Multi-scale fusion networks with enhanced attention mechanisms for auditory attention decoding from EEG
Document Type
Article
Publication Date
2-1-2026
Abstract
Humans have the remarkable ability to focus on a single speaker's voice in a noisy environment, demonstrating the brain's sophisticated auditory attention mechanisms. However, accurately decoding this attention, particularly from complex and variable EEG signals, remains a significant challenge. Current models struggle to capture the dynamic multi-scale features of EEG signals across different cognitive states or tasks, often focusing on local features while neglecting global spatial correlations. To address these challenges, this study introduces a novel multi-scale fusion network integrated with advanced attention mechanisms, designed specifically to enhance the decoding of auditory attention from EEG signals by capturing both local and global multi-scale features while dynamically and adaptively fusing them. The approach combines Convolutional Neural Networks (CNN) with Multi-Head Attention Graph Convolutional Networks (MGCN) to extract multi-scale features. Additionally, an Enhanced Multi-Aspect Attention Block (EMAAB) is introduced to prevent overemphasis on local features, ensuring the capture of global information in dynamic EEG signals. Finally, a Dynamic Fusion Module (DFM) leveraging Kolmogorov–Arnold Network (KAN) layers enhances adaptability to varying tasks. Experiments on the KUL and DTU public databases demonstrate that the proposed model outperforms existing methods in both recognition accuracy and generalization ability. For the 1-second decision window, it achieved peak accuracies of 97.03% on the KUL database and 90.69% on the DTU database under 10-fold cross-validation. Under the more challenging leave-one-subject-out (LOSO) setting, the model still maintained strong performance, reaching 74.66% on KUL and 61.04% on DTU, further confirming its robustness across subjects. Furthermore, this study also presents promising applications for real-time auditory attention recognition systems, such as brain–computer interfaces, hearing aids, and neurofeedback devices.
Keywords
Auditory attention, Dynamic fusion, Electroencephalography (EEG), Enhanced attention mechanisms, Multi-scale feature
Publication Title
Biomedical Signal Processing and Control
ISSN
1746-8094
DOI
10.1016/j.bspc.2025.108703
Recommended Citation
Zhang, Ke; Yan, Shang; Li, Zenan; Wang, Xin; Ji, Junyu; Siu, Alice Kwai Yee; Lu, Yun; Ting, Hua Nong; Cao, Song; Pan, Hongguang; and Chen, Shixiong, "Multi-scale fusion networks with enhanced attention mechanisms for auditory attention decoding from EEG" (2026). Research Publications (2026 to 2030). 1399.
https://knova.um.edu.my/research_publications_2026_2030/1399
Funders
National Key Research and Development Program of China (2022YFE0197500), China Postdoctoral Science Foundation (2024M75 3402), National Natural Science Foundation of China (62401558, 62471422), CPSF (GZC20241849), Science and Technology Planning Project of Shenzen Municipality (JCYJ20230807093819039, JCYJ20230807093701003, JCYJ202103 24143008022), Shenzhen Medical Research Fund (D2402003)
Volume
112
Publisher
Elsevier Ltd