Code for the paper: Incorporating real-world evidence into the development of patient blood glucose prediction algorithms for the ICU.
See https://mimic.physionet.org/gettingstarted/access/
For this analysis the MIMIC-III database was stored locally using a postreSQL RDBMS.
There are details here on how to set it up. As a first step towards creating an analysis dataset a series of views were created. The scripts for these views are found in the ./sql folder.
The code in https://github.com/MIT-LCP/mimic-code formed the basis for much of the SQL code used in the project.
The following table or views were output as CSVs to a folder ./data/raw/VIEWNAME:
- icustays
- pt_icu_outcome
- admissions
- patients
- diabetes
- pv_mechvent
- pv_nutrition
- ent_nutrition
- vasopressors
- output_hourly
- glycaemic_analysis_hr
The process for saving to CSV will differ between database management software. I used DBeaver.
The CSVs created from the database views were combined into an analysis file using
the code in scripts/create-dataset/static-dataset.R and scripts/create-dataset/analysis-dataset-discrete-time.R.
The test-train split was performed using the file scripts/create-dataset/test-train-discrete-time.R.
Since there was a bug in the R version of Catboost 0.24.2 preventing use of the 'UncertaintyRegression' loss function I switched to using the Python version through R's reticulate package. Installation details for reticulate can be found here.
The experiments in determining model parameters are found in scripts/modeling/experiments.R.
The file scripts/modeling/models.R fits the final models as evaulated in the paper. The file scripts/modeling/evaluation.R outputs the evaluation metrics.
The numbers in the table can be found by running the file ./scripts/table1.R
The hypoglycaemia prediction model code is found in ./scripts/appendices/hypoglycaemia-discrete-time-models.R
If you have any issues running the code feel free to get in touch through opening an issue.