Deep Learning Approach for the Precision Medicine Track

The paper describes the system presented by the University of L’Aquila in collaboration with the University of Havana - team named UNIVAQ - to the TREC 2019 Precision Medicine Track. The proposed solution, maps any kind of documents - Scientific Abstract, Clinical trials, and Topics - into a multi-dimensional common general representation. Each document is described by five primitive features. The values of each feature are extracted from the original documents using deep learning and machine learning text processing based techniques. To recognize Genes and Diseases, we have trained our models using the PubTator annotated corpus. Instead, to derive demographics information, we have trained the employed deep learning models using the documents -obtained from the Relevance and Raw judgements of the past edition of TREC Precision Medicine /Clinical Decision Support Track 2018- considered “relevant” or “partially relevant”. The results of the Track clearly show that applying a system (as our) made solely by a tagging based approach to the Precision Medicine task, is not sufficient to achieve the performances gained by other systems presented in the TREC Precision Medicine Track 2019.