Using Deep Learning Towards Biomedical Knowledge Discovery

A vast amount of knowledge exists within biomedical literature, publications, clinical notes and online content. Identifying hidden, interesting or previously unknown biomedical knowledge from free text resources using an automated approach remains an important challenge. Towards this problem, we investigate the use of deep learning methods that have shown significant promise in identifying hidden patterns from large corpus of text in an unsupervised manner. For example, it can deduce that „husband‟ „man‟ + 'woman' = 'wife'. We use the text corpus from MRDEF file in the Unified Medical Language System (UMLS) dataset as training set to discover potential relationships. To evaluate our approach, we cross-verify new relationships against the UMLS MRREL dataset and conduct a manual evaluation from a sample of the non-overlapping set. The algorithm found 32% of new relationships not originally represented in the UMLS. The deep learning methods provide a promising approach in discovering potential new biomedical knowledge from free text.