Metaheuristics for Discovering Favourable Continuous Intravenous Insulin Rate Protocols from Historical Patient Data

摘要

Metaheuristic search algorithms such as particle swarm optimisation algorithm and covariance matrix adaptation evolution strategy are used to discover improved strategies for setting intravenous insulin rates of hospital in-patients with diabetes. We describe an approach combining and extending two existing methods recently reported in the literature: the Glucose Regulation for Intensive Care Patients (GRIP) method, and a favourability metric used for comparing competing strategies using historical medical records. We demonstrate with a dataset of blood glucose level/insulin infusion rate time series records from sixteen patients that new and significantly better insulin infusion strategies than GRIP can be discovered from this data.