Deep Learning of Clinical Relation Identification in Health Narratives

The worldwide adoption of electronic health records (EHRs) to document patient data enables the use of big-data methods to harness the medical information contained therein for secondary use. While most medical informatics research focuses on using the structured data found in EHRs, there is a substantial amount of information in the narratives of the records that is inaccessible without processing. This dissertation focuses on extracting this information in the form of medical concepts and relations between them. Specifically, deep learning methods are presented to perform (1) concept detection; and (2) relation extraction. Multiple deep learning methods are explored including recurrent neural networks, convolutional neural networks, memory networks, and attention networks. Methods are presented for performing these tasks in multiple genres of EHRs with differing target concept and relation schemata. Moreover, due to the data-hungry nature of deep learning models and the expertise necessary to annotate medical narratives, this dissertation addresses the problem of training deep learning models using multi-task active learning. These methods can be used to extract data-driven knowledge encoding clinical expertise from practicing clinicians. As such, this dissertation explores methods for representing such knowledge in graphical structures that can be used for question answering and to infer new knowledge. Moreover, these techniques are extended to represent biomedical knowledge from expert-curated ontologies as embeddings. These embeddings expose otherwise inaccessible biomedical knowledge to deep learning models for relation identification.