Knowledge graph representation of electronic health records for clinical predictions

In many countries, key clinical and administrative data of the patients are now systematically collected, recorded, and stored in digital formats. These patient-specific medical data are referred to as electronic health records (EHR). EHR data are rich; they capture patient-health care provider interaction at many encounters over time. This systematic digital collection of medical data presents a significant opportunity for developing data-driven technologies for transforming healthcare. Especially for high-stake situations with high uncertainty, such as in intensive care units (ICUs), these systems have the potential to reduce medical errors by assisting health care providers throughout their decision-making process. While EHRs have the potential to bring solutions to diverse problems in the healthcare ecosystem, their use direct in predictive models is not trivial. Among many properties that yield technical challenges in machine learning systems, we address its sparse and heterogeneous nature. In this study, we propose a strategy where one can unify the heterogeneous data types in a knowledge graph framework and learn a dense patient representation that encodes meaningful information from patient EHRs. Our framework employs widely adapted knowledge graph embedding methods and deploys them in different ICU prediction tasks. These tasks comprise mortality prediction and binarized length of stay prediction tasks. We augment the learned patient representation from the knowledge graphs with lab measurements and vital signs. Compared to a state-of-the-art model, the proposed representation achieves superior performance in three of the four different classification tasks.