TY - GEN
T1 - SSA-HAN
T2 - 2026 2nd International Conference on Artificial Intelligence and Computational Intelligence, AICI 2026
AU - Yuan, Yanxue
AU - Xing, Kai
AU - Yin, Libo
AU - Yang, Hui
AU - Shi, Xin
AU - Zhou, Liping
AU - Chang, Huanhuan
AU - Wang, Ruyuan
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/8/19
Y1 - 2026/8/19
N2 - Sleep apnea severity assessment plays a central role in large-scale screening and longitudinal monitoring, but remains challenging due to pronounced intra-night variability and long-range temporal dependencies that are difficult to model with conventional detect-then-count or flat temporal approaches. In this work, we propose SSA-HAN, a sleep-stage-aware hierarchical attention network that formulates apnea-hypopnea index estimation as a structured whole-night reasoning problem. The proposed framework explicitly integrates sleep-stage context through cross-attention and aggregates information across multiple temporal scales, enabling consistent severity reasoning over long recordings. By jointly modeling fine-grained event evidence and global temporal structure, SSA-HAN produces severity-aware representations that align with clinical assessment principles. Evaluations on a private PPG-based dataset show that SSA-HAN achieves superior performance in both continuous AHI estimation and multi-class severity classification compared with representative baselines, with improved accuracy, agreement, and robustness across clinically relevant thresholds. These results suggest that incorporating sleep-stage awareness and hierarchical temporal aggregation provides an effective and clinically plausible solution for patient-level sleep apnea severity assessment.
AB - Sleep apnea severity assessment plays a central role in large-scale screening and longitudinal monitoring, but remains challenging due to pronounced intra-night variability and long-range temporal dependencies that are difficult to model with conventional detect-then-count or flat temporal approaches. In this work, we propose SSA-HAN, a sleep-stage-aware hierarchical attention network that formulates apnea-hypopnea index estimation as a structured whole-night reasoning problem. The proposed framework explicitly integrates sleep-stage context through cross-attention and aggregates information across multiple temporal scales, enabling consistent severity reasoning over long recordings. By jointly modeling fine-grained event evidence and global temporal structure, SSA-HAN produces severity-aware representations that align with clinical assessment principles. Evaluations on a private PPG-based dataset show that SSA-HAN achieves superior performance in both continuous AHI estimation and multi-class severity classification compared with representative baselines, with improved accuracy, agreement, and robustness across clinically relevant thresholds. These results suggest that incorporating sleep-stage awareness and hierarchical temporal aggregation provides an effective and clinically plausible solution for patient-level sleep apnea severity assessment.
KW - AHI estimation
KW - Hierarchical attention
KW - PPG-based monitoring
KW - Sleep apnea assessment
KW - Sleep-stage-aware learning
KW - Structured temporal modeling
UR - https://www.scopus.com/pages/publications/105048684663
U2 - 10.1145/3810280.3810395
DO - 10.1145/3810280.3810395
M3 - 会议稿件
AN - SCOPUS:105048684663
T3 - Proceedings of 2026 2nd International Conference on Artificial Intelligence and Computational Intelligence, AICI 2026
SP - 761
EP - 769
BT - Proceedings of 2026 2nd International Conference on Artificial Intelligence and Computational Intelligence, AICI 2026
PB - Association for Computing Machinery, Inc
Y2 - 6 February 2026 through 8 February 2026
ER -