 folds. Reproduce published numbers on one or work on ECGBench's folds on the other, but never cross them.

from ecgbench.labels.mimic_iv_ecg_ext_icd import label_set, load_ext_icd, multi_hot

# prefix= because MIMIC-IV-ECG's own label frame also carries ecg_time.
icd = load_ext_icd("/data/mimic-iv-ecg-ext-icd-labels/1.0.1/", prefix="icd_")
codes = label_set(icd, prefix="icd_")          # 1076, the published label set
targets = multi_hot(icd.head(1000), codes, prefix="icd_")

Only 58.5% of its records carry a diagnosis at all, and the empty ones are empty lists rather than nulls — see examples/load_mimic_iv_ecg_ext_icd.py.

Symile-MIMIC is the same shape with one difference: it is a cohort, not a layer over the whole host. It covers 11,610 of MIMIC-IV-ECG's 800,035 studies, so a partial join is the correct result rather than a broken one.

from ecgbench.labels.symile_mimic import by_study_id, chexpert_targets, load_cohort

host = ECGDataset("mimic_iv_ecg", split="train", fold_numbers=[1],
                  data_path="/data/mimic-iv-ecg/1.0/", metadata_source="local")
cohort = load_cohort("/data/symile-mimic/1.0.0/", prefix="sym_")   # (11622, 92)
# Rows are admissions, so 12 ECG studies appear twice; the default policy keeps
# the earliest admittime, and on_duplicate="raise" refuses instead.
keyed = by_study_id(cohort, prefix="sym_")                        # (11610, 92)
joined = keyed.reindex(host.metadata_df["study_id"].values)       # 1,135 of 78,655
targets = chexpert_targets(joined, uncertain="nan", prefix="sym_")  # 14 CXR findings

Two traps of its own: the column literally named study_id is the CXR's, not the ECG's (the loader drops it), and the CheXpert labels have four states — −1.0 means uncertain and NaN means not mentioned, so chexpert_targets() makes you resolve both. The shipped data_npy ECG tensors are min-max normalised to [−1, 1] with the scale discarded, so they are not millivolts and cannot be converted back — read MIMIC-IV-ECG for those. See examples/load_symile_mimic.py.

Datasets with no waveforms at all

A dataset can also lack recordings without annotating anyone else's. The Eye Tracking Dataset for 12-Lead ECG Interpretation ships ten printed ECGs and the gaze behaviour of 63 clinicians reading them — 630 sessions, scored against 16–25 areas of interest per image. There is no sampled signal, no sampling rate, and no patient behind a record, so it too gets no config and no splits: the unit of observation is a reader session, and folds over "records" would be partitioning ten pictures. How to split a reader study — by reader or by image — depends on the task, so ECGBench ships tables and leaves that choice open.

from ecgbench.labels.eye_tracking_ecg import load_eye_tracking_ecg

df = load_eye_tracking_ecg("/data/eye-tracking-ecg/1.0.0/")

# Group by aoi_lead, not Label: labels are scoped per image ("V1 NSR" vs "V1 AFib"),
# and 1/2/3 are leads I/II/III rather than indices.
leads = df[df.aoi_kind == "lead"]
leads.groupby("Group")["Hit_time_G"].mean().round(0)   # Consultant 7266 ms, Med 1 11305 ms

Its -1 "never happened" codes and 0 ages are converted to NaN on load — being sentinels rather than blanks, they make every column look fully populated. See examples/load_eye_tracking_ecg.py and the dataset page.

Restricted and credentialed datasets

Most datasets' fold CSVs are published to the HuggingFace Hub and download automatically. Some are deliberately not, and those you generate yourself.

Fold CSVs carry identifiers only — record ID, patient ID, signal path, fold, split. For an openly licensed source that is uncontroversial. For a credentialed or restricted source those identifiers are still data derived under a use agreement, and the ECGBench Hub repository is public and ungated, so ECGBench does not publish them. Two datasets are in this category: mimic_iv_ecg, whose 800,035 study_ids and 161,352 subject_ids stay with the people who signed the PhysioNet DUA, and echonext, under the PhysioNet Restricted Health Data License whose clause 3 forbids sharing access to the data at all.

Such a dataset declares this in its config, and the tooling enforces it in both directions — ecgbench upload refuses to publish it, and ECGDataset raises SplitsNotPublishedError (carrying the command below) instead of a 404:

publish_fold_csvs: false
no_publish_reason: >
  MIMIC-IV-ECG is credentialed under the PhysioNet Credentialed Health Data
  Use Agreement, so ECGBench does not republish its identifiers ...

