ECG Multi-Label Classfication

This repo is to archive code of a competition I attented.
The competition is about multi-label classfication of 8-leads ECGs.
This repo does not provide a top solution, but for backup and sharing.

1. Dataset

Page: "合肥高新杯"心电人机智能大赛”

Round	Classes	Train*	TestA*	TestB*
Round 1	55	24106	8036	5435
Round 2	34	20036	9918	12622

* Number of samples

Round 2 TestA and Round 2 TestB are not available for downloading.
Round 1 TestB does not have labels.
Access to Round 1 Train, Round 1 TestA and Round2 Train should be required.

2. Code and Command

Round	Path	Classes	TestA*	TestB*	Rank
Round 1	[Link]	55	0.8201	0.8424	28/2353
Round 2	[Link]	34	0.9251	0.9225	35/2353

* F1-score

3. Brief Intro

• Preprocessing:
  • merge Round 1 Train and Round 1 TestA, remove duplicates;
  • compute additional 4 leads (III, aVR, aVL, aVF) on basis of given 8 leads;
  • standard normalization;
  • (only in round 2) compute weights for samples in Round 1 data, merged with Round 2 Train.
• Model Structure:
  • 1D DenseNet with slight modification.
• Train:
  • 5-fold split by scikit-multilearn;
  • 5-fold cross validation;
  • augmentation: slightly adjust amplitude and baseline of ECG;
  • compound loss function: F1Loss, FocalLoss and MultiLabelSoftMarginLoss;
  • search best threshold on validation set after each fold training.
• Prediction:
  • ensemble 5-fold predictions by votting.

4. Dependencies

Package	Version	Comment
conda	4.5.11
python	3.7.3
tqdm	4.32.1
numpy	1.16.4	conda install
scipy	1.3.0
pandas	0.24.2
pytorch	1.1.0	conda install with cuda 9.0
scikit-learn	0.21.2
scikit-multilearn	0.2.0	pip install scikit-multilearn