Skip to content

Latest commit

 

History

History
43 lines (34 loc) · 1.65 KB

data_notes_root.md

File metadata and controls

43 lines (34 loc) · 1.65 KB

data

0. ECGTransform pre-processed data, .pt format

PyTorch library torch.save(x.pt), torch.load(x.pt)
MIT - train.pt, test.pt, val.pt 79mb
PTB - train.pt, test.pt, val.pt 10mb

MIT data 5 classes, PTB data only 2 classes (N normal, M mal)
Most training was done on MIT data, then generalized to PTB data,
but PTB data is of much higher quality, 15 channels vs 2 channels, and measuring device is medical grade vs ambulatory (noisy).
Also 2 doctors "corrected" many unreadable data for AI Computer training.

1. MIT BIH

48 30-minute recordings from 2-channel ECG ambulatory 24-hour monitor (noisy), from 47 unique people.
Recorded at BIH Lab (Beth Isreal Hospital), Boston 1975-79 period. Old tech, very old.
Digitized to 360 samples per second, 11 bit resolution, 2 physicians annotated each sample, clear up unreadable sections for computer to read.

2. PTB

594 high resolution 15 lead ECG from 294 subjects, digitized to 1000 samples per second, recording dates 1990 - 1997, in Germany. Much better, newer tech, high precision measuring device.
Used in 1995 paper, Germany.

Original data downloaded, binary format except header file in text format.

Physionet/CinC Challenge 2017

MatLab all formats, data, code. .m format
Single channel data.

About this file Each row is a different ECG. The last two columns contain the identifier of the original file and the label

This is the code for labels: labels: normal = 0 AF = 1 other = 2 noisy = 3