A relevance feedback-based system for biomedical literature search

There exist many online databases such as IEEE Xplore and PubMed. Their search engines produce a list of publications that satisfy user´s search criteria specified primarily in keywords. A common problem is that the list can be (very) long if the user fails to use specific keywords in the criteria. With this general problem in mind, we attempted to establish the feasibility of a new relevance feedback-based system for biomedical literature search. The system utilized techniques in relevance feedback and text mining and employed Unified Medical Language System (UMLS). We used keyword “dobutamine”, name of a popular drug suggested by the two physicians on the team, and the criterion “recent 20 years” to form a query to search PubMed. The citations returned by PubMed led us to download 1,099 full-text papers in PDF format. We used these full-text publications to set up a database. The two physicians used it to independently evaluate the feedback relevance approach by using keywords “30 day” and “mortality.” In the first round, they received the same 10 publications from the system. In each of the subsequent rounds (there were total three rounds), the system generated publications according to the relevance feedback of the previous round. We used the precision to measure the system performance. For physician #1, the precisions for rounds 1, 2 and 3 were 0.3, 0.4, and 0.6, respectively. For physician #2, the respective precisions were 0.1, 0.5, and 0.8. In either case, the precision improved over rounds. These preliminary results are encouraging.