Deep Learning to Predict Mortality Prior to Tracheostomy Decision in Adults with Respiratory Failure
This github repo represents the code base used to perform this study.
This code base relies on MIMIC-III for data substrate, as well as queries from the MIMIC-III Code Repository. Any SQL queries refer to MIMIC-III's postgres build.
To recreate the analysis:
cohort/Cohort Selection.Rmd
is used to generate the cohort from MIMIC-III flat files.
- a
MIMIC-III
folder containing the database's flat files is required
-
cohort/cohort.sql
is used to perform more complex queries for MIMIC-III-derived data, including illness severity scores, mechanical ventilation status, etc. -
cohort/Clinical Tables.Rmd
is used to generate the entire data set from the tables which are output fromcohort.sql
. -
modeling/Preprocessing for Modeling.ipynb
is used to preprocess data (scaling, normalization) and perform the training/testing split. -
modeling/modelfinding_1yr_mort.ipynb
andmodeling/modelfinding_3mo_mort.ipynb
are used to conduct a neural architecture search for their respective models. -
modeling/Model Evaluation.ipynb
is used to analyze and visualize resulting model metrics.