The split is distributed as a recipe instead. Because fold assignment is a deterministic function of the input table and a fixed seed, regenerating locally reproduces the canonical partition exactly:

# 1. Generate — writes output/<slug>/ plus a manifest.json
ecgbench splits --dataset mimic_iv_ecg --data-path /path/to/mimic-iv-ecg/1.0/

# 2. Verify it is the canonical partition, not merely a plausible one
python -c "from ecgbench import verify_splits; \
           print(verify_splits('mimic_iv_ecg', 'output/mimic_iv_ecg')['ok'])"

# 3. Point the loader at your generated folds
cp -r output/mimic_iv_ecg/{clean,original} /path/to/mimic-iv-ecg/1.0/

ds = ECGDataset("mimic_iv_ecg", split="train", metadata_source="local",
                data_path="/path/to/mimic-iv-ecg/1.0/", labels=True)

manifest.json is what makes "regenerate it yourself" trustworthy. ecgbench splits writes one for every dataset, recording the seed, fold count, grouping column, a SHA-256 of each input file, the record counts, and a fold digest — a hash over the entire record-to-fold mapping in canonical order. Two runs agree on that digest if and only if they produced the same partition. verify_splits() compares yours against a reference manifest shipped in the package and, on mismatch, names the input file that differs.

That last part is the common failure. A split only reproduces if the input is byte-identical, and local copies get filtered: we found a machine_measurements.csv cut to 789,481 of 800,035 rows, which silently changes the stratification and hence the folds. Verify your download against the provider's own checksums before generating.

Output format

Each sample is a dict:

• signal -- float32 tensor (leads, samples), in millivolts unless units="uV"
• record_id -- record identifier
• split, fold -- split name and fold number
• labels -- dict of the dataset's label and metadata fields (only with labels=True)
• All other CSV columns as tensors (numeric) or raw values (str/dict)

The dataset object also carries ds.lead_names and ds.units, so the tensor is self-describing.

Data Versions

• clean (default): only records that pass all quality checks
• original: all records with is_valid and quality_issues columns

Both versions share identical fold assignments. Use original when you need all records or want to filter manually; use clean for standard benchmarking.

Validation

ECGBench validates every signal file before splitting:

• missing_leads -- lead entirely NaN or all-zero
• nan_values -- any NaN in signal
• truncated_signal -- fewer samples than expected
• flat_line -- lead with near-zero variance
• corrupt_header -- unreadable signal file
• amplitude_outlier -- samples outside physiological range

Results are saved in validation_report.json with per-record details.

Croissant Metadata

Both clean/ and original/ versions include MLCommons Croissant 1.1 JSON-LD metadata (croissant.json) with SHA-256 hashes for reproducibility. The full pipeline generates both automatically. For standalone generation:

ecgbench croissant --dataset ptbxl --splits-dir output/ptbxl/clean/ --version clean
ecgbench croissant --dataset ptbxl --splits-dir output/ptbxl/original/ --version original

Adding a New Dataset

1. Copy ecgbench/data/configs/_template.yaml to <slug>.yaml, fill in fields
2. Run ecgbench splits --dataset <slug> --data-path /path/to/data/
3. Check validation_report.json -- review excluded records
4. If custom logic needed, create ecgbench/splitting/strategies/<slug>.py with @register("<slug>")
5. Run pytest
6. Upload: ecgbench upload --data-dir output/ --datasets <slug>

CLI

Installing ecgbench adds a single ecgbench console command with three subcommands:

ecgbench --help               # top-level help
ecgbench <command> --help     # per-subcommand flags
ecgbench --version            # package version

Every subcommand has an equivalent Python function (run_splits, run_croissant, run_upload) with the same arguments, so the same workflow can be driven from a notebook or downstream code.

ecgbench splits

Runs the full pipeline: validate -> split -> export -> Croissant. Writes output/<dataset>/{original,clean}/ by default.

ecgbench splits --dataset ptbxl --data-path /path/to/ptb-xl/1.0.3/
ecgbench splits --dataset ptbxl                        # auto-download
ecgbench splits --dataset chapman_shaoxing \
    --data-path /data/chapman/ \
    --output-dir /data/outputs/chapman/ \
    --n-folds 10 --max-workers 8

# PhysioNet ecg-arrhythmia (45,152 records, Chapman-Shaoxing + Ningbo).
# Ships no metadata CSV — the splitter builds ecgbench_metadata.csv from the
# per-record WFDB headers on first run, so the data directory must be writable.
ecgbench splits --dataset ecg_arrhythmia \
    --data-path /data/ecg-arrhythmia/1.0.0/ --max-workers 32

