/UmlsBERT

Primary LanguagePythonMIT LicenseMIT

UmlsBERT: Clinical Domain Knowledge Augmentation of Contextual Embeddings Using the Unified Medical Language System Metathesaurus

General info

This is the code that was used of the paper : UmlsBERT: Augmenting Contextual Embeddings with a Clinical Metathesaurus (NAACL 2021).

In this work, we introduced UmlsBERT, a contextual embedding model capable of integrating domain knowledge during pre-training. It was trained on biomedical corpora and uses the Unified Medical Language System (UMLS) clinical metathesaurus in two ways:

  • We proposed a new multi-label loss function for the pre-training of the Masked Language Modelling (Masked LM) task of UmlsBERT that considers the connections between medical words using the CUI attribute of UMLS.

  • We introduced a semantic group embedding that enriches the input embeddings process of UmlsBERT by forcing the model to take into consideration the association of the words that are part of the same semantic group.

Technologies

This project was created with python 3.7 and PyTorch 0.4.1 and it is based on the transformer github repo of the huggingface team

Setup

We recommend installing and running the code from within a virtual environment.

Creating a Conda Virtual Environment

First, download Anaconda from this link

Second, create a conda environment with python 3.7.

$ conda create -n umlsbert python=3.7

Upon restarting your terminal session, you can activate the conda environment:

$ conda activate umlsbert 

Install the required python packages

In the project root directory, run the following to install the required packages.

pip3 install -r requirements.txt

Install from a VM

If you start a VM, please run the following command sequentially before install the required python packages. The following code example is for a vast.ai Virtual Machine.

apt-get update
apt install git-all
apt install python3-pip
apt-get install jupyter

Dowload pre-trained UmlsBERT model

In order to use pre-trained UmlsBERT model for the word embeddings (or the semantic embeddings), you need to dowload it into the folder examples/checkpoint/ from the link:

 wget -O umlsbert.tar.xz https://www.dropbox.com/s/kziiuyhv9ile00s/umlsbert.tar.xz?dl=0

into the folder examples/checkpoint/ and unzip it with the following command:

tar -xvf umlsbert.tar.xz

Reproduce UmlsBERT

Pretraining

  • The UmlsBERT was pretrained on the MIMIC data. Unfortunately, we cannot provide the text of the MIMIC III dataset as training course is mandatory in order to access the particular dataset.

  • The MIMIC III dataset can be downloaded from the following link

  • The pretraining an UmlsBERT model depends on data from NLTK so you'll have to download them. Run the Python interpreter (python3) and type the commands:

>>> import nltk
>>> nltk.download('punkt')
  • After downloading the NOTEEVENTS table in the examples/language-modeling/ folder, run the following python code that we provide in the examples/language-modeling/ folder to create the mimic_string.txt on the folder examples/language-modeling/:
python3 mimic.py

you can pre-trained a UmlsBERT model by running the following command on the examples/language-modeling/:

Example for pretraining Bio_clinicalBert:

python3 run_language_modeling.py --output_dir ./models/clinicalBert-v1  --model_name_or_path  emilyalsentzer/Bio_ClinicalBERT  --mlm     --do_train     --learning_rate 5e-5     --max_steps 150000   --block_size 128   --save_steps 1000     --per_gpu_train_batch_size 32     --seed 42     --line_by_line      --train_data_file mimic_string.txt  --umls --config_name  config.json --med_document ./voc/vocab_updated.txt

Downstream Tasks

MedNLi task

  • MedNLI is available through the MIMIC-III derived data repository. Any individual certified to access MIMIC-III can access MedNLI through the following link

