ecgdig / data /README.md
RyanIRL's picture
ecgdig: data
a6bc474 verified
|
Raw History Blame Contribute Delete
2.59 kB
# ecgdig dataset
Recover 12-lead ECG voltage time series from images of printed ECG pages.
## Layout
```
train.csv one row per training record: id, fs, sig_len
train/<id>/<id>.csv ground-truth signal table for that record
train/<id>/<id>-v*.{png,jpg} nine images of the same printed record
test.csv one row per test record and lead:
id, lead, fs, number_of_rows
test/<id>.jpg ONE image per test record (condition varies)
sample_submission.csv valid all-zeros submission (the format contract)
```
## Records
A record is one 12-lead ECG acquisition, 10 seconds long, sampled at `fs`
Hz (`fs` is one of 250, 256, 500, 512, 1000, 1025; `sig_len = fs * 10`).
Each printed page uses the standard 3x4 layout plus a lead II rhythm
strip at 25 mm/s and 10 mm/mV: lead II is visible for the full 10 s;
every other lead is visible for one 2.5 s column (I, II, III during
0 to 2.5 s, aVR, aVL, aVF during 2.5 to 5 s, V1, V2, V3 during 5 to
7.5 s, V4, V5, V6 during 7.5 to 10 s).
## Signal tables (train/<id>/<id>.csv)
`sig_len` rows, twelve columns I, II, III, aVR, aVL, aVF, V1..V6, values
in millivolts. Each lead holds numbers only inside its visible window and
NaN outside it; lead II is fully populated. The visible windows are what a
submission must reconstruct for test records.
## Image conditions
Every training record carries nine images of the same printed page:
- `v1` clean digital render of the page (PNG)
- `v2` color print, scanned in color (JPEG)
- `v3` color print, scanned in black and white (PNG)
- `v4` phone photo of the printed page (JPEG)
- `v5` phone photo of the page shown on a laptop screen (JPEG)
- `v6` phone photo of a stained and soaked printout (JPEG)
- `v7` phone photo of a heavily damaged printout (JPEG)
- `v8` color scan of a mold-degraded printout (JPEG)
- `v9` black and white scan of a mold-degraded printout (JPEG)
Each test record provides exactly ONE image (JPEG), drawn from these same
conditions; the condition is not labeled. The pages carry printed header
furniture (an id string, age, sex, a date, sometimes a QR code); it is
page decoration, and the rules forbid using it to try to identify source
recordings.
## Submission format
CSV with header `id,value`; `id = {record}_{row}_{lead}` covering, for
every test record, `fs * 10` rows for lead II and `floor(fs * 2.5)` rows
for each other lead (`number_of_rows` in test.csv is authoritative).
`value` is the predicted voltage in mV at that sample. Every expected id
exactly once; finite values only.