Extracting structured information from free-text medication prescriptions using dependencies

Abstract

We explore an information extraction task where the goal is to determine the correct values for fields which are relevant to prescription drug administration such as dosage amount, frequency and route. The data set is a collection of prescriptions from a long-term health-care facility, a small subset of which we have manually annotated with values for these fields. We first examine a rule-based approach to the task, which uses a dependency parse of the prescription, achieving accuracies of 60-95% over various different fields, and 67.5% when all fields of the prescription are considered together. The outputs of such a system have potential applications in detecting irregularities in dosage delivery.