TAPER-EHR code (Electronic Health Record Representation Learning)
This repository contains the code for (arxiv link)
article{darabi2019taper,
title={TAPER: Time-Aware Patient EHR Representation},
author={Darabi, Sajad and Kachuee, Mohammad and Fazeli, Shayan and Sarrafzadeh, Majid},
journal={arXiv preprint arXiv:1908.03971},
year={2019}
}
Dependencies
This project was run in a conda virtual environment on Ubuntu 16.04 with CUDA 10.0. Dependencies:
- pytorch with torchvision. Note that you need a fairly recent version of PyTorch (e.g., 1.0).
- tensorflow with tensorboard (This is not critical, however, you can comment out all references to TF and TB)
Checkout the requirements.txt
file.
The Pretrained Bert Model
Bert models are typically trained in tensorflow and hence require to be ported into pytorch.
For example, download the Biobert-Base v1.1
Once you've downloaded the model, to convert it into pytorch run the following command from ./model/bert_things/pytorch_pretrained_bert/
folder
python convert_to_pytorch.py --path <path-to-biobert-folder>/biobert_model.ckpt --config <path-to-biobert-folder>/bert_config.json --save <path-to-save-converted-model>
Once the pretrained model has been converted, we can load it into our bert model in pytorch.
Running the experiments
Make sure you have downloaded the MIMIC III dataset (specifically the csv files). Once you have downloaded the dataset run the following to compile a dataframe where rows correspond to admissions:
python gen_data_df.py -p <path-to-mimic-folder> -s <path-to-save-files> -min-adm 1 -ft "Discharge summary" Physician ECG Radiology Respiratory Nursing Pharmacy Nutrition
The above will generate a dataframe containing demographics, medical codes, and medical texts for each admission. The min-adm
argument is used to filter out patients with less than the specified number of admissions and ft
argument is used to filter texts that are included in the patients row . Next, we will generate a dictionary containing patient id's as keys and each patient will have a list of dicts, where these dicts contain the patients data & label for that admission.
python gen_text_plus_code_data.py -p '<path-to-where-df-were-saved>/df_less_1.pkl' -s '<path-to-save-output>' -et -cpb './data/biobert_pubmed/bert_config.json' -sdp './data/biobert_pubmed_model' -vpb './data/biobert_pubmed/vocab.txt' -bsl 512 -diag -proc -med -cpt -sc -vp <path-to-medical-code-vocabulary>
In the above command
- et is a flag to embed text [this is done to speed up training later, otherwise you'd have to embed text as you are training your network for downstream tasks]
- cbp is the path to the pretrained bert config
- sdp is the path to the pretrained bert state_dict
- vbp is the path to the vocab on which bert was trained on
- bsl is the maximum sequence length of the bert model
- diag is a flag to include diagnoses codes in patient's data output
- proc is a flag to include procedure codes in patient's data output
- med is a flag to include medical codes in patient's data output
- cpt is a flag to include CPT codes in patient's data output
- sc is a flag to use short codes reducing the sparsity of the codes as codes have heirarchies; if you specify this you must add vp parameter
- vp path to diagnoses and cpt maps. These should be dicts that map a medical code -> id (an integer).
Take a look here for HCUP diagnosis/CPT code map examples.
Once the above script finishes, generate kfold splits by running
python kfold.py -p <path-to-data.pkl> -s <path-to-directory-data.pkl-is-stored>/splits
These splits will be used to determine the train and validation idx, and to avoid cross-contamination in the pre-training phases.
Note that in the dataloaders it is assumed the splits_X.pkl are stored in the splits
directory inside the same directory in which data.pkl was saved.
we can pretrain both code and text models:
Running Code Training
To run the code pretraining step, there are example config files in ./configs/codes
For example to run a pretraining step for diagnoses codes run the following from train.py directory
python train.py -c ./configs/codes/diag.json
Running Text Training
To run the code summarizer training step, there are example config files in ./configs/text
For example, to train the text summarizer on discharge run the following (make sure the paths are correctly set in the config file)
python train.py -c ./configs/text/pubmed_discharge.json
Running Classification
To run classification runs examples config files are in ./configs/taper. For example to run the readmission task run
python train.py -c ./configs/taper/red.json
Acknowledgments
This repository includes code from:
- The huggingface transformer very initial commits.
- TensorboardLogger from this gist