J-MeDic: A Japanese Disease Name Dictionary based on Real Clinical Usage

J-MeDic: A Japanese Disease Name Dictionary based on Real Clinical Usage Kaoru Ito, Hiroyuki Nagai, Taro Okahisa, Shoko Wakamiya, Tomohide Iwao, Eiji Aramaki Nara Institute of Science and Technology 8916-5 Takayama, Ikoma, Nara 630-0192, Japan {kito, hironagai, taro-o, wakamiya, iwao, aramaki}@is.naist.jp Abstract Medical texts such as electronic health records are necessary for medical AI development. Nevertheless, it is difficult to use data directly because medical texts are written mostly in natural language, requiring natural language processing (NLP) for medical texts. To boost the fundamental accuracy of Medical NLP, a high coverage dictionary is required, especially one that fills the gap separating standard medical names and real clinical words. This study developed a Japanese disease name dictionary called “J-MeDic” to fill this gap. The names that comprise the dictionary were collected from approximately 45,000 manually annotated real clinical case reports. We allocated the standard disease code (ICD-10) to them with manual, semi-automatic, or automatic methods, in accordance with its frequency. The J-MeDic covers 7,683 concepts (in ICD-10) and 51,784 written forms. Among the names covered by J-MeDic, 55.3% (6,391/11,562) were covered by SDNs; 44.7% (5,171/11,562) were covered by names added from the CR corpus. Among them, 8.4% (436/5,171) were basically coded by humans), and 91.6% (4,735/5,171) were basically coded automatically. We investigated the coverage of this resource using discharge summaries from a hospital; 66.2% of the names are matched with the entries, revealing the practical feasibility of our dictionary. Keywords: Medical NLP, Case reports, Discharge summary, Named entity, ICD-10 1. Introduction Medical data are fundamentally important resources for the development of medical AI and information extraction tools. Among various data, Electronic Health Records (EHR) are a promising resource because they include detailed information about a patient and diagnosis processes. Nevertheless, it is difficult to extract information from EHR because several expressions refer to the same concept. Orthographical variations present particular difficulty, especially for the Japanese language, in which characters of at least five kinds are used: Hiragana, Katakana, Kanji, Latin alphabet, and Arabic Numerals. In addition to orthographic difficulties, variations of expression for the same concept delivered by other reasons such as abbreviations are included in medical texts produced at clinics, hospitals, and other medical institutes. These variations present obstacles that are encountered when developing medical AI or information extraction tools, although several studies have been undertaken to solve these and related difficulties by developing or assisting automatic coding systems (Fabry et al., 2003; Yamada et al., 2010; Bouchet et al., 1998). Actually, several medical resources exist for non-Japanese languages, such as the International Classification of Diseases (ICD; World Health Organization, 2004) 1 , Medical Subject Headings (MeSH; Lipscomb, 2000), and Systematized Nomenclature of Medicine Clinical Terms (SNOMED-C2; Benson, 2012). SNOMED-CT, the largest, includes approximately 308,000 concepts and 777,000 expressions, and officially supports English and Spanish. Also, projects are translating SNOMED-CT into other languages (Abdoune et al., 2011; Zhu et al., 2012). Currently, medical language resources for Japanese language are smaller than those for other major languages 1 http://www.who.int/classifications/icd/en/ 2 http://www.snomed.org/snomed-ct/ …

Content truncated to 500 words