Stratification

How a site's OMOP data becomes a balanced training dataset.

Status of this page

To be filled.

This page will describe how data held locally in the OMOP Common Data Model is selected and balanced so that it can be used for training, and why the step is necessary.

Intended contents

Why stratification is needed. What goes wrong when sites contribute unbalanced data to a shared model, and what the step is intended to prevent.

Starting point. Which OMOP tables and fields are used, and what state the data must be in beforehand.

Selecting a cohort. How the population for a study is defined, and how that definition is expressed.

Balancing. The variables data is balanced across, how the strata are chosen, and what happens where a site has too few records in a stratum.

Producing the dataset. The procedure, where the result is written, and what format it takes.

Verification. How to confirm the resulting dataset is balanced as intended, with the expected output of each check.

Consistency between sites. How sites agree on the same definition, so that their contributions are comparable.

Privacy. What the dataset contains, whether it leaves the site, and what protections apply.