Package Dependencies

• allennlp == 0.8.4
• gensim == 3.8.3
• pytorch==1.7.1
• spacy == 2.1.9
• nltk == 3.6.1
• python == 3.8.8
• pytorch-pretrained-bert == 0.6.2
• transformers == 4.9.2
• scikit-learn == 0.24.1
• dgl == 0.6.1
• torch-scatter == 2.0.6
• torch-sparse == 0.6.9
• torch-geometric == 1.7.0

You can use the following command (recommended):

pip install -r requirements.txt

Preprossing

Clinical Document

The data can be downloaded from MIMIC-III. The structure of data files can be shown like:

data
|   D_ICD_DIAGNOSES.csv
|   D_ICD_PROCEDURES.csv
|   ICD9_descriptions.txt
└───mimic3/
|   |   NOTEEVENTS.csv
|   |   DIAGNOSES_ICD.csv
|   |   PROCEDURES_ICD.csv
|   |   *_hadm_ids.csv (get from CAML)

The hierarchical ICD graph can be obtained by running python icd_graph.py.

The corresponding ICD code file can be obtained by running python preprocess_mimic3.py.

To obtain the clause interaction graph, firstly run the python preprocess_mimic3_raw.py to obtain raw discharge summaries. And then run the python clause.py to get clause clustering results. Finally, run python code_matching.py to obtain interaction connection.

Training

MIMIC-III-50

python main.py --MAX_LENGTH 2500 --n_epochs 200 --batch_size 4 --model GRU --lr 1e-4 --clause True --gpu 1 --bidirectional --criterion prec_at_5

MIMIC-III-Full

python main.py --MAX_LENGTH 2500 --n_epochs 200 --batch_size 4 --model GRU --lr 1e-4 --clause True --gpu 1 --bidirectional --criterion prec_at_8 --Y full --data_path ./data/mimic3/train_full.csv

Testing

MIMIC-III-50

python main.py --MAX_LENGTH 2500 --n_epochs 200 --batch_size 4 --model GRU --lr 1e-4 --clause True --gpu 1 --bidirectional --test_model=save_model_path

MIMIC-III-Full

python main.py --MAX_LENGTH 2500 --n_epochs 200 --batch_size 4 --model GRU --lr 1e-4 --clause True --gpu 1 --bidirectional --Y full --data_path ./data/mimic3/train_full.csv --test_model=save_model_path