I am an ICU dataset developer and publisher, with a research focus on data bias in routinely collected clinical data. When transforming the dataset to the OMOP CDM, I encountered the following issue.
Documentation notes, ETL should use "as much information as possible from the data" [1]. Nevertheless, it gives no guidance on data validity levels, although data that is not used in clinical practice is often highly invalid.
An example is the granularity of drug_concept_id.
Validity Estimation: (absolute numbers represent Salzburg University Hospital ICU Data (SICdb[2]) 2013-2021)
Ingredient: 3.8 % invalid entries (outliers etc.), (MAR estimated ~1%, MNAR probably insigificant)
Clinical Drug Component: Estimated 5-10% invalid data
Branded X: Estimated >50-80% invalid data
Example for invalidity Ingredient: 1 kg of paracetamole documented (wrong unit)
Example for Clinical Drug Component: 1.5 Units of Enoxaparin 40 mg applied (when changing dosage, product is not updated)
For Branded Drugs, as long as no closed-loop documentation is established, it is actually quite unlikely the exact product is applied. At least in Austria almost every branded drug changes regularily, different brands in same drawer. Without closed-loop it is not possible to document this.
At least at my units, the product information in templates is (almost at the point of setting up) outdated.
So, guidance would be needed. "...as much information as possible from the data" or "...as much reliable information as possible from the data", and where to cut of?
This may appliy to other topics too, but drug_concept_id is a good example.
- https://ohdsi.github.io/CommonDataModel/cdm54.html#drug_exposure
- https://doi.org/10.1038/s41597-024-03164-9
I am an ICU dataset developer and publisher, with a research focus on data bias in routinely collected clinical data. When transforming the dataset to the OMOP CDM, I encountered the following issue.
Documentation notes, ETL should use "as much information as possible from the data" [1]. Nevertheless, it gives no guidance on data validity levels, although data that is not used in clinical practice is often highly invalid.
An example is the granularity of drug_concept_id.
Validity Estimation: (absolute numbers represent Salzburg University Hospital ICU Data (SICdb[2]) 2013-2021)
Ingredient: 3.8 % invalid entries (outliers etc.), (MAR estimated ~1%, MNAR probably insigificant)
Clinical Drug Component: Estimated 5-10% invalid data
Branded X: Estimated >50-80% invalid data
Example for invalidity Ingredient: 1 kg of paracetamole documented (wrong unit)
Example for Clinical Drug Component: 1.5 Units of Enoxaparin 40 mg applied (when changing dosage, product is not updated)
For Branded Drugs, as long as no closed-loop documentation is established, it is actually quite unlikely the exact product is applied. At least in Austria almost every branded drug changes regularily, different brands in same drawer. Without closed-loop it is not possible to document this.
At least at my units, the product information in templates is (almost at the point of setting up) outdated.
So, guidance would be needed. "...as much information as possible from the data" or "...as much reliable information as possible from the data", and where to cut of?
This may appliy to other topics too, but drug_concept_id is a good example.