Fast Model Adaptation for Automated Section Classification in Electronic Medical Records

摘要

Medical information extraction is the automatic extraction of structured information from electronic medical records, where such information can be used for improving healthcare processes and medical decision making. In this paper, we study one important medical information extraction task called section classification. The objective of section classification is to automatically identify sections in a medical document and classify them into one of the pre-defined section types. Training section classification models typically requires large amounts of human labeled training data to achieve high accuracy. Annotating institution-specific data, however, can be both expensive and time-consuming; which poses a big hurdle for adapting a section classification model to new medical institutions. In this paper, we apply two advanced machine learning techniques, active learning and distant supervision, to reduce annotation cost and achieve fast model adaptation for automated section classification in electronic medical records. Our experiment results show that active learning reduces the annotation cost and time by more than 50%, and distant supervision can achieve good model accuracy using weakly labeled training data only.