Administrative and claims databases are attractive for safety studies of anticoagulant and antithrombotic drugs. However, the validity of such data is often uncertain. It was our aim to assess the usefulness of the Swedish administrative health databases for detection of major bleeding events. All individuals with atrial fibrillation in Stockholm County from 2006 to 2013 (n=78,022) were identified from the Swedish Patient register. A search for bleeding diagnoses was done in the Patient register and in the Cause of Death register. The medical records of a random sample of 761 patients were studied and classified in a blinded and pre-specified way. The highest sensitivity (99.5 %) and specificity (94.0 %) were obtained by counting fatal bleeding events with the bleeding diagnosis recorded as first or second cause of death, and all hospitalisations without regard to the position of the diagnosis. Codes for transfusions were unreliable and did not increase accuracy. The registries identified 99.4 % of intracranial bleeding events and 82.6 % of gastrointestinal bleeding events correctly. All patients classified as "no bleedings" were indeed without bleeding. Overall the sensitivity was 85.5 % and the specificity 95.9 % for major bleeding events. In conclusion, Swedish nationwide health registries are well suited for conducting outcome studies involving identification of bleeding events. The use of a diagnostic code for bleeding, irrespective of its position as primary or secondary diagnosis, provided the best sensitivity and specificity for detection of bleeding events, as long as it was limited to contacts resulting in hospital admission.
Keywords: Bleeding; database; safety study; validity.