Python equivalent:

import ecgbench

result = ecgbench.run_splits(
    dataset="ptbxl",
    data_path="/path/to/ptb-xl/1.0.3/",
    output_dir=None,          # -> output/ptbxl/
    sampling_rate=None,       # -> config default_sampling_rate
    n_folds=10,
    max_workers=4,
    skip_validation=False,
    skip_croissant=False,
)
# result is a dict with: dataset, dataset_name, output_dir,
# original={total,train,val,test}, clean={total,train,val,test}, excluded

ecgbench croissant

Standalone Croissant 1.1 JSON-LD generator for an existing splits directory. Run once per version (clean and original).

ecgbench croissant --dataset ptbxl --splits-dir output/ptbxl/clean/    --version clean
ecgbench croissant --dataset ptbxl --splits-dir output/ptbxl/original/ --version original
ecgbench croissant --dataset ptbxl --splits-dir output/ptbxl/clean/ --validate

Python equivalent:

from pathlib import Path
import ecgbench

saved_path: Path = ecgbench.run_croissant(
    dataset="ptbxl",
    splits_dir="output/ptbxl/clean/",
    output=None,              # -> splits_dir/croissant.json
    version="clean",
    validate=True,            # raises RuntimeError if the file does not validate
)

Requires the croissant extra (pip install ecgbench[croissant]).

ecgbench upload

Uploads each dataset's original/ and clean/ CSV folds, plus validation_report.json and croissant.json if present, to a HuggingFace Hub dataset repository. One or more dataset slugs can be uploaded in a single call.

ecgbench upload --data-dir output/ --datasets ptbxl
ecgbench upload --data-dir output/ --datasets ptbxl chapman_shaoxing
ecgbench upload --data-dir output/ --datasets ptbxl --dry-run
ecgbench upload --data-dir output/ --datasets ptbxl \
    --hf-repo-id your-org/ECGBench

Authentication resolves in this order: token= argument (Python API only) -> HF_TOKEN env var -> HUGGINGFACE_HUB_TOKEN env var -> .env file in the current working directory. Run with --dry-run first to review the file list.

Python equivalent:

import ecgbench

counts: dict[str, int] = ecgbench.run_upload(
    data_dir="output/",
    datasets=["ptbxl", "chapman_shaoxing"],
    hf_repo_id="vlbthambawita/ECGBench",
    dry_run=False,
    token=None,               # falls back to env / .env
)
# counts: {"ptbxl": 42, "chapman_shaoxing": 42}

Requires the hf extra (pip install ecgbench[hf]).

API Reference

Config

• load_config(slug) -- load DatasetConfig from YAML
• list_available_configs() -- list dataset slugs with configs

Catalogue

• list_datasets() -- all 64 datasets as CatalogueEntry objects
• search(query, category, access) -- filter datasets
• get_dataset(name) -- look up by name
• categories() -- unique categories
• to_dataframe() -- as pandas DataFrame

Dataset

• ECGDataset(dataset, split, ...) -- unified PyTorch Dataset
• ecg_collate_fn(batch) -- custom collate for DataLoader
• WindowOutOfRangeError -- raised when a window= does not fit a record

Validation

• validate_dataset(data_path, config) -- run quality checks
• generate_report(result, config) -- generate report dict
• save_report(result, config, path) -- save report JSON

Splitting

• split_dataset(df, labels, config) -- generate folds
• export_splits(split_result, val_result, output_dir, config) -- write CSVs
• get_splitter(slug) -- get dataset-specific splitter

Croissant

• generate_croissant(config, splits_dir) -- generate JSON-LD
• save_croissant(config, splits_dir) -- save to file
• validate_croissant(path) -- validate JSON-LD

Download

• download_dataset(config) -- download from source
• resolve_data_path(path, config) -- resolve or download

Pipelines (CLI + Python API)

• run_splits(dataset, ...) -- full validate + split + Croissant pipeline (same as ecgbench splits)
• run_croissant(dataset, splits_dir, ...) -- standalone Croissant generation (same as ecgbench croissant)
• run_upload(data_dir, datasets, ...) -- HuggingFace Hub upload (same as ecgbench upload)

Development

uv pip install -e ".[dev]"
ruff check ecgbench/
black ecgbench/
pytest

Citation

If you use ECGBench in your research, please cite:

@software{ecgbench,
  author = {Thambawita, Vajira},
  title = {ECGBench: Reproducible ECG Benchmark Datasets},
  url = {https://github.com/vlbthambawita/ECGBench}
}

License

MIT License -- see LICENSE for details.