Physiology-informed Gaussian process meta-learning for postprandial glucose trajectory prediction

Advanced wearable sensors have enabled the continuous monitoring of physiological data under the Internet of Medical Things framework, which promotes the development of intelligent prediction technology for healthcare applications. This work concerns the problem of interpretable personalized blood glucose (BG) prediction that is crucial for diabetes management. Although encouraging results of machine-learning (ML) methods have been shown in short-term prediction, the lack of interpretability in models trained on limited data reduces the acceptance for practice use, especially for long-term postprandial BG prediction. To address this problem, a physiology-informed Gaussian process (GP) model with a neural mean function is proposed, in which the prior knowledge of glycometabolism is incorporated into the model optimization process to guarantee interpretable model learning and strengthen the ability of capturing key glucose dynamics. Besides, a model-agnostic meta-learning (MAML) is utilized to enhance the model adaptability for personalized prediction. The effectiveness of the proposed method is evaluated through the comparative results obtained by using the FDA-accepted UVA/Padova T1DM simulator and the clinical datasets. Among all the considered approaches, the proposed approach achieved the lowest RMSE on the in silico dataset, as well as in 3 and 4 h prediction on the clinical dataset. These results indicate that the proposed method has superior performance in predicting long-term postprandial BG trajectories.