[BibTeX] [RIS]
Blood Glucose Prediction with Convolutional Neural Networks
Type of publication: Mastersthesis
Type: Masters Thesis
Year: 2018
School: TU Darmstadt
Abstract: Type 1 diabetes mellitus is a metabolic disorder in which the body’s own insulin regulation is disturbed. To stay within healthy blood glucose levels, patients have to do this regulation themselves by injecting doses of insulin. To prevent hyperglycemia as well as hypoglycemia, it is helpful to know the further progression of blood glucose. Previous work did this successfully by using support vector regression. Since convolutional neural networks recently achieved extraordinary performance in time series regression tasks, it is reasonable to use them to forecast blood glucose. The integration of discrete events measured by smart devices could further increase the prediction quality. Therefore this work uses convolutional neural networks with shared weights on historical data to predict the blood glucose levels. By this method, it was possible to achieve a mean absolute error in prediction of 8.74 where currently sold solutions lead to an error of 10.1.
Authors Slawik, Albert