17 Seconds: Referential Matching Helps Reduce Duplicate Patient Records

In June 2018, Northwell turned to another approach to resolve its most-challenging match problems. Referential matching uses data from outside sources—for example, the U.S. Postal Service, other governmental agencies, credit reporting services, and cell phone providers—to create a fuller picture of past addresses, nicknames, or alternative name spellings and other demographic data for patients. 

That allows the health system to identify patient matches that its own records would not reveal. For example, if a new record is created for Paul Smith who lives at 361 Catalpa, probabilistic matching within Northwell’s own master patient index would not link it with the record for Paul E. Smith who lives at 1400 Market. But referential matching could show that Paul E. Smith, now living at 361 Catalpa, used to live at 1400 Market and that both records include identical birthdates and other demographic information, increasing the likelihood that the records belong to a single individual.   

In probabilistic matching, Northwell is limited to the data contained in the 6 million patient records in its master patient index. Referential matching gives it access to many more data points for more than 300 million lives.

“In our master patient index, patient information may not be always accurate and up to date, and we are only as good as the data we can match upon,” Aarnes says. “The key to this vendor is they have correct data and incorrect data on almost any adult living in the United States.”

Northwell does not use referential matching to automatically merge duplicate records. Rather, the vendor’s answer key—”yes, this is a match” or “no match”—is used as a job aid for Northwell staff members who manually review the new patient records created each day.

Before referential matching was introduced, members of the master patient index team took, on average, one minute and 42 seconds to determine if a new record was a duplicate, not a duplicate, or could not be resolved. With the referential-matching answer as another information source, the average time per record is now 17 seconds, Aarnes says.

Interviewed for this article:

Keely N. Aarnes is assistant vice president of business operations, Northwell Health, Lake Success, N.Y.