| Format | 12-lead · 10 s · 500 Hz · CSV (µV) |
|---|---|
| Patients | 10,646 |
| Records | 10,646 |
| Leads | 12 |
| License | CC0 1.0 |
| Origin | Chapman University; Shaoxing People's Hospital — China / USA |
The original figshare release of the Chapman-Shaoxing database: 10,646
twelve-lead ECGs, one per patient, each 10 seconds at 500 Hz. Signals are
CSV rather than WFDB — one file per record, 5,000 rows × 12 lead columns
with a header, in the standard lead order — and are stored in
microvolts, which ECGBench converts to millivolts via
signal_unit_scale: 0.001.
Labels are unusually rich for an open ECG dataset: a single-label Rhythm
class (11 values), a space-separated multi-label Beat annotation (57
codes), age and sex, and eleven automated measurements — ventricular and
atrial rate, QRS duration, QT and QTc, R and T axes, QRS count, and the
QRS/T fiducial offsets. A denoised copy of every signal ships alongside as
ECGDataDenoised/.
The collection is six separate files and its metadata is .xlsx, so
acquisition needs examples/download_chapman_figshare.py rather than
ECGBench’s single-archive auto-download.
| Code | Rhythm | Records |
|---|---|---|
| SB | Sinus bradycardia | 3,889 |
| SR | Sinus rhythm | 1,826 |
| AFIB | Atrial fibrillation | 1,780 |
| ST | Sinus tachycardia | 1,568 |
| SVT | Supraventricular tachycardia | 587 |
| AF | Atrial flutter | 445 |
| SA | Sinus irregularity | 399 |
| AT | Atrial tachycardia | 121 |
| AVNRT | AV node reentrant tachycardia | 16 |
| AVRT | AV reentrant tachycardia | 8 |
| SAAWR | Sinus atrium to atrial wandering rhythm | 7 |
These are recomputed from the shipped Diagnostics.xlsx, and they match the
paper’s figures — 10,646 records over 11 rhythm classes. Rhythm is
single-label, so the counts sum exactly to the record total.
Beat is a different quantity: space-separated and multi-label, averaging
0.84 codes per record with 5,419 records carrying NONE. Do not add the two
tables together.
Note the class imbalance: the rarest rhythm (SAAWR) has 7 records, fewer
than the 10 folds, so StratifiedKFold warns and those records cannot appear
in every fold. That is a property of the data, not a splitting fault.
| Version | Records | Note |
|---|---|---|
| original | 10,646 | all records, with is_valid + quality_issues |
| clean | 10,607 | 99.6% pass rate |
| excluded | 39 | 38 amplitude outliers, 49 flat-lead issues (overlapping) |
# Six files, 2.8 GB. md5-verified, resumable, converts the .xlsx metadata.
python examples/download_chapman_figshare.py --dest /path/to/chapman-figshare/
ecgbench splits --dataset chapman_shaoxing --data-path /path/to/chapman-figshare/from ecgbench import ECGDataset
ds = ECGDataset(
"chapman_shaoxing",
split="train",
version="clean",
data_path="/path/to/chapman-figshare/",
labels=True,
)
sample = ds[0]
sample["signal"].shape # (12, 5000), float32, millivolts
sample["labels"]["Rhythm"] # 'AFIB'
sample["labels"]["Beat"] # 'LVHV' (space-separated, multi-label)
sample["labels"]["QTCorrected"] # 454
# Signals are stored in microvolts; signal_unit_scale converts them to mV,
# so the tensor above is already in mV. Ask for microvolts back with
# units="uV", and select leads by name:
two = ECGDataset("chapman_shaoxing", split="train",
data_path="/path/to/chapman-figshare/",
leads=["II", "V5"], units="uV")
two[0]["signal"].shape # (2, 5000)
two.units # 'uV'