科研速览 · Science Skim继续刷下去 · Keep skimming →
◆ IEEE journal of biomedical and health informatics2026-09-10

EEG-Based Auditory Attention Recognition Using DSCANet: An Innovative Incremental Learning Framework.

Ke Zhang, Juanjuan Li, Zenan Li, Shang Yan, Desheng Jia, Yongchao Chen, Xinyue Liang, Xin Wang, Fei Chen, Alice Ky Siu, Hongguang Pan, Shixiong Chen

原始摘要(英文原文)· Original abstract
Auditory attention recognition from electroencephalography (EEG) is fundamental for understanding how the brain selectively processes auditory stimuli in complex acoustic environments. Nevertheless, conventional models usually require retraining the entire network whenever new auditory attention patterns need to be incorporated. This retraining strategy is not only inefficient but also prone to catastrophic forgetting, as previously learned patterns may be overwritten when the model adapts to new information. To address the above challenges, a theoretically supported architecture Dynamic Self-Organizing Convolutional Attention Network (DSCANet) is proposed. The model integrates deep learning with incremental learning mechanisms. The deep learning module is responsible for effectively extracting key EEG features, while the incremental learning module combines the Dynamic Wasserstein Self-Organizing Incremental Neural Network (DWSOINN) with Elastic Weight Consolidation (EWC) to enable continuous adaptation to new data without forgetting previously acquired knowledge. Specifically, DWSOINN autonomousl detects novel patterns in the input and triggers incremental updates at the appropriate time, whereas EWC constrains parameter updates to preserve information from earlier learning phases. Experiments on KUL and DTU show that DSCANet improves classification accuracy and incremental learning stability. The main subject-independent evaluation was conducted under leave-one-subject-out cross validation (LOSO-CV), with additional validation strategies used to assess distributional adaptation and rapid personalization. Specifically, under the 1-second decision window in LOSO-CV, DSCANet achieves 81.51% accuracy on the KUL dataset and 63.03% on the DTU dataset, exceeding the strongest baseline by 3.41 and 6.23 percentage points, respectively. Meanwhile, the model shows stronger stability and resistance to forgetting during incremental learning, which supports its adaptability and robustness in dealing with non-stationary EEG patterns.
读原文 · Read the paper ↗

AI 追问PRO

登录后使用 AI 追问

讨论区

登录后参与讨论

相关论文 · Related

EEG-Based Auditory Attention Recognition Using DSCANet: An Innovative Incremental Learning Framework. — 科研速览 Science Skim