an easy-to-use interface to fully trained BERT based models for multi-class and multi-label long document classification.
pre-trained models are currently available for two clinical note (EHR) phenotyping tasks: smoker identification and obesity detection.
To sustain future development and improvements, we interface pytorch-transformers for all language model components of our architectures. Additionally, their is a blog post describing the idea behind the architecture.
This repository contains an updated implementation that corrects are error found in the original version of the preprint
Install with pip:
pip install bert_document_classification
or directly:
pip install git+https://github.com/AndriyMulyar/bert_document_classification
Maps text documents of arbitrary length to binary vectors indicating labels.
from bert_document_classification.models import SmokerPhenotypingBert
from bert_document_classification.models import ObesityPhenotypingBert
smoking_classifier = SmokerPhenotypingBert(device='cuda', batch_size=10) #defaults to GPU prediction
obesity_classifier = ObesityPhenotypingBert(device='cpu', batch_size=10) #or CPU if you would like.
smoking_classifier.predict(["I'm a document! Make me long and the model can still perform well!"])
More examples.
Go to the directory /examples/ml4health_2019_replication. This README will give instructions on how to appropriately insert data from DBMI to replicate the results in the paper.
- For training you will need a GPU.
- For bulk inference where speed is not of concern lots of available memory and CPU cores will likely work.
- Model downloads are cached in
~/.cache/torch/bert_document_classification/
. Try clearing this folder if you have issues.
If you found this project useful, consider citing our extended abstract.
@misc{mulyar2019phenotyping,
title={Phenotyping of Clinical Notes with Improved Document Classification Models Using Contextualized Neural Language Models},
author={Andriy Mulyar and Elliot Schumacher and Masoud Rouhizadeh and Mark Dredze},
year={2019},
eprint={1910.13664},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
Implementation, development and training in this project were supported by funding from the Mark Dredze Lab at Johns Hopkins University.