Session Title: Building the Health Knowledge Graph: From Linked Data to Knowledge Graphs to Machine Leaning and back again


Weaving a web of reliable health knowledge has never been more important - as machine learning automation replaces computational health care tasks that used to be performed by hospital staff, the datasets used to train them need to be pristine, trustworthy, reliably updated and unbiased. Today, clinicians still rely heavily on written documents by the FDA/EMA, medical association guidelines or reviews published in reputable journals to make the correct clinical decisions. However, it is nearly impossible for a primary care physician today to stay current with all this documentation, which often results in medical errors and missed opportunities for workflow improvements. Interoperable, linked and FAIR data are supposed to address this problem - however, in spite of occasional successes, the large majority of medical knowledge is still not machine readable. Knowledge graphs are set to play an important role in the future of medical knowledge dissemination - not just as the intermediate engine for improved search but also as sources of data for training machine learning models to extract knowledge from written documents. This talk will describe a few of the challenges - and the solutions - that staff at Elsevier has come across while using an expert curated knowledge graph as the source of training data to extraction of triples from medical literature.