A Deep Learning Knowledge Graph Approach to Drug Labelling

Ensuring the accuracy and completeness of drug labels is a labour-intensive and potentially error prone process, as labels contain unstructured text that is not suitable for automated processing. To address this, we have developed a novel deep learning system that uses a bidirectional LSTM model to extract and structure drug information in a knowledge graph-based embedding space. This allows us to evaluate drug label consistency with ground truth knowledge, along with the ability to predict additional drug interactions. Annotated sentences from 7,117 drug labels sentences were used to train the LSTM model and 1,779 were used to test it. The drug entity extraction system was able to correctly detect relevant entities and relations with a F1 score of 91% and 81% respectively. The knowledge graph embedding model was able to identify inconsistent facts with ground truth data in 76% of the cases tested. This demonstrates that there is potential in building a natural language processing system that automatically extracts drug interaction information from drug labels and embeds this structured data into a knowledge graph embedding space to help evaluate drug label accuracy. We note that the accuracy of the system needs to be improved significantly before it can fully automate drug labeling related tasks. Rather such a system could provide best utility within a human-in-the-loop approach, where operators augment model training and evaluation.