A Switched Neural Process for Long-Term Postprandial Glucose Trajectory Prediction Based on Insulin Sensitivity Estimation

Advanced sensing technology has enabled the real-time monitoring of blood glucose (BG), making it possible to explore the BG prediction problem based on continuous observations. In this work, we propose a novel switched neural process (SNP) approach for long-term postprandial BG prediction, which has the probabilistic inference ability conditioned on preprandial BG measurements. Distinct representative neural process (NP) models are constructed to capture the nonlinear BG dynamics based on a designed indicator. This indicator is developed to estimate insulin sensitivity (IS) at meal time by using physiological models, in which the effect of preprandial BG variation on IS estimation is evaluated. Besides, a robust model switching mechanism is designed to determine the appropriate NP model for prediction. The effectiveness of the proposed method is illustrated through the comparative results by using both the in silico data from the U.S. Food and Drug Administration (FDA)-accepted UVA/Padova type 1 diabetes mellitus (T1DM) simulator and the clinical data from 47 T1DM subjects. The obtained results indicate that the proposed method outperforms existing methods, especially for 3- and 4-h prediction, which shows promising potential in predicting postprandial BG levels.