How to solve the core problem of patient record matching

Healthcare organizations are becoming increasingly sensitive about the problems involved in matching multiple patient identities within the healthcare system, and most solutions rely on record linkage. However, there’s another way to look at it: what we ultimately need are tools that intelligently adapt and learn variations in ID record fields, and can take advantage of all the available information to properly identify someone.

Consider what we learned in the final report on the Patient Identification and Matching Initiative sponsored by the Office of the National Coordinator for Health Information Technology (ONC). The report found health systems can achieve matching rates of 90 percent or better within a single hospital. Not bad. But the rate drops dramatically when matching records with other providers—even those using the same MDM/MPI vendor.

As an example, the report says Kaiser Permanente achieved a matching success rate greater than 90 percent within instances of a single EMR. Yet matching fell to between 50 percent and 60 percent across regions and outside partners with separate instances of that same EMR. Other health systems and vendors saw the same kind of drop-off in matching records across systems and networks.

Why the dramatic difference in matching rates? It’s all in the way health systems handle duplicate records. Each hospital usually has a dedicated team which cleans up duplicates. If there are four John Smith records with significant variations in data fields, team members look at ID records to see if at least some are for the same John Smith. This is all done manually, with trained professionals comparing demographic fields to make the determination.

The problem is that it’s not practical to have a single team cleaning up records across all hospitals in a network or ACO. So record matching is usually done by machines that lack the same critical thinking skills as dedicated hospital staff. It’s further complicated by the many different ways people describe themselves over time.

For example, John Smith’s record at Hospital A might have his home phone number and his given name. His record at Hospital B might have his cell phone number and list him as Jack Smith, which might be his preferred nickname. Both records are for the same person, but they don’t match. An automated matching solution won’t know Jack is a common substitute for John, and won’t reason that records with the same birthdate and address are probably the same person.

The most common approach has been to use Social Security numbers (SSNs) as a unique identifier. This is no longer an option, as large numbers of data system breaches have driven CMS to ban the use of SSNs, as specified in the The Medicare Access and CHIP Reauthorization Act of 2015.

Common automated matching systems also won’t realize duplicate records are created when a patient changes addresses, or gets married and changes her last name. Because simple field matching can’t reconcile these and many other data variations, the match rate plunges to KP’s example of 50 percent to 60 percent across hospitals and ambulatory settings.

Clinicians know providing the highest level of care quality requires a complete picture of a patient’s health and medical history. Treating people with the highest level of integrity requires deep commitment to accuracy. The best way to get there is with new solutions that intelligently adapt and learn about individuals, not simply ID record fields. This requires innovative tools that move beyond the rules-based and probabilistic models of the past to utilizing advances made in artificial, neural networks and machine learning. The evolving medical networks require new entity resolution approaches to fulfill their promise.