Boolean and ranked information retrieval for biomedical systematic reviewing

Abstract

Evidence-based medicine seeks to base clinical decisions on the best currently available scientific evidence and is becoming accepted practice. A key role is played by systematic reviews, which synthesize the biomedical literature and rely on different information retrieval methods to identify a comprehensive set of relevant studies. With Boolean retrieval, the primary retrieval method in this application domain, relevant documents are often excluded from consideration. Ranked retrieval methods are able to mitigate this problem, but current approaches are either not applicable, or they do not perform as well as the Boolean method.

In this thesis, a ranked retrieval model is identified that is applicable to systematic review search and also effective. The p-norm approach to extended Boolean retrieval, which generalizes the Boolean model but, to some extent, also introduces ranking, is found to have a particularly promising prospect: identifying a greater fraction of relevant studies when typical numbers of documents are reviewed, but also possessing properties important during the query formulation phase and for the overall retrieval process. Moreover, efficient methods available for ranked keyword retrieval models are adapted to extended Boolean models. The query processing methods presented in this thesis result in significant speed ups of a factor of 2 to 9, making this retrieval model an attractive choice in practice. Finally, in support of the retrieval process during the subsequent update of systematic reviews, a query optimization method is devised that makes use of the knowledge about the properties of relevant and irrelevant studies to boost the effectiveness of the search process.