Combining short and long-term models for predicting blood glucose concentrations on diabetic patients

This work presents a study about identification techniques applicable to predict the blood glucose concentration (glycaemia) taking into account data of two free-living patients with type 1 diabetes (T1D). The main objective is to have an efficient way to model the inter-patient and intra-day variability to improve the design of control algorithms to be implemented in the context of artificial pancreas (AP). Historically, multiple models have been proposed to attack this kind of problems. The novelty here is new since we propose to generate a family of long and short term models which predictions are selected according to specific conditions closely related with the historical data of the patient. Therefore, a systematic methodology is developed firstly obtaining an average long term prediction model (ALTPM), considering insulin and carbohydrates as inputs and glycaemia as output. Secondly, the ALTPM glycaemia predictions together with historical data allows to obtain statistically an interval long term prediction model (ILTPM), to be able to define the range of blood glucose variations at least for the 70% of the time considered for the experiments. To improve the short term predictions we use a particular approach based on Kalman Filter which helps us to tracking glycaemia every five minutes and estimate two weight parameters. These weights affect the predictions of the ILTM to improve the quality of the estimations which are evaluated through the commonly used Root Mean Square Error (RMSE) metric.