VitalML: Predicting Patient Decompensation from Continuous Physiologic Monitoring in the Emergency Department
After cloning the repo, to install dependencies from the environment.yml file, run:
conda env create -f environment.yml
Update all data source paths in path_configs.json
See transformer/training_example.sh for an example of a slurm script that runs the transformer/training.py program. Arguments are as follows:
python training.py <waveform embedding size> <task> <waveform lead> <timepoint>
Waveform embedding size is the embedding size in the layer directly prior to the output. Task is the decompensation prediction task, so one of HR, SpO2, MAP or MEWS. Waveform lead specifies the usage of ECG, Pleth or All. Lastly, timepoint indicates the length of the evaluation period for predicting decompensation and is one of 60min, 90min or 120min.
To train and tune models across all configurations of input data, run:
python lgbm/prediction.py
The above script will also evaluate the best model on the test set after tuning and selection of the best random seed. However, to bootstrap the best model's performance across the test set, run the following:
python lgbm/bootstrap_results.py
To run secondary analyses on the your models of interest (in our case the baseline and best-performing models shown in best_model_configs.json, run:
python analysis/run_analysis <analysis_mode>
The following analysis modes are available:
shap: Identifies highest contributing features to model predictionmews: Performs subgroup analysis on model decompensation predictions based on ground truth MEWS score labelscalibration_curve: Graphs calibration plotscharacteristic: Computes model characteristics such as sensitvity, specifity, PPV and NPV on the test datasetconfusion: Calculates confusion matrix of model predictions on the test dataset
The source code for the site is licensed under the MIT license, which you can find in the LICENSE file.