Data mining clinical records to detect adverse events

Abstract

Background Longitudinal clinical data can provide new insights regarding the safety of drugs through the application of data mining. The objective of our project was to develop new technologies to detect adverse drug events when applied to electronic medical records. Methods Longitudinal clinical test data was obtained from the Department of Defense's Composite Health Care System. We developed a set of adverse drug event detection rules (likely to be an event, not likely to be an event). A time interval was selected to group an occurrence of a diagnosis with the drugs taken. We applied a prototype software to the test data using a new Bayesian Logistic Regression (BLR) approach to compute safety signal scores for the events “lactic acidosis” and “thrombocytopenia” and compared these results to the method (Multi-item Gamma Poisson Shrinker) currently used by the FDA to detect safety signals in their Adverse Event Reporting System (AERS) database. Results We demonstrated an ability to distinguish both “innocent bystanders” (drugs whose high safety signal scores were most likely due to co-prescription with other drugs more strongly associated with the event) and “guilty bystanders” (drugs which received lower signal scores due to the presence in the analysis of other drugs with stonger associations which masked the effect) for both of our test events. Conclusions We demonstrated the utility of using data mining applied to longitudinal medical records to detect adverse drug events. Clinical Pharmacology & Therapeutics (2005) 77, P1–P1; doi: 10.1016/j.clpt.2004.11.007