Data-Driven Methods for Daily Glucose Prediction in People with Type 1 Diabetes

Low-error prediction of glucose levels is fundamental for the effective management of Type 1 Diabetes Mellitus (T1D), especially with the variability induced by physical activity. This work performs a detailed comparison of different types of artificial neural networks for the prediction of blood glucose during and after physical exercise. Four neural network architectures were evaluated: two based on static structures with temporal information at the input (classical and convolutional) and two based on recurrent neural networks. Furthermore, for each architecture, two variants were analyzed: one using only historical data from continuous glucose monitor (CGM) and another that incorporates auxiliary signals related to insulin, food intake, and physical activity. The predictive models were trained and validated using the dataset from the The Type 1 Diabetes and Exercise Initiative (T1DEXI) clinical trial. The evaluation of the methods was carried out using predictive performance metrics and clinical relevance using the Clarke Error Grid tool. The results obtained from this analysis indicate that the incorporation of auxiliary signals and the use of long short-term memory (LSTM) recurrent architectures significantly improve the accuracy of the predictions, highlighting their potential in clinical applications and automatic glucose control systems in the management of T1D.