Clinical Natural Language Processing Transfer Learning based on ULMFit.
For details on ULMFit, see:
- Introducing state of the art text classification with universal language models - Blog Post
- Universal Language Model Fine-tuning for Text Classification - arXiv paper; also peer reviewed and accepted for the 2018 Annual Meeting of the Association for Computational Linguistics.
This project uses data from MIMIC-III. The data is freely available but does require pre-registration and some training before access will be granted.
Once you have access, you can download the .gz version of the data files with these commands - replace mimicusername with your actual username:
wget --user mimicusername --ask-password https://physionet.org/works/MIMICIIIClinicalDatabase/files/version_1_4/NOTEEVENTS.csv.gz
wget --user mimicusername --ask-password https://physionet.org/works/MIMICIIIClinicalDatabase/files/version_1_4/ADMISSIONS.csv.gz
gunzip *.csv.gz
In the Python Jupyter notebooks, the data is assumed to be located in the directory Path.home()/'mimic'
conda create -y -n fastai python=3.6
conda activate fastai
pip install dataclasses gpustat
conda install -y -c pytorch pytorch torchvision cudatoolkit=9.0
conda install -y -c fastai fastai
conda install -y ipykernel nbconvert ipywidgets scikit-learn
conda install -y -c conda-forge jupytext
conda install -y -c conda-forge jupyterlab
conda install -y -c conda-forge altair vega_datasets
Project Structure and Organization based off https://github.com/callahantiff/Abra-Collaboratory/