Comprehension of Contextual Semantics Across Clinical Healthcare Domains

The widespread lack of adoption of clinical notetaking standards has rendered information retrieval from Electronic Health Records (EHRs) especially challenging using traditional Natural Language Processing (NLP) techniques. Clinical note authors too commonly adopt their own note-taking structures and styles, limiting the applicability of rule-based and statistical models. While the context of any given sentence within a note carries important implied information, context is notoriously difficult for a language model to infer. However, recent advances in deep learning NLP methods such as pre-training on domain-specific corpora, novel embedding structures, and transformer architectures have enabled an awareness of context not previously attainable. In this work, I study the application of these evidenced NLP approaches to a gold standard annotated corpus of primary care notes of multiple Mayo Clinic EHR systems. The strongly labelled data will be supplemented with large volumes of weakly labelled data curated using distant supervision. The combined dataset will be used to train and evaluate context classification and section boundary detection models that classify the current context of a sentence given adjacent text segments. Once validated against primary care corpora, transfer learning methods will enable access to shared knowledge across more specific clinical domains, enabling generalizability across clinical domains and a degree of transparency into the shared aspects of the integrated model.