Browsing by Subject "glucose forecast"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Design and Training Optimization of Neural Networks for Reducing Sensor Location Error in Glucose Forecasting
    (2022-05) Tucker, Aaron
    Diabetes mellitus is a disease in which insufficient blood glucose control leads to chronic hyperglycemia which has devastating side effects such as cardiovascular disease and kidney disease. Control of diabetes relies on accurate and precise monitoring of glucose values. Continuous glucose monitors (CGMs) provide glucose measurements using a small, skin--mounted sensor. Accurate glucose forecasts are desirable since diabetes treatment decisions are made based on measures and estimates of glucose values. Recently, neural networks (NNs) have been used to forecast glucose values using data from CGMs, but current methods do not account for changes in CGM location on the body when producing glucose forecasts despite evidence that location changes can cause variation in glucose estimation. To investigate this consideration, a time--delay feedforward NN was constructed and trained with data from study participants with Type 2 Diabetes who wore CGMs on both arms for 12 weeks. Results of glucose forecasting indicated that changes in CGM location can significantly (p < 0.05) increase prediction error. This indicates that NNs which accurately forecast glucose for CGMs located on the same arm may not be accurate for other locations. A NN with gated recurrent units (GRUs) was examined as a method for reducing location induced error increases. Notably, NNs with GRUs did not produce error increases due to sensor location changes (p < 0.05). However, this NN structure required an increase in computational cost relative to the time--delay feedforward NN. Subsequently, a novel, adaptive, linear quadratic regulator (LQR) controlled learning rate for NN training was implemented which reduced training time by an order of magnitude for the NN with GRUs. These results indicate that a NN with GRUs can effectively forecast glucose for low computational cost while remaining robust to changes in CGM location.