Speaker: Dr Christopher Yau, University of Birmingham and The Alan Turing Institute
Title: “Probabilistic approaches for optimal sequential feature acquisition”
Abstract: In many real-world data-driven problems, the evidence to support decisions is gathered sequentially and not all measurements are available immediately. For instance, in medical diagnosis, a clinician may order a series of tests and, based on their outcomes, order further tests to determine the disease state of a patient. Each patient disease classification is therefore associated with a “diagnostic trajectory” charting the series of measurements that were recorded to reach their diagnostic conclusion. Whilst much focus in recent developments in medical artificial intelligence have focused on predictive modelling and automation of decision making processes, in the context of complete data, there has been relatively less attention paid to the sequential data acquisition processes that operate in reality. In this talk, I will describe a novel and generic Bayesian optimisation approach that we have developed to integrate sequential feature acquisition processes into predictive models. I will demonstrate the optimality properties of this algorithm and illustrate its use on critical care data from the publicly available MIMIC-III database. Finally I discuss how the framework can be used to construct personal machine learning-based diagnostic tools.