Dynamic-static Feature Fusion with Multi-scale Attention for Continuous Blood Glucose Prediction

Accurate continuous blood glucose prediction is an effective and direct method for treating type 2 diabetes mellitus. However, current methods are commonly single-domain single-scale blood glucose prediction models. That is, they only learn time correlations within constant time steps of continuous blood glucose data, to mine fluctuation patterns, which limits the model effectiveness and robustness. To this end, a novel dynamic-static feature fusion with multi-scale attention method (DSfusion) is proposed for accurately predicting continuous blood glucose. Specifically, DSfusion designs the cross-domain complementary augmentation via Fourier cycle transformation for learning inherent frequency features. Then, DSfusion devises the multi-scale dependence aggregation based on the attention mechanism with various scale receptive fields to capture time correlations and feature correlations across different time steps. Meanwhile, DSfusion utilizes the static feature enhancement to integrate multi-domain information, i.e., static and dynamic features, into a unified prediction architecture, which greatly boosts the model performance. Finally, comprehensive experiments are performed on the real-world blood glucose dataset, and the results confirm the excellence of DSfusion.