Semantic and Machine Learning Methods for Mining Connections in the UMLS

UMLS 中挖掘连接的语义和机器学习方法

• 批准号: 7299922
• 负责人: JAMES J CIMINO
• 金额: $ 18.11万
• 依托单位: COLUMBIA UNIVERSITY HEALTH SCIENCES
• 依托单位国家: 美国
• 财政年份: 2007
• 资助国家: 美国
• 起止时间: 2007-09-30 至 2009-09-29
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/7299922
• 关键词: Algorithms; Communities; Complex; Data; Data Set; Dependency; Disease; Graph; Healthcare; Information Retrieval; Knowledge; Language; Learning; Literature; Machine Learning; Manuals; Maps; Medical; Methods; Mining; Nature; Ontology; Organism; Pathway interactions; Property; Proxy; Relative (related person); Research; Resources; Retrieval; Sampling; Semantics; Social Network; Solutions; Source; System; Techniques; Terminology; Training; Translating; Translations; Unified Medical Language System; Work; base; biomedical resource; concept; interest; metathesaurus; microbial alkaline proteinase inhibitor; novel; success

DESCRIPTION (provided by applicant): The Unified Medical Language System (UMLS) is an invaluable resource for the biomedical community. One of the intended uses of the UMLS Metathesaurus is to support the translation of terms from a source terminology into terms in a target terminology. It is evident from the research literature on the UMLS that users generally need to perform more broader types of "translations" that involve finding terms with closest meaning to source term (mapping), finding terms that are related to source term and can serve as proxy for various functions (e.g. information retrieval, knowledge discovery) or finding target terms that satisfy some structural or semantic constraint (e.g. information theoretic distance). The methods for finding such "translations" or connections between terms in Meta (other than the case of one-to-one synonymy) are not at all clear. Previous attempts to exploit such connections have depended on either manual selection of relevant connections, or problem-specific algorithms that use expert knowledge about the relative suitability of various inter-concept relationships. We believe that machine learning techniques offer automated, generalizable approaches that are appropriate for use with the UMLS, given the large set of potential connections and the need for a problem-independent approach. We hypothesize that learning strategies that exploit the relational features, scale free properties and probabilistic dependencies of connections in the UMLS will identify meaningful inter-term relationships and that a combined approach will perform better across different problem domains when compared to any of the approaches in isolation. We will evaluate the proposed learning algorithms with training connections from a variety of problem domains in biomedicine. We will disseminate the successful algorithms via the UMLS Knowledge Source API toolkit for mining and visualizing the connections. We believe that the UMLS provides a unique fertile ground to develop novel semantic relational mining methods and advance our understanding of mining large biomedical concept graphs.

描述（由申请人提供）：