Transfer learning

Transfer learning model from lima_deep_2021 and CODE-15% to ptb_xl.

Functions

There are several functionalities, including 1) converting wfdb files to hdf5, 2) making predictions of age based on predefined model, 3) training model based on ECG data. To view different functions use:

1. Convert .dat to hdf5

PTB-XL data is stored as .dat files, and can not natively be incorperated into the Lima. et. al. model. One must first convert the files to HDF5 using the convert function. Basic syntax:

# Function structure
python main.py convert --data DATA_PATH --metadata METADATA_FILE --save_hdf SAVE_HDF_PATH --save_csv SAVE_CSV_PATH

# View help
python main.py convert --help

Example: Convert .dat data in the 'data-raw/ptb-xl/data/' and store as 'data/ptbxl.hdf5'.

python main.py convert --data "data-raw/ptb-xl/data/" --metadata "data-raw/ptb-xl/metadata.csv" --save_hdf "data/ptbxl.hdf5" --recursive True --replace True --save_csv "data/metadata.csv"

2. Train model

Trains model based on predefined model arcitechture. Can be created by fine tuning or from scratch. Basic synstax:

# Function structure
python main.py train --data DATA_PATH --metadata METADATA_PATH --model MODEL_PATH

# View help
python main.py train --help

Example: Train model using fine tuning based on model stored in 'model/' folder.

python main.py train --data "data/ptbxl.hdf5" --metadata "data/metadata.csv" --model "model_ptbxl" --replace True --validation_percentage 0.2 --batch_size 6 --epochs 2 --tune True --tune_model "model/"

3. Predict age based on model

Create predictions of age based on predefined model. Basic syntax:

# Function structure
python main.py predict --data DATA_PATH --save PREDICTION_SAVE_PATH

# View help
python main.py predict --help

Example: Predict ages in 'data/ptbxl.hdf' and store predictions as 'data/predictions.csv'.

python main.py predict --data "data/ptbxl.hdf5" --save "data/predictions.csv"

Data

PTB-xl

Data format

• Data is formated as .dat binary files.
• Contains both 500hz and 200hz ECG measurments.
• Data is split in a bunch of ways in relation to quality of ECG (Folds)
  • Fold 9 + 10 underwent at least one human evaluation (High label quality)
  • Keep in mind when getting to cross-validation

Additional data

• ECG statements: A bunch of SCP-ECG statements
• Signal metadata: Stats related to:
  • Static noise
  • Burst noise
  • Baseline drift
  • Electrodes problems
  • Extra beats

Download raw data Raw data can be downloaded by bash command: wget -r -N -c -np https://physionet.org/files/ptb-xl/1.0.1/

Citations

Model and study citation:

Lima, E.M., Ribeiro, A.H., Paixão, G.M.M. et al. Deep neural network-estimated electrocardiographic age as a mortality predictor. Nat Commun 12, 5117 (2021). https://doi.org/10.1038/s41467-021-25351-7. Read more at code base https://github.com/antonior92/ecg-age-prediction

Training data citation:

Wagner, P., Strodthoff, N., Bousseljot, R., Samek, W., & Schaeffter, T. (2020). PTB-XL, a large publicly available electrocardiography dataset (version 1.0.1). PhysioNet. https://doi.org/10.13026/x4td-x982. Read more at https://physionet.org/content/ptb-xl/1.0.1/