CHIMERA: Dr Andrey Kormilitzin, University of Oxford

Recent years have seen an explosion of computational methods to efficiently learn representations of complex longitudinal electronic health records for downstream analytical tasks. While deep representation learning is the method of choice in many scenario, it requires large amounts of data to achieve robust results. In this talk I will discuss the path signature method, that provides a succinct summary of sequential data and can be used as an effective and learnable feature set for describing sequentially ordered data, such as multidimensional physiological time-series or texts in clinical notes. I will introduce the fundamentals of the paths signature method and discuss their applications in two clinical use-cases: the winning solution to the PhysioNet 2019 Challenge on early prediction of sepsis in ICU and mortality risk prediction for patients diagnosed with Alzheimer's disease using their secondary care medical records.