ICU predictions on MIMIC-III using discrete event input with distributed representations, resulting in state-of-the-art results for MIMIC-III Benchmark (Harutyunyan et al.):

You need MIMIC-III access to run the code.

1. Setup the benchmark and FastText submodule:

git submodule update

Setup the conda environment:

conda env create -f environment.yml

2. Generate csv files for the benchmark tasks as explained in mimic3-benchmarks/README.md under Building a benchmark. You need to generate for all tasks not only multitask as the evaluation scripts in the benchmark depend on label files generated for each task.

3. Install FastText:

cd fastText
make

4. Extract demographic information:

python -m dataloader.generate_demographic_csv.py

1. Setup a virtual environment with conda for easy CUDA support:

conda env create --name ehr --file=environments.yml
conda activate ehr

2. Extract bin-edges and patient sentences using the notebook dataloader/extract.ipynb:

This generates following files:

med_values.<table>*.txt
med_bin_edges.<table>*.txt
dem.*.params
data/sentences.mimic3.txt
embeddings/sentences.mimic3.counts

3. Train fasttext embeddings. You can skip this step as pretrained vectors are included in embeddings/:

./fastText/fasttext skipgram -input embeddings/sentences.mimic3.txt -output embeddings/sentences.mimic3.txt.100d.Fasttext.15ws

4. (Optional step) Pretrain encoders with contrastive predictive coding:

python -m pretrain cpc
<logs wandb run id>

Default parameters are in pretrain.py file.

5. Finetune on benchmark tasks with a multitask model as explained in the paper:

Without contrastive pretraining:

python -m finetune base -e 20

Using contrastive pretraining step, with its wandb run id:

python -m finetune base --finetune=<WANDB_ID>

Default parameters are in finetune.py file.

First we generate validation and test predictions as csv, and then use the evaluation scripts provided by Harutyunyan et al.

python -m collect <wandb_id>; sh evaluate.sh <wandb_id>

This results in files for predictions and the evaluation result for each task:

wandb/*-<wandb_id>/files/test_listfile_predictions/<task>-*.csv
wandb/*-<wandb_id>/files/test_listfile_predictions/<task>-*.csv.json