The challenge
After three mergers, Tidewater had 1.9 million patient records and no confident count of its patients. Nearly one record in five was a duplicate, split across systems that each spelled names and addresses their own way.
The data team maintained over a thousand mapping tables to keep the systems talking, and onboarding a new source took six weeks of engineering.
“We had 1.9 million records and no confident count of our patients. Atlas matched them in a week, and nobody on my team wrote a single mapping table.”
What they did
- 1.
Matching on evidence
Atlas matched records on date of birth, MRN, address history, and more, and kept every source value alongside the merged one.
- 2.
Humans for the hard cases
Matches Atlas couldn't settle with confidence went to a reviewer queue rather than being guessed.
- 3.
No mapping tables
New sources map themselves into the shared model, so the data team stopped maintaining translations by hand.
The result
In thirty days, Atlas resolved 412,000 duplicates into one person each, bringing the duplicate rate from 18% to 0.4%. Onboarding a new source now takes three days, and the team retired all 1,140 mapping tables.