Title: Theory and Practice for Risk Score Updating – A Report from the Scottish Population

Abstract: Avoiding emergency admissions is advantageous both to individuals’ health and for the overall efficiency of the healthcare system. In this talk, I present an ensemble model which predicts the risk of an emergency admission for each of the 4.3 million people in Scotland, with predictors derived from their electronic health records using a novel application of topic modelling. This builds on an existing simpler model which has been in clinical use since 2012. Motivated by this application we study potential problems in the updating of a predictive score for a binary outcome when an existing predictive score forms part of the standard workflow. In this setting, the existing score induces an additional causative pathway which can lead to problems when the original score is replaced. Moreover, there is no immediate apparent solution, which has significant real-world implications in light of the EU attempts to determine how best to regulate software as a medical device.