EUVIMEDEUVIMED entdecken
Crossref

MIHNet-ACSleep: A Multimodal Image-Enhanced Hybrid Network With Adaptive Cross-Attention for Wearable Sleep Stage Classification

Yuxuan Zhao, Haitao He, Ke Xu, Pengwei Zhang, Jiadong Ren

IEEE Internet of Things Journal · 2026-11 · 13 (22) · 55177-55190

DOI: 10.1109/jiot.2026.3733463 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Internet of Things Journal, 2026-11. Verfasst von Yuxuan Zhao und 4 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Innovation Capability Improvement Plan Project of Hebei Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate sleep staging using consumer-grade wearable devices is important for continuous and unobtrusive sleep monitoring. However, limited cardiovascular and motion signals make it difficult to capture complex sleep-related dynamics and effectively integrate complementary modalities. To address these challenges, we propose MIHNet-ACSleep, a multimodal image-enhanced hybrid network with adaptive cross-attention for wearable sleep stage classification. Interbeat interval (IBI) sequences derived from photoplethysmography (PPG) and composite acceleration (CA) sequences derived from triaxial accelerometry (Tri-ACC) are represented in both time-series and image forms. The image representations are constructed using Gramian Angular Field (GAF), Recurrence Plot (RP), and Spectrogram (SPCT) encodings to capture complementary temporal and time–frequency characteristics. A hybrid dual-stream architecture with representation-specific attention and hierarchical residual connections is designed to extract discriminative features, while an adaptive cross-attention module dynamically models intermodal dependencies and reweights multimodal information. On the DREAMED dataset, subject-wise fivefold cross-validation shows that MIHNet-ACSleep achieves an accuracy of 86.4% and a Cohen’s Kappa of 0.755 for four-stage sleep classification. External evaluation on the independent Walch dataset further demonstrates its adaptability across heterogeneous wearable devices. These results indicate that MIHNet-ACSleep provides an accurate and generalizable solution for wearable-based sleep monitoring.

Förderung

  • National Natural Science Foundation of China
  • Innovation Capability Improvement Plan Project of Hebei Province

Zitieren

APA

Zhao, Y., He, H., Xu, K., Zhang, P., & Ren, J. (2026). MIHNet-ACSleep: A Multimodal Image-Enhanced Hybrid Network With Adaptive Cross-Attention for Wearable Sleep Stage Classification. IEEE Internet of Things Journal, 13(22), 55177-55190. https://doi.org/10.1109/jiot.2026.3733463

Vancouver

Zhao Y, He H, Xu K, Zhang P, Ren J. MIHNet-ACSleep: A Multimodal Image-Enhanced Hybrid Network With Adaptive Cross-Attention for Wearable Sleep Stage Classification. IEEE Internet of Things Journal. 2026;13(22):55177-55190. doi:10.1109/jiot.2026.3733463

BibTeX
@article{Zhao2026,
  author = {Zhao, Yuxuan and He, Haitao and Xu, Ke and Zhang, Pengwei and Ren, Jiadong},
  title = {MIHNet-ACSleep: A Multimodal Image-Enhanced Hybrid Network With Adaptive Cross-Attention for Wearable Sleep Stage Classification},
  journal = {IEEE Internet of Things Journal},
  year = {2026},
  volume = {13},
  number = {22},
  pages = {55177--55190},
  doi = {10.1109/jiot.2026.3733463}
}
Kostenlos starten →