Blood Glucose Prediction for Type-1 Diabetics using Deep Reinforcement Learning
An accurate prediction of blood glucose levels for individuals affected with type-1 diabetes mellitus helps to regulate blood glucose through specific insulin delivery. In our work, we propose the design of a densely-connected encoder-decoder network in conjunction with Long-Short Term Memory networks. We formulate the blood glucose prediction as a deep reinforcement learning problem and evaluate our results on the OhioT1DM dataset. The OhioT1DM dataset contains blood glucose monitoring records in intervals of 5 minutes over 8 weeks for 12 patients affected with type-1 diabetes mellitus. Prior works aim to predict the blood glucose levels in prediction horizons of 30 and 45 minutes, corresponding to 6 and 9 data points, respectively. Compared to prior work with the best prediction accuracy so far with respect to the mean absolute error, we improve by 18.4% and 22.5% in 30-minute and 45-minute prediction horizons, respectively. Furthermore, for risk assessment in our predictions, we visualize the error and evaluate clinical risk through a surveillance error grid approach.
