Data Mining Technique for Signal Detection of Drug Adverse Events Using Health Insurance Claims / 薬剤疫学

ABSTRACT

Objective: To detect signals of potential drug adverse events (DAEs) through data mining of health insurance claims.<BR>Design and Data Retrospective observational study. The data used were the database of health insurance claims collected and maintained by the Japan Medical Data Center consisting of 312, 797 medical and pharmaceutical claims in one year (August 2003 through July 2004) linked uniquely for 35, 410 patients using an encryption technique to ensure privacy.<BR>Methods: We counted all combinations (cross product or Cartesian product) of drugs and diagnoses appearing in the same claims and counted the number of times a given drug was prescribed preceding the suspected diagnosis in all combinations of the drug and the diagnosis appearing in a claim, i.e., the prescription date precedes the diagnosing date (the preceding number). We calculated the expected preceding number from the overall prevalence of drugs and diagnoses, and then calculated the observed and expected ratio, which was used as the signal indices. We calculated the signal indices on the health insurance claims data to detect DAEs of psychiatric drugs.<BR>Results: Amoxapine and trazodone HCL showed high signal indices with paralytic ileus and convulsion (epilepsy) as documented in their package inserts. However, paroxetine HCL and etizolam showed high signal indices with these potential adverse events although no such DAEs are documented in their package inserts.<BR>Conclusions: The undocumented high signal indices observed between the drugs and diagnoses indicate the potential DAEs and warrant in-depth pharmacovigilance. Given the strength of health insurance claims with a well-defined source population and accurate drug exposure, the proposed signal index will likely prove to be an effective data mining technique when combined with nested case-control analysis and counter-matching.