Leave No Patient Behind: Enhancing Medication Recommendation for Rare Disease Patients

This repository provides the official PyTorch implementation and reproduction for our SIGIR'24 paper titled "Leave No Patient Behind: Enhancing Medication Recommendation for Rare Disease Patients".

More descriptions are available via the paper.

If this work helps you, please kindly cite our papers:

@article{zhao2024leave,
  title={Leave No Patient Behind: Enhancing Medication Recommendation for Rare Disease Patients},
  author={Zhao, Zihao and Jing, Yi and Feng, Fuli and Wu, Jiancan and Gao, Chongming and He, Xiangnan},
  journal={arXiv preprint arXiv:2403.17745},
  year={2024}
}

Installation

1. Clone this git repository and change directory to this repository:

   git clone https://github.com/zzhUSTC2016/RAREMed.git
   cd RAREMed/

2. A new conda environment is suggested.

   conda create --name RAREMed

3. Activate the newly created environment.

   conda activate RAREMed

4. Install the required modules.

   sh install.sh

Download the data

1. You must have obtained access to MIMIC-III and MIMIC-IV databases before running the code.

2. Download the MIMIC-III and MIMIC-IV datasets, then unzip and put them in the data/input/ directory. Specifically, you need to download the following files from MIMIC-III: DIAGNOSES_ICD.csv, PRESCRIPTIONS.csv, and PROCEDURES_ICD.csv, and the following files from MIMIC-IV: DIAGNOSES_ICD.csv, PRESCRIPTIONS.csv, and PROCEDURES_ICD.csv.

3. Download the drugbank_drugs_info.csv and drug-DDI.csv files, and put them in the data/input/ directory.

Process the data

Run the following command to process the data:

python process.py

If things go well, the processed data will be saved in the data/output/ directory. You can run the models now!

Run the models

usage: main_RAREMed.py [-h] [-n NOTE] [--model_name MODEL_NAME] [--dataset DATASET] [--early_stop EARLY_STOP] [-t] [-l LOG_DIR_PREFIX] [-p PRETRAIN_PREFIX] [--cuda CUDA] [-s] [-e] [-nsp] [-mask] [--pretrain_epochs PRETRAIN_EPOCHS] [--mask_prob MASK_PROB] [--embed_dim EMBED_DIM] [--encoder_layers ENCODER_LAYERS] [--nhead NHEAD] [--batch_size BATCH_SIZE] [--adapter_dim ADAPTER_DIM] [--lr LR] [--dropout DROPOUT] [--weight_decay WEIGHT_DECAY] [--weight_multi WEIGHT_MULTI] [--weight_ddi WEIGHT_DDI]

RAREMed model training and evaluation script

optional arguments:
  -h, --help            show this help message and exit
  -n NOTE, --note NOTE  User notes
  --model_name MODEL_NAME
                        model name (default: 'RAREMed')
  --dataset DATASET     dataset (default: 'mimic-iii')
  --early_stop EARLY_STOP
                        early stop after this many epochs without improvement (default: 10)
  -t, --test            test mode
  -l LOG_DIR_PREFIX, --log_dir_prefix LOG_DIR_PREFIX
                        log dir prefix like "log0", for model test (default: None)
  -p PRETRAIN_PREFIX, --pretrain_prefix PRETRAIN_PREFIX
                        log dir prefix like "log0", for finetune (default: None)
  --cuda CUDA           which cuda (default: 6)
  -s, --patient_seperate
                        whether to combine diseases and procedures
  -e, --seg_rel_emb     whether to use segment and relevance embedding layer (default: True)
  -nsp, --pretrain_nsp  whether to use nsp pretrain
  -mask, --pretrain_mask
                        whether to use mask prediction pretrain
  --pretrain_epochs PRETRAIN_EPOCHS
                        number of pretrain epochs (default: 20)
  --mask_prob MASK_PROB
                        mask probability (default: 0)
  --embed_dim EMBED_DIM
                        dimension of node embedding (default: 512)
  --encoder_layers ENCODER_LAYERS
                        number of encoder layers (default: 3)
  --nhead NHEAD         number of encoder head (default: 4)
  --batch_size BATCH_SIZE
                        batch size during training (default: 1)
  --adapter_dim ADAPTER_DIM
                        dimension of adapter layer (default: 128)
  --lr LR               learning rate (default: 1e-5)
  --dropout DROPOUT     dropout probability of transformer encoder (default: 0.3)
  --weight_decay WEIGHT_DECAY
                        weight decay (default: 0.1)
  --weight_multi WEIGHT_MULTI
                        weight of multilabel margin loss (default: 0.005)
  --weight_ddi WEIGHT_DDI
                        weight of ddi loss (default: 0.1)

Example:

python main_RAREMed.py -nsp -mask    # pretrain and train
python main_RAREMed.py -t -l=log0    # test

For Baselines:

python main_GBert_pretrain.py
python main_GBert.py -p=0
python main_GBert.py -t -l=log1

python main_GAMENet.py
python main_GAMENet.py -t -l=log0

Acknowledgement

Thanks to Jing Yi(井怡) for her help in the implementation of the RAREMed model. This repository is partially based on the SafeDrug repository, you can find some additional details in the original repository.

Welcome to contact me [email protected] for any question.