Multiple models for artificial pancreas predictions identified from free-living condition data: A proof of concept study

The artificial pancreas (AP) is a closed-loop system to automatically regulate the glucose concentration in patients with type 1 diabetes (T1D). Model predictive control (MPC) revealed to be one of the most promising approaches for this control problem. Several MPC algorithms have been tested in clinical trials with satisfactorily results. However, the inter-patient variability characterising T1D patients limits the performance of MPC algorithms synthesized on average models and calls for patient-tailored models. The availability of experimental data on long outpatient trials motivated the study of identification techniques applicable to free-living patient data. Moreover, a detailed data analysis can be used to improve model identification. Considered that the postprandial (PP) glucose control is one of the most critical aspects of glucose regulation, the analysis was focused on the PP period. The intra-day variability is investigated via ANOVA test that highlighted a correlation between PP glucose profiles and different day periods (DPs). A data-driven multiple-model predictor (MMP) based on real-data analysis is proposed in this work. It exploits different identified models on the basis of the knowledge acquired through the data analysis. In particular, the MMP uses three basic models specific of each DP. These models have been identified through the impulse-response technique that achieved promising results in model identification from real-data. The prediction capabilities of the MMP are compared to the performance of a predictor built using a single model identified on a daily subset, showing an improvement in terms of predictions capabilities in the breakfast DP.