Skip to content

IJCAI2021: Code for SafeDrug, MIMIC data processing, Medical code mapping

Notifications You must be signed in to change notification settings

ycq091044/SafeDrug

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 
 
 
 
 

Repository files navigation

Data and Code for IJCAI'21 paper - SafeDrug

YOU NEED TO KNOW FIRST!

  • The results using the archived branch (results of the paper):

    DDI: 0.0589 (0.0005) Ja: 0.5213 (0.0030) F1: 0.6768 (0.0027) PRAUC: 0.7647 (0.0025)
    
  • The results using this master branch (@J-Zhangg obtained this in issue #23, thanks):

    # When the learning rate is set to 5e-4:
    DDI: 0.0632 (0.0003) Ja: 0.5114 (0.0026) F1: 0.6676 (0.0023) PRAUC: 0.7649 (0.0028)
    
    # When the learning rate is set to 2e-4:
    DDI: 0.0607 (0.0005) Ja: 0.5089 (0.0022) F1: 0.6659 (0.0019) PRAUC: 0.7632 (0.0022)
    

[Implementation difference] Here are two main differences:

  1. As we mentioned below, the main difference of two branches is in how we get the drug SMILES string (the paper crawl methods misses a lot of molecules, while the current branch uses drugbank method, which gives more comprehensive sets).
  2. The data processing scripts are also a bit difference, and thus output data statistics differ from the ones reported in the paper.

[which branch to use?] General guidance:

  1. This master branch contains more descriptions (to learn how to use our codes), and the folder structures are very similar to archived branch.
  2. Use the archived branch to reproduce the results in the paper.

Folder Specification

  • data/
    • procesing.py our data preprocessing file.
    • Input/ (extracted from external resources)
    • Output/
      • atc3toSMILES.pkl: drug ID (we use ATC-3 level code to represent drug ID) to drug SMILES string dict
      • ddi_A_final.pkl: ddi adjacency matrix
      • ddi_matrix_H.pkl: H mask structure (This file is created by ddi_mask_H.py)
      • ehr_adj_final.pkl**: used in GAMENet baseline (if two drugs appear in one set, then they are connected)
      • records_final.pkl: The final diagnosis-procedure-medication EHR records of each patient, used for train/val/test split (NOTE: we only provide the first 100 entries as examples here. We cannot distribute the whole MIMIC-III data https://physionet.org/content/mimiciii/1.4/, then please download the dataset by yourself and use our processing code to obtain the full records.).
      • voc_final.pkl: diag/prod/med index to code dictionary
  • src/
    • SafeDrug.py: our model
    • baselines:
      • GAMENet.py
      • DMNC.py: there are some issues for the latest dnc package, please refer to the original DMNC repo https://github.com/thaihungle/DMNC
      • Leap.py
      • Retain.py
      • ECC.py
      • LR.py
    • setting file
      • model.py
      • util.py
      • layer.py

Note that we previously use ./data/get_SMILES.py for getting SMILES strings from drugbank. However, due to the web structure change of drugbank, this crawler is not used in the current pipeline. Now, we are using drugbank_drugs_info.csv to obtain the SMILES string for each ATC3 code, thus, the data statistics differ a bit from the paper. The current statistics are shown below:

#patients  6350
#clinical events  15032
#diagnosis  1958
#med  112
#procedure 1430
#avg of diagnoses  10.5089143161256
#avg of medicines  11.647751463544438
#avg of procedures  3.8436668440659925
#avg of vists  2.367244094488189
#max of diagnoses  128
#max of medicines  64
#max of procedures  50
#max of visit  29

High-level Clarifications on How to Map ATC Code to SMILES

ALSO, check out this tool for easy medication code mapping https://github.com/ycq091044/MedCode.

Step 1: Package Dependency

  • first, install the rdkit conda environment
conda create -c conda-forge -n SafeDrug  rdkit
conda activate SafeDrug

# can also use the following in your current env
pip install rdkit-pypi
  • then, in SafeDrug environment, install the following package
pip install scikit-learn, dill, dnc

Note that torch setup may vary according to GPU hardware. Generally, run the following

pip install torch

If you are using RTX 3090, then plase use the following, which is the right way to make torch work.

python3 -m pip install --user torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
  • Finally, install other packages if necessary
pip install [xxx] # any required package if necessary, maybe do not specify the version, the packages should be compatible with rdkit

Here is a list of reference versions for all package

pandas: 1.3.0
dill: 0.3.4
torch: 1.8.0+cu111
rdkit: 2021.03.4
scikit-learn: 0.24.2
numpy: 1.21.1

Let us know any of the package dependency issue. Please pay special attention to pandas, some report that a high version of pandas would raise error for dill loading.

Step 2: Data Processing

  • Go to https://physionet.org/content/mimiciii/1.4/ to download the MIMIC-III dataset (You may need to get the certificate)

    cd ./data
    wget -r -N -c -np --user [account] --ask-password https://physionet.org/files/mimiciii/1.4/
  • go into the folder and unzip three main files

    cd ./physionet.org/files/mimiciii/1.4
    gzip -d PROCEDURES_ICD.csv.gz # procedure information
    gzip -d PRESCRIPTIONS.csv.gz  # prescription information
    gzip -d DIAGNOSES_ICD.csv.gz  # diagnosis information
  • download the DDI file and move it to the data folder download https://drive.google.com/file/d/1mnPc0O0ztz0fkv3HF-dpmBb8PLWsEoDz/view?usp=sharing

    mv drug-DDI.csv ./data
  • processing the data to get a complete records_final.pkl

    cd ./data
    vim processing.py
    
    # line 323-325
    # med_file = './physionet.org/files/mimiciii/1.4/PRESCRIPTIONS.csv'
    # diag_file = './physionet.org/files/mimiciii/1.4/DIAGNOSES_ICD.csv'
    # procedure_file = './physionet.org/files/mimiciii/1.4/PROCEDURES_ICD.csv'
    
    python processing.py

Step 3: run the code

python SafeDrug.py

here is the argument:

usage: SafeDrug.py [-h] [--Test] [--model_name MODEL_NAME]
               [--resume_path RESUME_PATH] [--lr LR]
               [--target_ddi TARGET_DDI] [--kp KP] [--dim DIM]

optional arguments:
  -h, --help            show this help message and exit
  --Test                test mode
  --model_name MODEL_NAME
                        model name
  --resume_path RESUME_PATH
                        resume path
  --lr LR               learning rate
  --target_ddi TARGET_DDI
                        target ddi
  --kp KP               coefficient of P signal
  --dim DIM             dimension

Citation

@inproceedings{yang2021safedrug,
    title = {SafeDrug: Dual Molecular Graph Encoders for Safe Drug Recommendations},
    author = {Yang, Chaoqi and Xiao, Cao and Ma, Fenglong and Glass, Lucas and Sun, Jimeng},
    booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI} 2021},
    year = {2021}
}

Welcome to contact me [email protected] for any question. Partial credit to https://github.com/sjy1203/GAMENet.

About

IJCAI2021: Code for SafeDrug, MIMIC data processing, Medical code mapping

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published