Machine learning methods to improve insulin sensitivity forecasting in glycemic control

Abstract

Stress-induced hyperglycemia is a common complication among critically ill patients in the in- tensive care unit. This health condition poses significant dangers, as it can increase the risks of morbidity and mortality. Glycaemic control has been shown beneficial for these patients but hard to achieve safely and consistently. STAR is a patient-specific, model-based glucose control protocol framework successfully providing safe and effective blood glucose control within nor- moglyceic range using insulin and nutrition as treatments. The STAR protocol has undergone clinical validation, a challenging task due to glycemic variability and the need to prevent hypo- glycemic episodes. Currently, the STAR protocol uses a kernel density approach to predict the 5th and 95th percentiles of the insulin sensitivity within 1-3 hourly intervals. Forecasting insulin sensitivity is an important criterion of the STAR protocol to give the best treatment for the patient. Specif- ically, the STAR 3D model leverages the past two values to predict the distributions of insulin sensitivity and recommend the optimal treatment based on predicted risks. However, there is an opportunity to enhance the current model by incorporating machine learning techniques. This thesis explores the application of various machine learning models using quantile regres- sion to forecast these percentiles. The aim is to improve prediction accuracy and subsequently enhance patient safety within the intensive care unit environment. The findings from these machine learning models highlight the challenge of achieving sub- stantial prediction improvements using only two input features. The solution lies in increasing the number of features, thereby enhancing prediction performance and potentially reducing the occurrence of both hypoglycemic and hyperglycemic episodes among patients.