Extraction of drug-drug interactions using all paths graph kernel

Abstract

Drug-drug interactions (DDIs) cause nearly 3% of all hospital admissions. Regulatory authorities such as the Food and Drug Administration (FDA) and the pharmaceutical companies keep a rigorous tab on the DDIs. The major source of DDI information is the biomedical literature. In this paper we present a DDI extraction approach based on all paths graph kernel [1] from the DrugDDI corpus [2]. We also evaluate the method on an in-house developed clinical in vivo pharmacokinetic DDI corpus. When the DDI extraction model was evaluated on the test dataset from both corpora we recorded a F-score of 0.658 on the clinical in vivo pharmacokinetic DDI corpus and 0.16 on the DrugDDI corpus.