Using large clinical corpora for query expansion in text-based cohort identification.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Information Retrieval

Background:

• Clinical text presents challenges like polysemy, synonymy, and hyponymy, complicating patient cohort identification.
• Existing information retrieval (IR) methods struggle with the nuances of clinical language.

Purpose of the Study:

• To investigate if using a large, in-domain clinical corpus for query expansion can improve patient cohort identification.
• To evaluate the effectiveness of auxiliary collections in IR-based cohort retrieval.

Main Methods:

• Utilized four auxiliary collections for query expansion in the Text REtrieval Conference (TREC) task.
• Employed a mixture of relevance models for cohort retrieval from the Pittsburgh NLP Repository.
• Assessed the impact of collection size, query difficulty, and collection interactions on retrieval performance, measured by mean average precision (MAP).

Main Results:

• Performance improved over the baseline query likelihood model (MAP=0.373) with any auxiliary resource (MAP=0.386+).
• Adding the Mayo Clinic collection significantly improved performance (MAP=0.4223 with all four collections).
• Retrieval gains plateaued after including 2.5 billion term instances from the Mayo Clinic collection, indicating diminishing returns and the importance of data curation.

Conclusions:

• Query expansion with a large clinical corpus, specifically the Mayo Clinic corpus, consistently and significantly improves IR-based cohort identification.
• The study highlights that "more data is not necessarily better," emphasizing the value of curated clinical datasets for effective query expansion.
• Access to large clinical corpora is beneficial for IR query expansion tasks in healthcare settings.