Clinical text annotation - what factors are associated with the cost of time?

Building high-quality annotated clinical corpora is necessary for developing statistical Natural Language Processing (NLP) models to unlock information embedded in clinical text, but it is also time consuming and expensive. Consequently, it important to identify factors that may affect annotation time, such as syntactic complexity of the text- to-be-annotated and the vagaries of individual user behavior. However, limited work has been done to understand annotation of clinical text. In this study, we aimed to investigate how factors inherent to the text affect annotation time for a named entity recognition (NER) task. We recruited 9 users to annotate a clinical corpus and recorded annotation time for each sample. Then we defined a set of factors that we hypothesized might affect annotation time, and fitted them into a linear regression model to predict annotation time. The linear regression model achieved an R² of 0.611, and revealed eight time-associated factors, including characteristics of sentences, individual users, and annotation order with implications for the practice of annotation, and the development of cost models for active learning research.