EUVIMEDEUVIMED entdecken
Crossref

A physiologically inspired dual-stream decoupled interaction fusion architecture for effective utilization of multimodal data in blood glucose prediction

Shuang Wen, Hongru Li, Xiaoyu Sun

Biomedical Signal Processing and Control · 2026-07 · 120 · 110203

DOI: 10.1016/j.bspc.2026.110203 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2026-07. Verfasst von Shuang Wen und 2 weiteren Autor:innen. Die Arbeit nennt 68 Literaturangaben. Gefördert von: Natural Science Foundation of Liaoning Province, Liaoning Province Doctoral Start-up Foundation, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate blood glucose (BG) prediction plays an increasingly important role in improving the daily glycemia management of people with type 1 diabetes (T1D). Multimodal data related to BG fluctuations has been extensively collected, and advanced deep learning methods have been applied to process them aiming to enhance BG prediction performance. However, the heterogeneous effects of exogenous variables (bolus and meals) on BG in terms of temporal and mechanistic aspects hinder these efforts in the effectiveness of multimodal data utilization. Therefore, based on physiological principles, we propose a dual-stream decoupled interaction fusion (DDIF) model to improve the effectiveness of multimodal data utilization from three aspects. Firstly, different variables are fed into separate streams for modality-specific encoding. For the dense series modality of BG variable, one-dimensional convolution and self-attention network are employed to extract local and global temporal contexts of the input BG. For the sparse event modality of exogenous variables, we design the dose-pattern-duration joint encoding strategy with learnable multi-pattern embedding matrix and time-course function to aggregate the personalized multi-dimensional information of exogenous variables. Secondly, we design a variable-specific head cross-attention to address the heterogeneous physiological interactions of different exogenous variables on BG, enabling more accurate estimation of BG changes under different coupling scenarios of these exogenous variables. Thirdly, considering the alternating dominance of internal metabolic regulations and external event effects on BG under different physiological states, we introduce the dynamic gated mechanism to adaptively fuse the contextual information from each modality. We evaluate the DDIF model using both numerical and clinical metrics on the real-world OhioT1DM dataset. Compared to other advanced deep learning-based methods, the proposed DDIF model using fewer exogenous variables produces more reliable BG predictions for T1D patients under free-living conditions.

Förderung

  • Natural Science Foundation of Liaoning Province
  • Liaoning Province Doctoral Start-up Foundation
  • National Natural Science Foundation of China

Zitieren

APA

Wen, S., Li, H., & Sun, X. (2026). A physiologically inspired dual-stream decoupled interaction fusion architecture for effective utilization of multimodal data in blood glucose prediction. Biomedical Signal Processing and Control, 120, 110203. https://doi.org/10.1016/j.bspc.2026.110203

Vancouver

Wen S, Li H, Sun X. A physiologically inspired dual-stream decoupled interaction fusion architecture for effective utilization of multimodal data in blood glucose prediction. Biomedical Signal Processing and Control. 2026;120:110203. doi:10.1016/j.bspc.2026.110203

BibTeX
@article{Wen2026,
  author = {Wen, Shuang and Li, Hongru and Sun, Xiaoyu},
  title = {A physiologically inspired dual-stream decoupled interaction fusion architecture for effective utilization of multimodal data in blood glucose prediction},
  journal = {Biomedical Signal Processing and Control},
  year = {2026},
  volume = {120},
  pages = {110203},
  doi = {10.1016/j.bspc.2026.110203}
}
Kostenlos starten →