Clinical Relation Extraction with Deep Learning

Relations between medical concepts convey meaningful medical knowledge and patients’ health information. Relation extraction on Clinical texts is an important task of information extraction in clinical domain, and is the key step of building medical knowledge graph. In this research, the task of relation extraction is based on the task of concept recognition and is implemented as relation classification by the adoption of a CRF model. The proposed CRF-powered classification model depends on features of context of concepts. To remedy the problem of word sparsity, a deep learning model is applied for features optimization by the employment of auto encoder and sparsity limitation. The proposed model is validated on the data set of I2B2 2010. The experiments give the evidence that the proposed model is effective and the method of features optimization with the deep learning model shows the great potential.