Recommendation of Clinical Diagnostic Literature based on Random Forest Model and Query Expansion

Abstract

The discovery of valuable content from mass medical literature to help clinicians in decision-making of diagnosis has become an important research area. Since 2014, it has been included in TREC Track for 5 consecutive years. The aim of this study is to make a new exploration in this research area based on a new method combining random forest model and query expansion. With the MeSH lexicon and automatically constructed acronyms dictionary, full relation of keywords and corresponding articles can be built on three levels including sentences, paragraphs and documents. Based on obtained relation of keywords and articles, many similarities between topics and articles can be calculated, which include cosine similarity, Jaccard similarity, Pearson correlation and Euclidean distance. For each article, PageRank weight and Authority weight of HITS can be also calculated by the citation network in the collection of literature. Based on these four similarities and two weights of citation relationship, the experiment of this paper shows that the effect of medical literature recommendation can be improved by re-ordering the result of query expansion using random forest methods.