Abstract 238: Automating Medical Record Matching: A Key Component of an Automated Cardiac Arrest Registry

Abstract

Introduction: An automated registry for out-of-hospital cardiac arrest can facilitate cardiac arrest research and emergency medical services (EMS) and community health quality improvement programs. A key element of a cardiac arrest registry is outcome, especially survival to hospital discharge, which is usually obtained from the hospital medical record. A crucial step in capturing survival data automatically is being able to reliably match EMS patient care reports with hospital medical records. Objective: To determine the accuracy of a probabilistic record-linkage approach to match EMS and hospital encounters and identify barriers of automation. Methods: This is a prospective feasibility study from the Dallas Fort-Worth Center for Resuscitation Research. Our cardiac arrest registry was part of the Resuscitation Outcomes Consortium (ROC) Epistry from 2006 to 2016. Since 2016, we maintained it separately from the ROC and have since enrolled over 20,000 cases. Record linkage was performed using demographic data of encounters from a single EMS agency within the registry and a community hospital in Dallas where these were transported from January 2017 to May 2018. An expectation maximization algorithm to aid in discerning matches and nonmatches was performed using the R package RecordLinkage identifying record pairs as a match, non-match, or possible match based on a pre-determined threshold. Match and Possible Match pairs were manually reviewed to assess the algorithm’s accuracy. Results: The EMS and hospital data sets had 67 and 6,050 unique encounters, respectively. The record linkage algorithm identified 40 EMS encounters as having a match (60%), 12 with a possible match (18%), and 15 with no identified match (22%). After manual review, the algorithm had a positive predictive value (PPV) of 100% (40/40) for the match group with a sensitivity of 60%. Conclusions: Applying probabilistic linkage to aid in automating a cardiac arrest registry was successful in linking over half of the EMS encounters with a PPV of 100%. When reviewing the unmatched encounters, many were adjacent to each other when sorted by time suggesting the poor sensitivity is more likely a reflection of the hospital’s data collection method as opposed to the linkage methodology.