Pressor Gauge is a web app designed to help physicians prioritize their care for patients who will soon become very sick and need life-saving blood pressure medication (pressors).
The web app can be found at pressorgauge.com, and the accompanying demo slides can be found here.
Clone this repository to the current working directory:
git clone git@github.com:wrs28/PressorGauge.git
Create a Conda environment (it will be called 'PressorGauge'):
conda env create -f environment.yml
Activate it with
conda activate PressorGauge
The MIMIC-III data are not included, as they require special access, though anyone can request access.
Once access is gained and the data are downloaded, set the MIMIC_DIR variable in src/building_database/directories.py to the location of the MIMIC-III data.
The shell script building_database.sh will build a SQL database locally, as specified in the engine variable in src/building_database/directories.py (the DB must be created by the user, and engine must be accordingly updated).
The other shell script building_model.sh will clean the data and build the features.
Logs of both scripts are stored in logs/build.log and logs/model.log.
- conda
- pandas
- numpy
- scipy
- scikit-learn
- psycopg2
- sqlalchemy
- shap
- streamlit
- joblib
- matplotlib
- seaborn
conda env create -f environment.yml
pip install -r requirements
To serve the model locally, simply run
streamlit run PressorGauge.py
from within the src directory.
To create and run a Docker image locally, can call the script local_build.sh.
It can be found at localhost:80
To publish a Docker image to dockerhub and push it to AWS, call the script aws-docker-build/aws_build.sh.
Note this file must be adapted for your local build.
The first line source insight activates the PressorGauge environment, changes directory to PressorGauge, and loads credentials as enviorenmental variables.
After container is activated on AWS, result can be found at pressorgauge.com.
The data has been aggregated in 12-hour windows, with a 6-hr overlap between each window. Those that end 12 hours or less from a pressor event define the positive class, everything else is negative.
Looking at the distribution between classes of some of the features shows correlation with pressor outcome. For example, this is visible in the partial-pressure O2 (blood gas test):
A preliminary model training gives random forest feature importances, which can be compared to the Shapley importances (explained here), which show that among the lab features considered here, a long PTT (prothrombin time [test]), a high partial pressure O2, and an abnormal lactate test all strongly influence the classification.
The model is trained on 11_777 test windows, and the learning curve suggests that the model does not suffer too much from overfitting.
The AUROC score of the model is .7, with mean time from prediction to event of 8 hours.
