Building Realistic Potential Patient Queries for Medical Information Retrieval Evaluation

To evaluate and improve medical information retrieval, benchmarking data sets need to be created. Few benchmarks have been focusing on patients ’ information needs. There is a need for additional benchmarks to enable research into effective retrieval methods. In this paper we describe the manual creation of patient queries and investigate their automatic generation. This work is conducted in the framework of a medical evaluation campaign, which aims to evaluate and improve technologies to help patients and laypeople access eHealth data. To this end, the campaign is composed of different tasks, including a medical information retrieval (IR) task. Within this IR task, a web crawl of medically related documents, as well as patient queries are provided to participants. The queries are built to represent the potential information needs patients may have while reading their medical report. We start by describing typical types of patients ’ information needs. We then describe how these queries have been manually generated from medical reports for the first two years of the eHealth campaign. We then explore techniques that would enable us to automate the query generation process. This process is particularly challenging, as it requires an understanding of the patients ’ information needs, and of the electronic health records. We describe various approaches to automatically generate potential patient queries from medical reports and describe our future development and evaluation phase