Adapting State-of-the-Art Deep Language Models to Clinical Information Extraction Systems: Potentials, Challenges, and Solutions

• DOI: 10.2196/11499
• Publication Date: 2019-04-25T13:15:37Z
• Authors: Liyuan Zhou; Hanna Suominen; Tom Gedeon
• Keywords: Original Paper; 03 medical and health sciences; 0302 clinical medicine; 3. Good health

Deep learning (DL) has been widely used to solve problems with success in speech recognition, visual object and detection for drug discovery genomics. Natural language processing achieved noticeable progress artificial intelligence. This gives an opportunity improve on the accuracy human-computer interaction of clinical informatics. However, due difference vocabularies context between a environment generic English, transplanting models directly from up-to-date methods real-world health care settings is not always satisfactory. Moreover, legal restriction using privacy-sensitive patient records hinders applying machine (ML) processing.The aim this study was investigate 2 ways adapt state-of-the-art extracting information free-form narratives populate handover form at nursing shift change automatically proofing revising by hand: first, domain-specific word representations second, transfer knowledge general English. We have described practical problem, composed it as ML task known extraction, proposed solving task, evaluated their performance.First, trained different domains served input DL system extraction. Second, model applied way learned text sources domain. The goal gain improvements extraction performance, especially classes that were topically related but did sufficient amount solutions available target A total 3 independent datasets generated they training (101 reports), validation (100 test reports) sets our experiments.Our now task. Domain-specific improved macroaveraged F1 3.4%. Transferring English corpora task-specific domain contributed further 7.1% improvement. best performance populating 37 headings 41.6% 81.1% filtering out irrelevant information. Performance differences its baseline statistically significant (P<.001; Wilcoxon test).To knowledge, first attempt deep specific tasks As shows advantage over other methods, limited data, less experts' time needed annotate data ML, which may enable good results even resource-poor domains.