Early experience in forecasting blood glucose levels using a delayed and auto-regressive jump neural network

Type 1 diabetes mellitus is a widespread chronic disease that, if not properly treated, can lead to short- and longterm complications. In recent years, continuous glucose monitoring has become very popular among patients since it allows to keep track of glucose levels for 24 hours. Nevertheless, hypo- and hyper-glycemic events are still widely reported, motivating the development of methods that forecast blood glucose levels at given prediction horizons, which usually range from 15 to 30 minutes. However their application, regardless of the approaches adopted, is limited in practice by the fact that they usually need a long training time and other information gathered from the patients. To overcome these issues in this work we present a new neural network that extends the jump network model by introducing time delays from the input to the hidden layer and auto-regressive feedback from the output to the hidden layer. The proposed neural network shows promising results at 15, 20 and 30-minute prediction horizons, which are comparable to or slightly better than the results reported in the literature, although the training period used in this work is considerably shorter.