Investigating the Impact of Query Representation on Medical Information Retrieval.

This study investigates the effect that various patient-related information extracted from unstructured clinical notes has on two different tasks, i.e., patient allocation in clinical trials and medical literature retrieval. Specifically, we combine standard and transformer-based methods to extract entities (e.g., drugs, medical problems), disambiguate their meaning (e.g., family history, negations), or expand them with related medical concepts to synthesize diverse query representations. The empirical evaluation showed that certain query representations positively affect retrieval effectiveness for patient allocation in clinical trials, but no statistically significant improvements have been identified in medical literature retrieval. Across the queries, it has been found that removing negated entities using a domain-specific pre-trained transformer model has been more effective than a standard rule-based approach. In addition, our experiments have shown that removing information related to family history can further improve patient allocation in clinical trials.