Teaching Computers to Match Patient Records Like a Human

For public health information systems, identifying and matching patient records from numerous providers poses a common challenge. How well the system can accomplish this task has a direct effect on its efficacy and credibility. In 2006, the Utah State Immunization Information System (USIIS) team was forced to address the integrity of its system when the patient duplication rate exceeded 21% and the number of records requiring manual review approached 1,000,000. This condition resulted in incomplete immunization records, frustrated users, inaccurate statistical measures, and excessive time devoted to manually reviewing records. The USIIS program staff identified its electronic record matching algorithm as the area where efforts would most improve the situation. A project was formed to develop and deploy a new record matching algorithm. The team conceived to design the replacement computer algorithm based on human record matching techniques. The foundation of this design was knowledge gained by staff during many years of evaluating and correcting record matching faults. The manual patient record matching practices of these expert de-duplication staff were observed and documented. Their knowledge of recognizing data patterns that identified matches and non-matches was employed to describe and program a computer algorithm. The project revealed patient record matching methods that were expected based on common knowledge, while others were not so obvious. Deploying the new record matching algorithm in December 2007, the USIIS team attained successful results. Patient record duplication rate was reduced by over 12% points. Furthermore, the number of records requiring manual review declined by over 90%.