With vast amount of biomedical literature available online, doctors have the benefits of consulting the literature before making clinical decisions, but they are facing the daunting task of finding needles in haystacks. In this situation, it would help doctors if an effective clinical decision support system could generate accurate queries and return a manageable size of highly useful articles. Existing studies showed the useful-ness of patients’ diagnosis information in such scenario, but diagnosis is often missing in most cases. Furthermore, existing diagnosis prediction systems mainly focus on predicting a small range of diseases with well-formatted features, and it is still a great challenge to perform large-scale automatic diagnosis predictions based on noisy pa-tient medical records. In this paper, we propose automatic diagnosis prediction meth-ods for enhancing the retrieval in a clinical decision support system, where the predic-tion is based on evidences automatically collected from publicly accessible online knowledge bases such as Wikipedia and Semantic MEDLINE Database (SemMedDB). The assumption is that relevant diseases and their corresponding symptoms co-occur more frequently in these knowledge bases. Our methods perfor-mance was evaluated using test collections from the Clinical Decision Support (CDS) track in TREC 2014, 2015 and 2016. The results show that our best method can au-tomatically predict diagnosis with about 65.56% usefulness, and such predictions can significantly improve the biomedical literatures retrieval. Our methods can generate comparable retrieval results to the state-of-art methods, which utilize much more complicated methods and some manually crafted medical knowledge. One possible future work is to apply these methods in collaboration with real doctors.
