Exploring information retrieval approaches for clinical decision support and biomedical search tasks

Carterette, Benjamin A.Health care practitioners seek out relevant information about providing the best possible care for their patients. Relevant information to those practitioners can vary between finding potential diagnoses, suggesting related lab tests, and determining treatments for a given patient. The role of a clinical search system, relying on information retrieval (IR), is to exploit the enormous volume of publications in the biomedical literature and make them efficiently accessible to those practitioners. In this thesis, we propose to develop several approaches which can improve the effectiveness of search results produced by clinical search systems. Our approaches can help in minimizing the overhead of manual settings in current search approaches and reduce the complexity of search tasks performed by physicians. ☐ We propose to design a clinical search system that extends traditional search strategies by incorporating relevant biomedical domain resources. In our system, we will rely on some essential retrieval models (e.g., Divergence From Randomness, Pseudo-Relevance Feedback) and some text reformulation tasks that are based on domain-specific knowledge (e.g., Unified Medical Language System, Medical Subject Headings) in order to improve our search strategy. Moreover, we show that our system can be further enhanced when using simulated user models in evaluating the relevance of medical concepts appearing in clinical search queries. ☐ Lastly, we study the usefulness of adopting learning to rank (LtR) framework for clinical search tasks. We propose a methodology for evaluating several LtR models which includes relying on a diverse set of training features as well as some feature selection techniques and show how this method can lead to learning effective ranking models for clinical search.University of Delaware, Department of Computer and Information SciencesPh.D