dayveedd/ecg-af-250hz

ECG-AF processed ECG collection

What this dataset is

This is a derived, harmonized release produced by the ECG-AF framework. The ECG-AF project did not collect new patient recordings. It downloaded the public source datasets listed below, converted their heterogeneous signal and annotation formats into one representation, and exported the framework's processed layer to Hugging Face.

This release contains 1,567 processed ECG recordings at 250 Hz with their available annotations, recording-level metadata, source provenance, and integrity checksums. The exact framework commit is recorded in release.json; this card links the same source revision above.

Source subset	Recordings	Processed payload	Source license
AF Termination Challenge Database	80	0.03 GB	ODC-By 1.0
Arrhythmias in Children and Patients with Congenital Heart Disease	39	2.71 GB	ODC-By 1.0
CPSC 2021 Paroxysmal AF Event Detection Dataset	1,425	9.18 GB	CC BY 4.0
MIT-BIH Atrial Fibrillation Database	23	1.47 GB	ODC-By 1.0
Total	1,567	13.39 GB	Mixed; see each source

Credit for the original recordings and annotations belongs to their respective creators. Each subset retains its source license; the license list in the card metadata is not a choice of license for every file.

How ECG-AF derived this release

Public source archive
  -> ECG-AF provider downloader and format reader
  -> source-specific annotation adapter
  -> signal resampling to 250 Hz (no amplitude normalization)
  -> one signal Parquet + optional annotation JSON per recording
  -> recording index, provenance manifest, and SHA-256 checksums

The framework preserves source recording/subject identifiers and records the original sampling rate, channel names, units, source format, and annotation availability. It does not add model predictions or impose benchmark labels or train/validation/test splits.

Data layout and use

The Dataset Viewer and datasets.load_dataset expose one metadata row per recording from each subset's JSONL index:

from datasets import load_dataset

records = load_dataset(
    "dayveedd/ecg-af-250hz",
    "cpsc-2021-paroxysmal-af-event-detection-dataset",
    split="recordings",
)
print(records[0]["parquet_path"])

parquet_path and annotations_path are paths relative to this repository, not external S3 addresses. Download the actual waveforms and annotations with huggingface_hub.snapshot_download:

import json
from pathlib import Path
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download

root = Path(snapshot_download(repo_id="dayveedd/ecg-af-250hz", repo_type="dataset"))
index = next((root / "metadata").glob("*.jsonl"))
recording = json.loads(index.read_text().splitlines()[0])
signal = pq.read_table(root / recording["parquet_path"])
annotations = (
    json.loads((root / recording["annotations_path"]).read_text())
    if recording.get("annotations_path") else None
)

Signal Parquet files are preserved byte for byte: one sample per row, a sample_index column, and one channel_<index>_<lead> column per lead. Footer metadata records sampling rates, channels, units, source format, and original filename. JSON sidecars contain beat positions/labels, rhythm episodes, delineation, and dataset-specific labels when available. Annotation sample positions refer to the processed sampling rate. Signals are resampled by the pipeline; its current default does not normalize amplitudes. Consult each recording's units and source documentation rather than assuming uniform units.

metadata/ retains recording and subject identifiers, sampling rates, channels, durations, annotation availability, and source licenses from processing reports. provenance/ contains source manifests and processing counts. release.json records the code revision and exported counts. SHA256SUMS covers release files. Raw archives, model outputs, database dumps, and credentials are not included. Storage filenames beginning x00 receive a recording- prefix for Hub path compatibility; recording identifiers and file contents are preserved.

Labels, splitting, and limitations

No training/validation/test split is imposed. The recordings split is an inventory, not a training recommendation. Group by dataset and subject when subject IDs exist, and by recording otherwise. Keep slices of a recording together and check source-specific identity conventions before evaluating models.

AF-vs-NSR window labels are a downstream benchmark policy, not the native class of every source. Missing annotations are unknown; do not interpret them as NSR. The AF Termination Challenge subset consists of AF examples and does not supply a normal control group. Populations, lead configurations, units, and annotation quality vary across sources. The AF Termination subset's full-record AF rhythm spans are generated by the adapter from challenge membership; its beat annotations come from the original WFDB QRS files. Consult annotation sidecar provenance fields when distinguishing derived and original labels. Inspect source documentation and processing summaries when interpreting cross-dataset results.

This release preserves the processed source content and identifiers; it does not perform an additional de-identification pass. It is a research dataset and has not been clinically validated by this project.

Sources, attribution, and licenses

AF Termination Challenge Database

• Source: https://physionet.org/content/aftdb/1.0.0/ (version 1.0.0)
• License: ODC-By 1.0
• Exported recordings: 80
• Citation: Moody GB. Spontaneous Termination of Atrial Fibrillation: A Challenge from PhysioNet and Computers in Cardiology 2004. Computers in Cardiology. 2004;31:101-104. Dataset: PhysioNet AFTDB v1.0.0. https://physionet.org/content/aftdb/1.0.0/

Arrhythmias in Children and Patients with Congenital Heart Disease

• Source: https://physionet.org/content/leipzig-heart-center-ecg/1.0.0/ (version 1.0.0)
• License: ODC-By 1.0
• Exported recordings: 39
• Citation: Klehs S, Franke D, Alhamad B, Gebauer R, Teich L, Teich T, Paech C. Leipzig Heart Center ECG-Database: Arrhythmias in Children and Patients with Congenital Heart Disease. PhysioNet (v1.0.0). 2025. https://doi.org/10.13026/7a4j-vn37

CPSC 2021 Paroxysmal AF Event Detection Dataset

• Source: https://physionet.org/content/cpsc2021/1.0.0/ (version 1.0.0)
• License: CC BY 4.0
• Exported recordings: 1425
• Citation: Wang X, Ma C, Zhang X, Gao H, Clifford GD, Liu C. Paroxysmal Atrial Fibrillation Events Detection from Dynamic ECG Recordings: The 4th China Physiological Signal Challenge 2021. PhysioNet (v1.0.0). 2021. https://doi.org/10.13026/ksya-qw89

MIT-BIH Atrial Fibrillation Database

• Source: https://physionet.org/content/afdb/1.0.0/ (version 1.0.0)
• License: ODC-By 1.0
• Exported recordings: 23
• Citation: Moody GB, Mark RG. A new method for detecting atrial fibrillation using R-R intervals. Computers in Cardiology. 1983;10:227-230. Dataset: PhysioNet AFDB v1.0.0 (https://physionet.org/content/afdb/1.0.0/).