Graph learning approaches to extract actionable insights from Electronic Health Records

This paper explores the application of machine learning techniques to knowledge graph representations of Electronic Health Records (EHRs). The problem is to extract actionable insights from EHRs to support decision making by medical staff. Healthcare providers look to better understand what influences patient journeys so they can improve individual outcomes as well as identify optimal treatment paths. Treatments are rarely a single event, especially for serious illnesses. The aim is to understand pathways of patient journeys through intensive care treatment by modelling and storing the patient journeys as a knowledge graph of symptoms, tests, treatment plans, and outcomes, together with secondary and following treatment plans and outcomes. We aim to then apply machine learning to the graphs. Knowledge graphs are a type of knowledge representation that can effectively organize and represent enormously large volumes of connected data so that the representation can efficiently be used by advanced applications. A critical element in machine learning on EHR knowledge graphs is to convert the patient journey to a machine learning friendly representation. The topology of the graph must be represented as a vector in a vector space suitable for machine learning. We explore the literature on the use of knowledge graphs for modelling EHR data focusing on Intensive Care Unit (ICU) care. We aim to identify the advantages and disadvantages of the use of embeddings (representation-based inference) compared to graph neural networks (neural network-based inference) for learning the vector representations of patient journeys and the intersection of machine learning on EHR graphs.