A Graph-Centric Neuro-Symbolic Architecture Applied to Personalized Sepsis Treatments

Recent research on intelligent healthcare employs Deep Reinforcement Learning (DRL) to personalize treatments according to patients’ physiological characteristics and thus render treatments more effective. However, the majority of approaches rely on the relational data model, that struggles with the representation of the complex relationships within medical data. Moreover, the output of these approaches is typically a recommended action, e.g., a dosage; however, clinicians need the contextualization, i.e., the provision of supporting information, of such recommendations in order to decide whether to follow it.We present a neuro-symbolic architecture for personalized treatments based on a graph-centric foundation. The architecture is based on representing medical data as a knowledge graph and learning via graph neural networks; their combination enables the inherent capturing of relationships and their native integration into reasoning, which may thus render recommendations more effective. Moreover, the architecture employs formally specified graph queries over the knowledge graph to contextualize personalized treatments. We exemplify the architecture by an application to sepsis treatments and based on a widely-used medical dataset.