Mining Longitudinal Data for Known and Unknown Adverse Events

Longitudinal clinical data resources such as the Composite Health Care System are important large-scale repositories of data from real-world medical practice that can provide significant new insights regarding the safety of approved drugs and vaccines through the application of automated data mining. Such "secondary" use depends, however, on overcoming substantial information-handling obstacles: the original data, collected primarily for patient care and administration, are highly complex and fine-grained and unsuitable for direct data mining analysis. The project objective is to develop innovative software capable of (1) extracting and standardizing relevant information from medical records in military health systems, (2) utilizing temporal abstraction and pattern matching to identify medical record segments that correspond to potential drug-related adverse events, (3) performing disproportionality analyses to detect drug-event associations with greater-than-expected frequencies that may signal previously uncharacterized adverse reactions and (4) identifying patterns involving known adverse reactions useful in providing alerts to physicians. Phase I focused on developing temporal abstraction, pattern matching, and disproportionality analysis technologies that show promise in identifying adverse drug events when applied to moderate-scale test data. Phase II focuses on creating, testing, and refining a data mining workbench and associated data resources interfaced to an analysis-oriented military medical records data mart.