This repository dedicated to the first release of Med7: a transferable clinical natural language processing model for electronic health records, compatible with spaCy v3+, for clinical named-entity recognition (NER) tasks. The en_core_med7_lg model is trained on MIMIC-III free-text electronic health records and is able to recognise 7 categories:
Both vector and transformer models are now hosted on Huggingface.
The trained model comprises three components in its pipeline:
- tagger
- parser
- clinical NER with seven categories.
November 2024: This readme file is updated to reflect new versions of pip that allow install. This is because pip >23.0 requires addition of "@" for dependency speficiations to work. See original pip documentations for details.
See tutorial for reproducible example which has a environment.yml that is a dump of conda virtual environment for replication. Note in the example.ipynb,transformer model "en_core_med7_trf" is used.
It is recommended to create a dedicated virtual environment and install all recent required packages in there. The trained model was tested with spaCy version >=3.1 and Python >=3.7. For example, if the anaconda distribution of Python is already installed:
create a new virtual environment:
(base) conda create -n med7 python=3.9
activate and install spaCy:
(base) conda activate med7
(med7) pip install -U spacy
once all went through smoothly, install the Med7 model from the Huggingface Models repo:
Vectors model:
pip install "en-core-med7-lg @ https://huggingface.co/kormilitzin/en_core_med7_lg/resolve/main/en_core_med7_lg-any-py3-none-any.whl"
Transformer-based model:
pip install "en-core-med7-trf @ https://huggingface.co/kormilitzin/en_core_med7_trf/resolve/main/en_core_med7_trf-any-py3-none-any.whl"
This is RoBERTa-base implementation. Future works will improve its performance and introduce new feautres. Some entities may not be identified correctrly.
Notice You can download en_core_med7_lg for spaCy v2 here: https://www.dropbox.com/s/xbgsy6tyctvrqz3/en_core_med7_lg.tar.gz?dl=1
and then
pip install /path/to/downloaded/spacy2_model
import spacy
med7 = spacy.load("en_core_med7_lg")
# create distinct colours for labels
col_dict = {}
seven_colours = ['#e6194B', '#3cb44b', '#ffe119', '#ffd8b1', '#f58231', '#f032e6', '#42d4f4']
for label, colour in zip(med7.pipe_labels['ner'], seven_colours):
col_dict[label] = colour
options = {'ents': med7.pipe_labels['ner'], 'colors':col_dict}
text = 'A patient was prescribed Magnesium hydroxide 400mg/5ml suspension PO of total 30ml bid for the next 5 days.'
doc = med7(text)
spacy.displacy.render(doc, style='ent', jupyter=True, options=options)
[(ent.text, ent.label_) for ent in doc.ents]The Med7 model identifies correctly all seven entities in the following example and highlights them in different colours for better visualisation:
and the resulting output:
[('Magnesium hydroxide', 'DRUG'),
('400mg/5ml', 'STRENGTH'),
('suspension', 'FORM'),
('PO', 'ROUTE'),
('30ml', 'DOSAGE'),
('bid', 'FREQUENCY'),
('for the next 5 days', 'DURATION')]
It is straightforward to extract relations between the entities, since Med7 has both parser and tagger pipelines, similar to this example.
The code in above can also be run in Colab
This model is the very first step in our programme on clinical NLP for electronic health records (cNLPEHR). We are committed to developing FAIR - Findable, Accessible, Interoperable and Reusable tools which will benefit the wider community.
If you found this model useful, please acknowledge by citing as:
@article{kormilitzin2020med7,
title={Med7: a transferable clinical natural language processing model for electronic health records},
author={Kormilitzin, Andrey and Vaci, Nemanja and Liu, Qiang and Nevado-Holgado, Alejo},
journal={arXiv preprint arXiv:2003.01271},
year={2020}
}