GAMENet : Graph Augmented MEmory Networks for Recommending Medication Combination

For reproduction of medication prediction results in our paper, see instructions below.

This repository contains code necessary to run GAMENet model. GAMENet is an end-to-end model mainly based on graph convolutional networks (GCN) and memory augmented nerual networks (MANN). Paitent history information and drug-drug interactions knowledge are utilized to provide safe and personalized recommendation of medication combination. GAMENet is tested on real-world clinical dataset MIMIC-III and outperformed several state-of-the-art deep learning methods in heathcare area in all effectiveness measures and also achieved higher DDI rate reduction from existing EHR data.

• Pytorch >=0.4
• Python >=3.5

In ./data, you can find the well-preprocessed data in pickle form. Also, it's easy to re-generate the data as follows:

1. download MIMIC data and put DIAGNOSES_ICD.csv, PRESCRIPTIONS.csv, PROCEDURES_ICD.csv in ./data/
2. download DDI data and put it in ./data/
3. run code ./data/EDA.ipynb

Data information in ./data:

• records_final.pkl is the input data with four dimension (patient_idx, visit_idx, medical modal, medical id) where medical model equals 3 made of diagnosis, procedure and drug.
• voc_final.pkl is the vocabulary list to transform medical word to corresponding idx.
• ddi_A_final.pkl and ehr_adj_final.pkl are drug-drug adjacency matrix constructed from EHR and DDI dataset.
• drug-atc.csv, ndc2atc_level4.csv, ndc2rxnorm_mapping.txt are mapping files for drug code transformation.

Traning codes can be found in ./code/baseline/

• Nearest will simply recommend the same combination medications at previous visit for current visit.
• Logistic Regression (LR) is a logistic regression with L2 regularization. Here we represent the input data by sum of one-hot vector. Binary relevance technique is used to handle multi-label output.
• Leap is an instance-based medication combination recommendation method.
• RETAIN can provide sequential prediction of medication combination based on a two-level neural attention model that detects influential past visits and significant clinical variables within those visits.
• DMNC is a recent work of medication combination prediction via memory augmented neural network based on differentiable neural computers.

python train_GAMENet.py --model_name GAMENet --ddi# training with DDI knowledge
python train_GAMENet.py --model_name GAMENet --ddi --resume_path Epoch_{}_JA_{}_DDI_{}.model --eval # testing with DDI knowledge
python train_GAMENet.py --model_name GAMENet # training without DDI knowledge
python train_GAMENet.py --model_name GAMENet --resume_path Epoch_{}_JA_{}_DDI_{}.model --eval # testing with DDI knowledge

Please cite our paper if you use this code in your own work:

@article{shang2018gamenet,
  title="{GAMENet: Graph Augmented MEmory Networks for Recommending Medication Combination}",
  author={Shang, Junyuan and Xiao, Cao and Ma, Tengfei and Li, Hongyan and Sun, Jimeng},
  journal={arXiv preprint arXiv:1809.01852},
  year={2018}
}