Generating cohorts¶
The YAIB-cohorts repository extracts the task cohorts from the ICU datasets in the format YAIB reads. It uses the ricu R package, which defines clinical concepts and how to load them from each dataset, so that every dataset is accessed through the same interface.
Clone it next to YAIB:
The cohorts can be extracted from R or from Python. The README of YAIB-cohorts is the authoritative reference; the steps below give an overview.
Run these commands from YAIB-cohorts/R.
1. Set up the environment. The extractions for the paper ran with R 4.2.2. An renv lock file pins all
package versions:
2. Load the datasets into ricu. Download the datasets you have access to and import them as described in
?ricu::import_src. ricu also ships two demo datasets, mimic.demo and eicu.demo, for quick experiments; renv
should install them (otherwise see mimic-demo and
eicu-demo). Then set the data path in .Rprofile:
3. Generate the cohorts. All tasks share a data-cleaning step, base_cohort.R, that must run first:
Replace mortality.R with the task and mimic_demo with the dataset: one of mimic_demo, eicu_demo, aumc,
hirid, eicu, mimic, miiv. To extract a task from all datasets at once, run bash gen_cohort.sh mortality.
Run these commands from YAIB-cohorts/Python.
1. Set up the environment:
2. Load the datasets into ricu. As for R: import the datasets with ?ricu::import_src and set
RICU_DATA_PATH in .Rprofile.
3. Generate the cohorts:
Replace mortality.py with the task and mimic_demo with the dataset, as above.
Tasks¶
| Script | Task | Type |
|---|---|---|
mortality |
ICU mortality after 24 hours | Classification |
aki |
Acute kidney injury within the next 6 hours | Classification |
sepsis |
Sepsis within the next 6 hours | Classification |
kidney_function |
Kidney function on the second day of the ICU stay | Regression |
los |
Remaining length of stay | Regression |
The output directory is set in config.yaml at the root of YAIB-cohorts. Each cohort is written as dyn.parquet,
sta.parquet and outc.parquet, which you pass to YAIB with -d. The exact label definitions, such as the time window
around AKI onset that counts as positive, are in the task scripts of YAIB-cohorts (for example R/aki.R).