Natural language processing in pathology: a scoping review. Issue 11 (22nd July 2016)

• Record Type: Journal Article
• Authors: Burger, Gerard; Abu-Hanna, Ameen; de Keizer, Nicolette; Cornet, Ronald
• Is Part Of: Journal of clinical pathology. Volume 69:Issue 11(2016)
• Page Start: 949
• Page End: 955
• Publication Date: 2016-07-22
• Subjects: COMPUTER SYSTEMS -- SURGICAL PATHOLOGY -- REPORTS; Pathology -- Periodicals; Pathology, Molecular -- Periodicals; 616.0705
• DOI: 10.1136/jclinpath-2016-203872
• Languages: English
• ISSNs: 0021-9746

Abstract

Abstract : Background: Encoded pathology data are key for medical registries and analyses, but pathology information is often expressed as free text. Objective: We reviewed and assessed the use of NLP (natural language processing) for encoding pathology documents. Materials and methods: Papers addressing NLP in pathology were retrieved from PubMed, Association for Computing Machinery (ACM) Digital Library and Association for Computational Linguistics (ACL) Anthology. We reviewed and summarised the study objectives; NLP methods used and their validation; software implementations; the performance on the dataset used and any reported use in practice. Results: The main objectives of the 38 included papers were encoding and extraction of clinically relevant information from pathology reports. Common approaches were word/phrase matching, probabilistic machine learning and rule-based systems. Five papers (13%) compared different methods on the same dataset. Four papers did not specify the method(s) used. 18 of the 26 studies that reported F-measure, recall or precision reported values of over 0.9. Proprietary software was the most frequently mentioned category (14 studies); General Architecture for Text Engineering (GATE) was the most applied architecture overall. Practical system use was reported in four papers. Most papers used expert annotation validation. Conclusions: Different methods are used in NLP research in pathology, and good performances, that is, high precision and recall, high retrieval/removal rates, are reported for all of these. Lack of validation and of shared datasets precludes performance comparison. More comparative analysis and validation are needed to provide better insight into the performance and merits of these methods.