I’m looking for any literature, published methodology, or real-world examples addressing this specific scenario: two or more datasets that have each already been independently converted to OMOP by different organizations or teams (separate ETL processes, no shared mapping), how would you go about integrating them, and what methods would you use to validate consistency/agreement between the two?
We know it’s common practice to combine OMOP datasets for analysis. What we’re specifically interested in is: what are people’s thoughts on checking for consistency between two datasets before combining them?
This matters critically for us because both datasets would be used together in the same analysis. Without confidence that the two or more OMOP instances are genuinely comparable, we risk a real problem: any differences we observe in our results could simply reflect differences in how each dataset was mapped or ETL’d, rather than a true difference in the underlying data, which would undermine the reliability of any conclusions we draw.
If anyone has worked through this, comparing or merging two or more independently-produced OMOP instances, or has a process for checking consistency between sources before combining, I’d really appreciate being pointed to any papers, internal writeups, tools, or lessons learned.
Welcome to the community! The problem you’re trying to solve is well known in the community. I don’t have a deeper answer beyond ‘Use DQD checks and make sure your dataset is fit for purpose’, which it seems you’ve already mastered. I just wanted to give your post some visibility.
The best people to contact would be Dr. @clairblacketer , who leads the Evidence Network Working Group (WG) and often coordinates network studies, and @MPhilofsky, who leads the THEMIS Working Group (WG) and our community ETL efforts.
Do your datasets must have full PHI and multiple attributes of a person: full name, date of birth, country level identifier (i.e. Social Security Number), etc.? If not, you shouldn’t combine them. And this sentence leads me to believe you do not have PHI/PII for your two OMOP CDMs:
If you combine them, then could be duplicating the data which will skew and invalidate your results.
What’s your use case for combining the datasets? Why not leave them separate?
In the OHDSI community, many health systems have EHR and other sources of data. It is common practice for them to combine more than one datasets (EHR, registry, claims, etc.) into one OMOP CDM to make their CDM more comprehensive. But they have full PHI, PII data.
Thanks for your response.
To clarify our situation a little more: both organisations do have PHI/PII for the individuals. We have performed an OMOP transformation for a dataset from community health records, and we are collaborating with a local hospital that also has PHI/PII for its patients and has independently transformed its data to OMOP.
Because both datasets cover the same geographical region, they contain many of the same individuals. Our goal is to link the records for these shared individuals and combine the datasets to create a more complete longitudinal view of their healthcare across both settings, rather than simply appending two datasets together.
We’re looking for guidance on the practical aspects of doing this after two independent OMOP transformations. Beyond running the Data Quality Dashboard (DQD), are there any additional checks or validation steps that are recommended? If inconsistencies arise between the two OMOP datasets (for example, differences in concept mappings, dates, or duplicate events), are there any best practices for resolving them? We’d also be interested to hear about others’ experiences and lessons learned from combining independently OMOP-converted datasets in this type of scenario.
, who leads the Evidence Network Working Group (WG) and often coordinates network studies, and