    • Converting into an appropriate format: After downloading and unzipping the MedNLi dataset (mednli-a-natural-language-inference-dataset-for-the-clinical-domain-1.0.0.zip) on the folder examples/text-classification/dataset/mednli/, run the following python code in the examples/text-classification/dataset/mednli/ folder that we provide in order to convert the dataset into a format that is appropriate for the UmlsBERT model
python3  mednli.py
  • This python code will create the files: train.tsv,dev_matched.tsv and test_matched.tsv in the text-classification/dataset/mednli/mednli folder
  • We provide an example-notebook under the folder experiements/:

or directly run UmlsBert on the text-classification/ folder:

python3 run_glue.py --output_dir ./models/medicalBert-v1 --model_name_or_path  ../checkpoint/umlsbert   --data_dir  dataset/mednli/mednli  --num_train_epochs 3 --per_device_train_batch_size 32  --learning_rate 1e-4   --do_train --do_eval  --do_predict  --task_name mnli --umls --med_document ./voc/vocab_updated.txt

NER task

  • Due to the copyright issue of i2b2 datasets, in order to download them follow the link.

    • Converting into an appropriate format: Since we wanted to directly compare with the Bio_clinical_Bert we used their code in order to convert the i2b2 dataset to a format which is appropriate for the BERT architecture which can be found in the following link: link

    We provide the code for converting the i2b2 dataset with the following instruction for each dataset:

  • i2b2 2006:

    • In the folder token-classification/dataset/i2b2_preprocessing/i2b2_2006_deid unzip the deid_surrogate_test_all_groundtruth_version2.zip and deid_surrogate_train_all_version2.zip
    • run the create.sh scrip with the command ./create.sh
    • The script will create the files: label.txt, dev.txt, test.txt, train.txt in the token-classification/dataset/NER/2006 folder
  • i2b2 2010:

    • In the folder token-classification/dataset/i2b2_preprocessing/i2b2_2010_relations unzip the test_data.tar.gz, concept_assertion_relation_training_data.tar.gz and reference_standard_for_test_data.tar.gz
    • Run the jupyter notebook Reformat.ipynb
    • The notebook will create the files: label.txt, dev.txt, test.txt, train.txt in the token-classification/dataset/NER/2010 folder
  • i2b2 2012:

    • In the folder token-classification/dataset/i2b2_preprocessing/i2b2_2012 unzip the 2012-07-15.original-annotation.release.tar.gz and 2012-08-08.test-data.event-timex-groundtruth.tar.gz
    • Run the jupyter notebook Reformat.ipynb
    • The notebook will create the files: label.txt, dev.txt, test.txt, train.txt in the token-classification/dataset/NER/2012 folder
  • i2b2 2014:

    • In the folder token-classification/dataset/i2b2_preprocessing/i2b2_2014_deid_hf_risk unzip the 2014_training-PHI-Gold-Set1.tar.gz,training-PHI-Gold-Set2.tar.gz and testing-PHI-Gold-fixed.tar.gz
    • Run the jupyter notebook Reformat.ipynb
    • The notebook will create the files: label.txt, dev.txt, test.txt, train.txt in the token-classification/dataset/NER/2014 folder
  • We provide an example-notebook under the folder experiements/:

or directly run UmlsBert on the token-classification/ folder:

python3 run_ner.py --output_dir ./models/medicalBert-v1 --model_name_or_path  ../checkpoint/umlsbert    --labels dataset/NER/2006/label.txt --data_dir  dataset/NER/2006 --do_train --num_train_epochs 20 --per_device_train_batch_size 32  --learning_rate 1e-4  --do_predict --do_eval --umls --med_document ./voc/vocab_updated.txt

If you find our work useful, can cite our paper using:

@inproceedings{michalopoulos-etal-2021-umlsbert,
    title = "{U}mls{BERT}: Clinical Domain Knowledge Augmentation of Contextual Embeddings Using the {U}nified {M}edical {L}anguage {S}ystem {M}etathesaurus",
    author = "Michalopoulos, George  and
      Wang, Yuanxin  and
      Kaka, Hussam  and
      Chen, Helen  and
      Wong, Alexander",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.naacl-main.139",
    doi = "10.18653/v1/2021.naacl-main.139",
    pages = "1744--1753",
}