Clinical Question Answering using Key-Value Memory Networks and Knowledge Graph

We describe our clinical question answering system implemented for the Text Retrieval Conference (TREC 2016) Clinical Decision Support (CDS) track. We submitted five runs using a combination of knowledge-driven (based on a curated knowledge graph) and deep learning-based (using key-value memory networks) approaches to retrieve relevant biomedical articles for answering generic clinical questions (diagnoses, treatment, and test) for each clinical scenario provided in three forms: notes, descriptions, and summaries. The submitted runs were varied based on the use of notes, descriptions, or summaries in association with different diagnostic inferencing methodologies applied prior to biomedical article retrieval. Evaluation results demonstrate that our systems achieved best or close to best scores for 20% of the topics and better than median scores for 40% of the topics across all participants considering all evaluation measures. Further analysis shows that on average our clinical question answering system performed best with summaries using diagnostic inferencing from the knowledge graph whereas our key-value memory network model with notes consistently outperformed the knowledge graph-based system for notes and descriptions. ∗The author is also affiliated with Worcester Polytechnic Institute (szhao@wpi.edu). †The author is also affiliated with Northwestern University (kathy.lee@eecs.northwestern.edu). ‡The author is also affiliated with Brandeis University (aprakash@brandeis.edu).