Interpret textual data generated from medical vocal memos

In the Library of Celsus in Ephesus, built in the 2nd century, there are four statues depicting wisdom (Sophia), knowledge (Episteme), intelligence (Ennoia) and excellence (Arete). Our project is named after this city and the goddess Sophia.

After visiting a patient nurses and doctors need to quickly and easily send information

So they record a vocal memo after each visit

Today these memos are read by humans and the infos are manually entered in the database

We want to ease their work by automatically extracting informations from the vocal memos and pre-filling the informations to be entered in the database

4000 vocal memo recordings (4000 sentences)

14 targets to predict (up to 14 different pieces of informations per memo)

Here is an example of a memo

And here is the corresponding informations we need to extract

Clean the data from stop words and punctuation

We identify which part of the memo (which group of words) corresonds to which information

For this, we build a Named Entity Recognition model (NER) using the spaCy library

We build models to convert each information into the target classes using the nltk library

You can play around with our demo here

In this demo, we let you try your own sentences and see the results from our models

Show our success percentage

Give our feedback on possible improvement points and share the hypotheses we used to build our models

Clone the project:

git clone [email protected]:GeoffroyGit/ephesus.git

We recommend you to create a fresh virtual environment

Create a python3 virtualenv and activate it:

cd ephesus
pyenv virtualenv ephesus
pyenv local ephesus

Upgrade pip if needed:

pip install --upgrade pip

Install the package:

pip install -r requirements.txt
pip install -e .

Run the API on your machine:

make run_api

Build the docker image:

make docker_build

Run a container on your machine:

make docker_run

Stop the container running on your machine

docker ps
docker stop <container id>

Push the image to Google Cloud Platform (GCP):

make docker_push

Run a container on GCP:

make docker_deploy

You'll need similar training data in order to train the models

We're sorry we can't share our data

mkdir models
mkdir models/config

Download base config on https://spacy.io/usage/training (select only French and NER) and save it to models/config/base_config.cfg

Fill config file with default values:

cd models/config/
python -m spacy init fill-config base_config.cfg config.cfg

Create train set and test set for the model:

cd ephesus/
python sentence.py

Create variable to host training data file name (put the same names as in ephesus/sentence.py):

export EPHESUS_TRAINING_DATA = "train_set_v2.spacy"
export EPHESUS_TEST_DATA = "test_set_v2.spacy"

Train the model:

cd models/
mkdir model_v2
cd models/config/
python -m spacy train config.cfg --output ../model_v2 --paths.train ../../raw_data/$EPHESUS_TRAINING_DATA --paths.dev ../../raw_data/$EPHESUS_TRAINING_DATA

Evaluate the model:

cd models/model_v2/
mkdir eval
cd models/config/
python -m spacy evaluate ../model_v2/model-best ../../raw_data/$EPHESUS_TEST_DATA -dp ../model_v2/EVAL -o ../model_v2/EVAL/model_v2_scores.json

Train and evaluate the models for treatment and location:

cd ephesus/
python nlp.py

Check the classes for date and time:

cd ephesus/
python timedate.py

Congratulations!