Atrial Fibrillation detection with a deep probabilistic model. Backend for Diagnose Report app. Use PhysioNet dataset for model training and testing.

The following dependencies are required.

• Python
• Tensorflow
• Keras
• Cardio Framework

I have already ran training for you. You can use the saved model in dirichlet_model folder to predict right away. But make sure to change the path in direchlet_model/checkpoint according to your path.

If you want to train again by yourself, run this following notebook file: dirichlet_model_training.ipynb, download the dataset and start training. Notice that on 1000 epochs, the training will take some time. Mine took about 8-9 hours on Tesla K80.

For the project I'm working on, I create some shell file and python files to convert and predict stuffs.

• To predict whether an image is AF (Atrial Fibrillation) or not:

predict <image file path>

It will return something like this

[{'target_pred': {'A': 0.021675685, 'NO': 0.9783243},
  'uncertainty': 0.0073926448822021484}]

Which A is – Atrial fibrillation N – Normal rhythm, O – Other rhythm, so NO is no problem.

See more in test.ipynb for more test case and example.

• To generate image from csv:

gnuplot -e "fileIn='csv/04015.csv'; fileOut='uploads/04015.png'" csv2img.gnuplot

• To convert single file to csv and image:

./raw2img <filename without extension>

• To convert image to csv:

python img2csv.py '<full path to file>'

• To convert MAT to csv:

python mat2csv.py "raw/A00001.mat"

• To convert csv to signal, with Gain equal 1000, Frequency 300Hz. Notice that you can specify -f mean from which line (remove if from beginning of file) and also -t mean to which line.

wrsamp -i raw/A00001.csv -o raw/A00001-converted -G 1000 -F 300 -z

• To convert signal back to MAT:

wfdb2mat -r raw/A00001-converted

This project is licensed under the terms of the MIT license.