13lheytens / micron

Processing files and Code for IJCAI'21 Paper: MICRON

Luke's Reproduction of MICRON findings

• src/ contains the original author's code.

• luke_src/ contains luke's code as he reproduces the findings of the original author.

• data/ contains the data for training and testing, along with a preprocessing script from the original author.

  Citation to the original paper and repo

  • This repo is a fork of the original author's (Chaoqi Yang) implementation found here: https://github.com/ycq091044/MICRON

  • The original paper can be found here: https://arxiv.org/pdf/2105.01876.pdf

    Chaoqi Yang, Cao Xiao, Lucas Glass, and Jimeng Sun.
      2021. Change matters: Medication change prediction 
      with recurrent residual networks.
    

  Dependencies

  I used:

  • MacOs 11.6.2

  • Python 3.9.4

  • The python dependencies I used for this project can be found in requirements.txt.

  • The dependencies can be installed using the following command:

    pip3 install -r requirments.txt
    

  Raw data download instruction

  • Download the following MIMIC-III Patient data files from PhysioNet. https://physionet.org/content/mimiciii/1.4/
    • PRESCIPTIONS.csv
    • DIAGNOSES_ICD.csv
    • PROCEDURES_ICD.csv
  • Place the above files in the data/ directory.
  • The preprocessed data is already available in the data/records_final.pkl file.

  Preprocessing code

  • I used the (modified) preprocessing script of the original author to ensure that I was using the same filtered dataset.

  • This script filters and combines MIMIC-III patient prescriptions, procedure ICD codes, diagnosis codes, as well as drug code mappings into the records_final.pkl file.

  • To process:

    cd data
    python3 preprocessing.py
    

  Training code

  • Each luke_src/predict_* file contains the model implementation, as well as the script for training and evaluating the model.

  • To train MICRON model:

    cd luke_src
    python3 predict_MICRON.py --save_trained
    

  • The "--save_trained" flag will save the trained model to the "pretrained_models/" folder.

  • Other models are trained in the same way.

  Evaluation code

  • To evaluate a pretrained model, use the flag "--test_only".
  • The models are also evaluated at the end of training.
  • To evaluate the pretrained MICRON model:

    cd luke_src
    python3 predict_MICRON.py --test_only
    

  • Other models are evaluated in the same way.

  Pretrained model

  • Pretrained models are found in the "luke_src/pretrained_models/" folder.
  • These are the result of training each model with the "--save_trained" flag.

  Table of results

  Model	F1 Score	Jaccard
  Baseline Model 1	.603	.444
  Baseline Model 2	.638	.479
  Baseline Model 3	.661	.505
  GameNet	.499*	.347*
  MICRON	.669	.513
  MICRON Ablation	.662	.505

  ** GameNet results are lower than expected, and could very likely be improved with more compute resources and iterations on the model.