Unifying Heterogeneous Electronic Health Record Systems via Clinical Text-Based Code Embedding

Kyunghoon Hur, Jiyoung Lee, Jungwoo Oh, Wesley Price, Young-Hak Kim, Edward Choi

This repository provides official Pytorch code to implement DescEmb, a code-agnostic EHR predictive model.

The paper can be found in this link: Unifying Heterogeneous Electronic Health Record Systems via Clinical Text-Based Code Embedding

Requirements

• PyTorch version >= 1.8.1
• Python version >= 3.7

Getting started

Prepare training data

First, download the dataset from these links:

MIMIC-III

eICU

ccs_multi_dx_tool_2015

icd10cmtoicd9gem

Second, make directory sturcture like below:

data_input_path
├─ mimic
│  ├─ ADMISSIONS.csv
│  ├─ PATIENTS.csv
│  ├─ ICUSYAYS.csv
│  ├─ LABEVENTES.csv
│  ├─ PRESCRIPTIONS.csv
│  ├─ PROCEDURES.csv
│  ├─ INPUTEVENTS_CV.csv
│  ├─ INPUTEVENTS_MV.csv
│  ├─ D_ITEMDS.csv
│  ├─ D_ICD_PROCEDURES.csv
│  └─ D_LABITEMBS.csv
├─ eicu
│  ├─ diagnosis.csv
│  ├─ infusionDrug.csv
│  ├─ lab.csv
│  ├─ medication.csv
│  └─ patient.csv
├─ ccs_multi_dx_tool_2015.csv
└─ icd10cmtoicd9gem.csv

data_output_path
├─mimic
├─eicu
├─pooled
├─label
└─fold

Then run preprocessing code

$ python preprocess_main.py
    --src_data $data
    --dataset_path $data_src_directory
    --dest_path $run_ready_directory 

Note that pre-processing takes about 1hours in 128 cores of AMD EPYC 7502 32-Core Processor, and requires 60GB of RAM.

Examples

Pre-training a model

Pre-train a DescEmb model with Masked Language Modeling (MLM)

$ python main.py \
    --distributed_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --src_data $data \
    --task mlm \
    --mlm_prob $percent \
    --model $model

Pre-train a CodeEmb model with Word2Vec

$ python main.py \
    --distributed_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --src_data $data \
    --task w2v
    --model codeemb

$data should be set to 'mimic' or 'eicu'

$percent should be set to probability (default: 0.3) of masking for MLM

$model should be set to 'descemb_bert' or 'descemb_rnn'

Training a new model

Other configurations will set to be default, which were used in the DescEmb paper.

$descemb should be 'descemb_bert' or 'descemb_rnn'

$ratio should be set to one of [10, 30, 50, 70, 100] (default: 100)

$value should be set to one of ['NV', 'VA', 'DSVA', 'DSVA_DPE', 'VC']

$task should be set to one of ['readmission', 'mortality', 'los_3day', 'los_7day', 'diagnosis']

Note that --input-path should be the root directory containing preprocessed data.

Train a new CodeEmb model:

$ python main.py \
    --distributed_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --model ehr_model \
    --embed_model codeemb \
    --pred_model rnn \
    --src_data $data \
    --ratio $ratio \
    --value_mode $value \
    --task $task

Train a new DescEmb model:

$ python main.py \
    --disrtibuted_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --model ehr_model \
    --embed_model $descemb \
    --pred_model rnn \
    --src_data $data \
    --ratio $ratio \
    --value_mode $value \
    --task $task

Note: if you want to train with pre-trained BERT model, add command line parameters --init_bert_params or --init_bert_params_with_freeze. --init_bert_params_with_freeze enables the model to load and freeze BERT parameters.

Fine-tune a pre-trained model

Fine-tune a pre-trained CodeEmb model:

$ python main.py \
    --distributed_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --model_path /path/to/model.pt \
    --load_pretrained_weights \
    --model ehr_model \
    --embed_model codeemb \
    --pred_model rnn \
    --src_data $data \
    --ratio $ratio \
    --value_mode $value \
    --task $task

Fine-tune a pre-trained DescEmb model:

$ python main.py \
    --distributed_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --model_path /path/to/model.pt \
    --load_pretrained_weights \
    --model ehr_model \
    --embed_model $descemb \
    --pred_model rnn \
    --src_data $data \
    --ratio $ratio \
    --value_mode $value \
    --task $task

Transfer a trained model

$ python main.py \
    --distributed_world_size $WORLDSIZE \
    --input_path /path/to/data \
    --model_path /path/to/model.pt \
    --transfer \
    --model ehr_model \
    --embed_model $embed_model \
    --pred_model rnn \
    --src_data $data \
    --ratio $ratio \
    --value_mode $value \
    --task $task \

Note that --embed_model and pred_model should be matched with the transferred model.

License

This repository is MIT-lincensed.

Citation

Please cite as:

@misc{hur2021unifying,
      title={Unifying Heterogenous Electronic Health Records Systems via Text-Based Code Embedding}, 
      author={Kyunghoon Hur and Jiyoung Lee and Jungwoo Oh and Wesley Price and Young-Hak Kim and Edward Choi},
      year={2021},
      eprint={2108.03625},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}