Semantics-driven Clinical Decision Support in Radiology

Abstract

Radiological reporting is generating large quantities of content within the electronic health record, which is a valuable source of information for improving clinical care. Harnessing the potential of novel applications in radiology requires automated conversion of report content into a semantic representation that can be processed computationally. Natural language processing (NLP) techniques can be leveraged to derive meaning from narrative input in such manner. Benefits for the use of NLP on diagnostic reporting have been described in the literature [1,2], identifying variable types of applications. One such unexplored category of applications is integration in the clinical workflow to support radiologists at the time of reporting. The interactive features of such applications can assist the radiologist in real time, for example by linking report content to background knowledge, guidelines and other reference material [1]. A high-quality resource for diagnostic guidance during radiologic reporting is Elsevier’s STATdx®. This diagnostic decision support system for radiologists gives you instant access to the collective clinical experience and knowledge of renowned sub-specialists in every field of radiology [3]. STATdx® is unique for its embedded knowledge structure, which maps radiographic finding to the associated differential diagnostic considerations. The current usage experience of STATdx® relies on active queries by radiologists. Performing such query temporarily interrupts the reporting workflow, where want radiologist to spent most of their time. Leveraging novel NLP technology has the potential to automatically and unobtrusively retrieve appropriate STATdx® topics. This upgrades the user experience from passive search to active recommendation, enabling seamless presentation of the right content for the reporting context. Providing content in context may improve utilizations of important information sources with the potential to reduce uncertainty in diagnostic reporting, improve the quality of care and provide a powerful experience for radiology education. Therefore, we performed a pilot that assesses technological feasibility of STATdx® Enhanced Radiology Reporting, which we consider a benchmark for the feasibility of semantics-driven clinical decision support applications. Given adequate retrieval performance user supervision of search may gradually decrease and contextual supervision gradually increase over time. In summary, the objective of the Proof-of-Concept (POC) for STATdx Enhanced Radiology Reporting was to assess technical feasibility using NLP and ML for automatically retrieving and ranking relevant STATdx reference material (findings and diagnosis) specific to a radiology report given as input.

Ähnliche Arbeiten