Repository for the manuscript entitled "Comparative neural word embeddings approaches for medical concept representation and patient trajectory prediction".
- This project aims to compare NLP models (word2vec, fastTex and GloVe), based on the quality of their representation of medical concepts.
- We use MIMIC-IV to train the models, from which we extract patient trajectories as sequences of (amongst others) ICD10 and ATC codes.
- We train the models with model-specific NLP tasks that use patient trajectory sequences as input.
- We evaluate the models by producing medical concept embeddings, clustering them, comparing them to existing biomedical terminologies, and using them for medical outcome and patient trajectory prediction tasks.
-
Clone the repository:
git clone git@github.com:ds4dh/medical_concept_representation.git # or https://github.com/ds4dh/medical_concept_representation.git cd medical_concept_representation
-
Install the required dependencies:
./create_env.sh conda activate medical_representation
If you have version issues, you can build an environment with the packages listed in environment.yml
-
The project uses WandbLogger for experiment tracking. Ensure you have a Weights & Biases account set up for logging.
You need to download the data yourself! Instructions for downloading and pre-processing the data are here: https://github.com/ds4dh/medical_concept_representation/tree/main/data
Once the pre-processed data is ready, train the models with:
python run_all_models.py # long step, best in screen https://linuxize.com/post/how-to-use-linux-screen/Once the models are trained, test the trained models with:
python run_all_models.py -t # long step, best in screen https://linuxize.com/post/how-to-use-linux-screen/Result figures 4, 5, and 7 will be available at your wandb log page.
For the other result figures, run:
python figures/figure_6.py
python figures/figure_8.py
python figures/figure_8_bis.py # supplementary figures
python figures/figure_9.py