Abstract 2721: Probabilistic Matching of Computerized Emergency Medical Services (EMS) records and Emergency Department and Patient Discharge Data: a Novel Approach to Evaluation of Prehospital Stroke Care

Background and Purpose: Emergency Medical Services is an important element of acute stroke care. However, evaluation of prehospital stroke care is limited by lack of exchange of patient outcome data between hospitals and emergency medical services (EMS) agencies. In this study, we describe and demonstrate the feasibility of linking county wide patient level ambulance data with emergency department (EDD) and patient discharge data (PDD) using a probabilistic matching algorithm. Methods: Probabilistic linkage was used to match county-wide ambulance data from 2005-2007 to hospital (EDD and PDD) records with a final ICD -9 diagnosis of stroke (430-436). The linkage model was based on the patient’s transport/admission date, date of birth, race, sex, county of residence, and destination hospital. Probabilistic linkage was performed using LinkSolv version 8.29746 which calculates the probability that a pair of records is a true match based on agreement/disagreement patterns of the linkage variables. Pairs of records with a match probability of 0.8 or higher were considered true matches. All other pairs were false matches and rejected. Results: During 2005 - 2007 there were 310,731 patients transported to a facility in county and 34,785 hospital records with a diagnosis of stroke. Using the linkage algorithm we identified 11,473 (33%) matches with EMS records. Linkage rates increased each year with 30%, 34%, and 36% of hospital patients matching EMS record for 2005, 2006, and 2007 respectively. The median match probability was 0.993 and the IQR was 0.974 to 0.9996. By taking the compliment of the match probability we estimate our linked sample to include 255 (2%) false matches. Date of treatment/admission and the patient’s sex were observed to be the most reliable, disagreeing on less than one percent (1%) of all matched pairs. Patient’s zip code was the least reliable, disagreeing on one third of matched pairs. Conclusions: Our study demonstrates that probabilistic matching can be used to create a comprehensive patient care record which in turn can provide opportunities for researchers to study different phases of stroke care.