medcat 0.1.7

Concept annotation tool for Electronic Health Records

oncept Annotation Tool

A simple tool for concept annotation from UMLS or any other source.

This is still experimental

How to use

There are a few ways to run CAT, simplest one being docker.

Docker

If using docker the appropriate models will be automatically downloaded, you only need to run:

docker build --network=host -t cat -f Dockerfile.MedMen .

Once the container is built start it using:

docker run --env-file=./envs/env_medann -p 5000:5000 cat

You can now access the API on

<YOUR_IP>:5000/api

OR a simple frontend

<YOUR_IP>:5000/api_test

Installation

pip install --upgrade medcat

Please install the langauge models before running anything

python -m spacy download en_core_web_sm

pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.2.0/en_core_sci_md-0.2.0.tar.gz

Building a new Concept Database (.csv) or using an existing one

First download the vocabulary from Vocabulary Download

Now in python3+

from cat.cat import CAT
from cat.utils.vocab import Vocab
from cat.prepare_cdb import PrepareCDB
from cat.cdb import CDB 

vocab = Vocab()

# Load the vocab model you just downloaded
vocab.load_dict('<path to the vocab file>')

# If you have an existing CDB
cdb = CDB()
cdb.load_dict('<path to the cdb file>') 

# If you need a special CDB you can build one from a .csv file
preparator = PrepareCDB(vocab=vocab)
csv_paths = ['<path to your csv_file>', '<another one>', ...] 
# e.g.
csv_paths = ['./examples/simple_cdb.csv']
cdb = preparator.prepare_csvs(csv_paths)

# Save the new CDB for later
cdb.save_dict("<path to a file where it will be saved>")

# To annotate documents we do
doc = "My simple document with kidney failure"
cat = CAT(cdb=cdb, vocab=vocab)
cat.train = False
doc_spacy = cat(doc)
# Entities are in
doc_spacy._.ents
# Or to get a json
doc_json = cat.get_json(doc)

# To have a look at the results:
from spacy import displacy
# Note that this will not show all entites, but only the longest ones
displacy.serve(doc_spacy, style='ent')

# To run cat on a large number of documents
data = [(<doc_id>, <text>), (<doc_id>, <text>), ...]
docs = cat.multi_processing(data)

Training and Fine-tuning

To fine-tune or train everything from the ground up (excluding word-vectors), you can use the following:

# Loadinga CDB or creating a new one is as above.

# To run the training do
f = open("<some file with a lot of medical text>", 'r')
# If you want fine tune set it to True, old training will be preserved
cat.run_training(f, fine_tune=False)

If building from source, the requirements are

python >= 3.5 [tested with 3.7, but most likely works with 3+]

All the rest can be instaled using pip from the requirements.txt file, by running:

pip install -r requirements.txt

Results

Dataset	SoftF1	Description
MedMentions	0.798	The whole MedMentions dataset without any modifications or supervised training
MedMentions	0.786	MedMentions only for concepts that require disambiguation, or names that map to more CUIs
MedMentions	0.92	Medmentions filterd by TUI to only concepts that are a disease

Models

A basic trained model is made public. It is trained for the 35K entities available in MedMentions. It is quite limited so the performance might not be the best.

Vocabulary Download - Built from MedMentions

Trained CDB Download

(Note: This is was compiled from MedMentions and does not have any data from NLM as that data is not publicaly available.)

Acknowledgement

Entity extraction was trained on MedMentions In total it has ~ 35K entites from UMLS

The dictionary was compiled from Wiktionary In total ~ 800K unique words For now NOT made publicaly available