Advanced methods for detection of adverse drug events in clinical notes.

Manual chart review is used as the gold standard in many adverse drug event (ADE) detection studies. Owing to large resource utilization and expense this method is generally reserved for research studies. Building an expert system capable of mimicking the human expert's decision pathway would increase the efficiency of ADE detection. The first step in building such an expert system was to identify the expert for the task of detecting ADEs in manual chart-review. A systematic review and meta-analysis of studies using chart review as the method of detection of ADEs, was conducted. Results showed that pharmacists were capable of detecting higher incidence rates than other clinical specialties. The next step was to evaluate the decision-making processes used by the pharmacists for AIDE detection. Think-aloud analysis was used to identify signals pharmacists looked for while using the method of chart-review. Verbal protocol analysis also gave an insight into the gaps that exist between pharmacists' information needs and existing clinical information systems. The textual signals extracted using think-aloud analyses were limited in their scope because they represented only the case-scenarios that were presented in the focus groups. In order to make these signals generalizable, the use of the method of propositional analysis to evaluate the semantic structure of the think-aloud protocols, was proposed. A proposition for the detection of ADEs in the clinical notes consists of two types of information, first, the concepts representing the `adverse event' and those representing the drugs'. A second type of information needed would be the relationship expressed between the drug and the adverse event. A comparison of text-based techniques for identifying the first type of information represented in the proposition was conducted. This study evaluated the feasibility of using propositional analysis for identification of ADEs in clinical notes. This study used a combination of methodologies from the domains of cognitive science and artificial intelligence for detecting ADEs in clinical notes. Future work will focus on two specific directions. First, the automated extraction of propositions for ADE detection. Second, the development of rules that combine textual signals with medication