Representing Outcome-driven Higher-order Dependencies in Graphs of Disease Trajectories

In this era of digital medicine, computational analysis of historical patient data is a foundational approach for generating evidence-based insights into patient care, as well as developing new knowledge surrounding the etiology, risk factors, and progression of health conditions [1, 2]. While each assessment of an individual occurs at a discrete point in time, it is critical to recognize that data collected from these observations are not independent. The nature of human disease and the structure of the healthcare system itself impose temporal dependencies that connect information across an individual's lifetime [3-5]. As a result, appropriately utilizing historical data requires the capability to model not only the incidence of prior events but also the relationships among data over time. To capture these complex interactions between events over time, researchers have widely adopted supervised neural architectures [6, 7]. In contrast to traditional, unsupervised trajectory models such as latent growth curves, group-based trajectory models, and temporal clustering [8, 9], these techniques are designed to directly learn relationships between patterns in sequential data and outcome incidence.