Toward Creating a Gold Standard of Drug Indications from FDA Drug Labels

Having quick access to trustworthy drug-disease relationships (which drug(s) are approved for treating or preventing which disease(s)) is one of the top information needs of health providers, consumers, and researchers. This paper presents a semi-automatic approach that can lead to the creation of a gold standard of drugs and their indications. As our system input, we use the Daily Med, which houses the most current drug labels submitted to FDA by pharmaceutical companies. Extraction of specific indications from FDA labels is a challenging problem that requires distinguishing indications from other disease mentions. In response, we first identify the candidate indications from drug labels using UMLS resources and BioNLP tools, and then rely on expert judgments to validate those pre-computed indications through an interactive Web interface. For preliminary analysis, we recruited two experts to manually annotate 100 labels of frequently sought human prescription drugs at PubMed Health. We find that the resultant expert-curated gold standard on drugs and their indications is high-quality (precision=97%, recall=94%), and differs from existing resources in that it is factual, structured, and dose-form specific. The study findings suggest the feasibility of the proposed method toward building a comprehensive resource of drug indications.