RESUMO
Radiology reports are written primarily in natural language. Automated extraction of malignant findings from narrative reports is an important technique for clinical support or alert generation for physicians. This paper proposes a method for automatically extracting malignant findings from narrative radiological reports written in Japanese. First, sentences are parsed and a medical attribute of each phrase is determined. Next, sub-trees related to radiological findings are extracted from a dependency tree using medical attributes. Finally, the malignant findings in each sub tree are extracted with their positive or negative assertions, each of which is determined by the multiplication of pos/neg signs along a path in a sub-tree. The recall and precision for the extraction of malignant findings with their positive or negative assertions were 76% and 91% respectively. The experimental results showed the validity of the proposed method for extracting malignant findings with correct assertions.
Assuntos
Armazenamento e Recuperação da Informação/métodos , Sistemas Computadorizados de Registros Médicos , Processamento de Linguagem Natural , Neoplasias/diagnóstico , Humanos , Imageamento por Ressonância Magnética , Neoplasias/diagnóstico por imagem , Serviço Hospitalar de Radiologia , Sistemas de Informação em Radiologia , Tomografia Computadorizada por Raios XRESUMO
This paper reports on the assigning of MeSH (Medical Subject Headings) categories to Japanese terms in an English-Japanese dictionary using the titles and abstracts of articles indexed in MEDLINE. In a previous study, 30,000 of 80,000 terms in the dictionary were mapped to MeSH terms by normalized comparison. It was reasoned that if the remaining dictionary terms appeared in MEDLINE-indexed articles that are indexed using MeSH terms, then relevancies between the dictionary terms and MeSH terms could be calculated, and thus MeSH categories assigned. This study compares two approaches for calculating the weight matrix. One is the TF*IDF method and the other uses the inner product of two weight matrices. About 20,000 additional dictionary terms were identified in MEDLINE-indexed articles published between 2000 and 2004. The precision and recall of these algorithms were evaluated separately for MeSH terms and non-MeSH terms. Unfortunately, the precision and recall of the algorithms was not good, but this method will help with manual assignment of MeSH categories to dictionary terms.
Assuntos
Indexação e Redação de Resumos , MEDLINE , Medical Subject Headings , Algoritmos , Dicionários Médicos como Assunto , Japão , Terminologia como Assunto , TraduçãoRESUMO
This paper introduces and reports the results for a project to map Japanese medical terms to the UMLS Metathesaurus. The "Thesaurus for Medical and Health related Terms version 5" published in 2003 by the Japan Medical Abstracts Society and UMLS version 2002AC provided by NLM were used in this study. The goal was to judge the validity of the correlation between the Japanese and English terms that belong to the same MeSH concept. Fifteen medicine, nursing, and library science professionals, excluding JAMAS, used a custom designed Web interface to perform this task. About 10% of the concepts were judged as invalid, and the reasoning behind these failures were analyzed. Experience from this project can be used to estimate the manpower required to revise the Japanese thesaurus after future revisions to UMLS or MeSH.
Assuntos
Unified Medical Language System , Vocabulário Controlado , Internet , Japão , Idioma , Medical Subject Headings , Interface Usuário-ComputadorRESUMO
This study is aimed at extracting diagnosis with positive or negative assertion from radiological report written in Japanese Natural Language. We get frequency of verb patterns that indicate pos/neg assertion, and extract a rule in order of the occurrence. We made customized dictionary of 36,152 terms relating to disease names or radiological findings, and tried to extract pairs of (pos/neg, disease and verb pattern ) by using rules according to the most frequent pattern from 1,524/5,000 CT reports (each report consists of 15.1 words on the average). We tried only a few rules so far, and continue to find other rules.