JayantSharma08 / RAGAT

A customized version of the Relational Aware Graph Attention Network for large scale EHR records.

Electronic Health Record graph completion using RAGAT: Relation Aware Graph Attention Network.

Data

The encoded data is uploaded in the ehr folder in the form of subject,relation,object triples:

• Diagnosis(PheWAS code), associated_with, Diagnosis(PheWAS code)
• Diagnosis(PheWAS code), treated_with, Drug(ATC code)
• Procedure(CPT code), performed_for, Diagnosis(PheWAS code)
• Drug(ATC code), causes, Side Effect(UMLS code)

Although the original data source from which these triples have been extracted is private, the links captured happened in real life patient encounters.

Dependencies

• Pytorch 1.5

Datasets

• FB15k-237
• WN18RR

Training model

# FB15k-237
python run.py -epoch 1500 -name test_fb -model ragat -score_func
interacte -opn cross -gpu 0 -data FB15k-237 -gcn_drop 0.4 -ifeat_drop 0.4 
-ihid_drop 0.3 -batch 1024 -iker_sz 9 -attention True -head_num 2
# WN18RR
python run.py -epoch 1500 -name test_wn -model ragat -score_func
interacte -opn cross -gpu 0 -data WN18RR -gcn_drop 0.4 -ifeat_drop 0.2 
-ihid_drop 0.3 -batch 256 -iker_sz 11 -iperm 4 -attention True -head_num 1
# EHR
#pretrained
python run.py -epoch 1 -name ehr_ragat_17_03_2022_12:29:23 -model ragat -score_func interacte -opn cross -gpu 0 -gcn_drop 0.4 -ifeat_drop 0.2 -ihid_drop 0.3 -batch 256 -iker_sz 11 -iperm 4 -attention True -head_num 1 -restore
python run.py -epoch 1 -name extended_ehr_29_06_2022_07:15:16 -model ragat -score_func interacte -opn cross -gpu 0 -gcn_drop 0.2 -ifeat_drop 0.2 -ihid_drop 0.3 -batch 128 -iker_sz 11 -iperm 4 -attention True -head_num 1 -lbl_smooth 0.125 -lr 0.0005 -restore
#Training with Hyperparamter Optimization using Optuna
python run.py -epoch 800 -model ragat -score_func interacte -opn cross -gpu 0 -gcn_drop 0.4 -ifeat_drop 0.2 -ihid_drop 0.3 -batch 256 -iker_sz 11 -iperm 4 -attention True -head_num 1
python run.py -epoch 5 -model ragat -score_func interacte -opn cross -gpu 0 -gcn_drop 0.4 -ifeat_drop 0.2 -ihid_drop 0.3 -batch 256 -iker_sz 11 -iperm 4 -attention True -head_num 1
python run.py -epoch 800 -model ragat -score_func interacte -opn cross -gpu 0 -gcn_drop 0.2 -ifeat_drop 0.2 -ihid_drop 0.3 -batch 128 -iker_sz 11 -iperm 4 -attention True -head_num 1 -lbl_smooth 0.125 -lr 0.0005
python run.py -epoch 1000 -name extended_ehr -model ragat -score_func interacte -opn cross -gpu 0 -gcn_drop 0.2 -ifeat_drop 0.2 -ihid_drop 0.3 -batch 128 -iker_sz 11 -iperm 4 -attention True -head_num 1 -lbl_smooth 0.125 -lr 0.0005

Acknowledgement

This code has been forked from the original RAGAT repo.RAGAT The project is built upon COMPGCN