Accurate and low clinical risk blood glucose level (BGL) prediction plays an important role in blood glucose management of type 1 diabetes (T1D). Due to the diurnal variation in physiological regulation and the influence of external life events, blood glucose patterns change over time and events within a day. The fact that humans have an inherent biological cycle of approximately one day and an individual has similar responses to the same life events leading to day-to-day similarities in blood glucose trajectories. However, the daily schedules are not entirely same under free-living conditions so that similar glycemic patterns may occur at different periods on each day. Thus, the mutable intra-day and inter-day temporal correlations in multi-day BGL data challenge the BGL prediction. To this end, we propose an inter-temporal dynamic joint learning method, named as iTDJL. Firstly, considering the variability of the association between days in the free-living, we design a discovery strategy based on discrete cross-correlation to dynamically select the highly correlated series slices from previous days. Then, based on the multi-head self-attention mechanism, the intra-day and inter-day attention layers are developed and connected in sequence to capture the correlation relationships between different time spans that varies with input. Finally, the future multi-step predictions are obtained through a multi-layer perception layer. We evaluate the iTDJL model on the real-world OhioT1DM dataset. Compared to other advanced methods, the iTDJL produces more reliable BGL predictions without requiring meal announcement and other lifestyle information from the patients under free-living conditions. Our proposed method may promote the T1D management, such as clinical decision support assisting insulin injection and abnormal glucose warning.
Read full abstract