2019/04/05 by Eben Holderness, Holderness, Eben, Philip Cawkwell +7 · 1 citation
Psychology · Computer Science · #Mental Health via Writing #Machine Learning in Healthcare #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1904.03225
Recently natural language processing (NLP) tools have been developed to\nidentify and extract salient risk indicators in electronic health records\n(EHRs). Sentiment analysis, although widely used in non-medical areas for\nimproving decision making, has been studied minimally in the clinical setting.\nIn this study, we undertook, to our knowledge, the first domain adaptation of\nsentiment analysis to psychiatric EHRs by defining psychiatric clinical\nsentiment, performing an annotation project, and evaluating multiple\nsentence-level sentiment machine learning (ML) models. Results indicate that\noff-the-shelf sentiment analysis tools fail in identifying clinically positive\nor negative polarity, and that the definition of clinical sentiment that we\nprovide is learnable with relatively small amounts of training data. This\nproject is an initial step towards further refining sentiment analysis methods\nfor clinical use. Our long-term objective is to incorporate the results of this\nproject as part of a machine learning model that predicts inpatient readmission\nrisk. We hope that this work will initiate a discussion concerning domain\nadaptation of sentiment analysis to the clinical setting.\n