This repository is the result of a work on ECG Classification, with a large focus on visualization and representation learning, carried out by me and Andrej Ivanov.
We used 2 databases, MIT-BIH and PTBDB. The first contains much over 100,000 samples and 5 different classes while the latter contains just under 15,000 samples divided into 2 classes. Both are ECG readings that have been pre-processed.
This repositiory contains the following approaches, which are easily accessible through the Index Notebook:
- Baseline CNN
- Residual CNN
- Deep residual CNN
- RNN
- Bidirectional RNN
- Transfer Learning from the larger dataset to the smaller one
- Ensemble method
The results from our comparison of the approaches:
Here are the initial results from our visualizations. These are pca and t-sne projections of the representations achieved by the baseline CNN.
PTBDB dataset
MITBIH dataset
And here are comparisons from the representations achieved by each deep model:
