Skip to content

Adding tasks

YAIB provides five clinical prediction tasks by default:

No Task Frequency Type
1 ICU mortality Once per stay (after 24 h) Binary classification
2 Acute kidney injury (AKI) Hourly (within 6 h) Binary classification
3 Sepsis Hourly (within 6 h) Binary classification
4 Kidney function (KF) Once per stay Regression
5 Length of stay (LoS) Hourly (within 7 days) Regression

A task has two parts: the cohort (which stays, which variables, which label), defined in YAIB-cohorts, and the task config in YAIB, which says how to read and model it.

Changing the cohort

Cohorts are built to be adapted. For example, predicting the ordinal KDIGO stage instead of the presence of AKI only requires changing the outcome variable from aki to kdigo. If the general setup stays the same (covariates, exclusion criteria, prediction times) but a new outcome should be predicted, define the outcome as a clinical concept in ricu and use it as the outcome variable. If the covariates, exclusion criteria or prediction times change, the existing helper functions can be reused; the existing task definitions in YAIB-cohorts are examples.

Writing a task config

The task configs in configs/tasks/ cover binary classification (BinaryClassification.gin), regression (Regression.gin) and imputation (DatasetImputation.gin). A new cohort in the standard format works with them directly. For a cohort with other files, columns or settings, copy one and change it:

configs/tasks/MyTask.gin
include "configs/tasks/common/Imports.gin"
include "configs/tasks/common/CrossValidation.gin"

Run.mode = "Classification"          # "Classification", "Regression" or "Imputation"
NUM_CLASSES = 2
HORIZON = 24
train_common.weight = "balanced"
train_common.ram_cache = True

DLPredictionWrapper.loss = @cross_entropy

preprocess.file_names = {
    "DYNAMIC": "dyn.parquet",
    "OUTCOME": "outc.parquet",
    "STATIC": "sta.parquet",
}
vars = {
    "GROUP": "stay_id",
    "SEQUENCE": "time",
    "LABEL": "my_outcome",
    "DYNAMIC": ["hr", "map", "resp", "o2sat", "temp"],
    "STATIC": ["age", "sex"],
}
preprocess.preprocessor = @base_classification_preprocessor
preprocess.vars = %vars
preprocess.use_static = True

include "configs/tasks/common/Dataloader.gin"

Then run it with -t MyTask. The file name is the task name.

Variants of the defaults show smaller changes: BinaryClassificationNoStatic.gin and BinaryClassificationNoDynamic.gin train without the static or the dynamic variables. You can also select variable groups at run time with --modalities and a modality_mapping, as in PredictionTaskVariables.gin.

For multiclass classification, set NUM_CLASSES to the number of classes; the metrics switch to the multiclass set (accuracy, balanced accuracy, AUC).