A Multi-Model Fusion Deep Learning Framework for Blood Glucose Prediction from Historical Glucose Signals

Abstract

Regular monitoring of blood glucose levels is essential for those with diabetes mellitus, a chronic condition. Neglecting to maintain a steady blood glucose level may have serious effects for the person. This is why blood glucose (BG) prediction has been the subject of a great deal of research. In order to make an accurate prediction of blood glucose levels, multiple criteria are typically used. To achieve this goal, blood glucose predictions were made using a mix of architectures such as Long Short-Term-Memory (LSTM), WaveNet, and Gated-Recurrent-Units (GRU), as well as combinations of these at the decision-level. Experimental experiments were carried out utilizing the enlarged OhioT1DM data set, which contains the blood glucose history of 12 diabetes patients, and hyper-parameters were chosen to ensure the most efficient operation of these network architectures. In our experimental experiments, we found that combining three networks yielded the lowest RMSE values of 21.90 mg/dl, 29.12 mg/dl, and 35.10 mg/dl over 30-, 45-, and 60-minute prediction horizons (PH), respectively. The results demonstrate that the suggested strategy is highly effective for BG prediction when the resulting RMSE value is compared to state-of-the-art studies. Furthermore, the suggested fusion method offers a fresh outlook for next research on BG prediction in the literature.