Explainable Prediction of Adverse Outcomes Using Clinical Notes
Code for the paper Explainable Prediction of Adverse Outcomes Using Clinical Notes.
Modify all of the file paths in resources/params.json according to your local file structure.
Processing Data
Preprocess the notes and extract the target labels.
python preprocess/extract_labels.py
Split the data into 5 splits to use for 5-fold cross validation.
python preprocess/cross_val_splits.py
Training Embeddings
In this work we use two different embedding techniques, Word2Vec and Starspace embeddings. Running word_embedding.py generates the Word2Vec embeddings and the files needed to generate the Starspace embeddings. Instructions for installing and training Starspace embeddings can be found in their repo. An example script with the settings we used to train the starspace embeddings can be found at preprocess/train_starspace.py.
python preprocess/word_embedding.py
Training Models
The training hyperparameters can be modified in resources/params.json. The models can be trained using the following command.
python train.py
Acknowledgments
We adapt our preprocessing code from that released by DeepEHR.