.png){kind=icon}
.png){kind=icon}
Return
An adaptive decoupling learning system informed by the brain functional structure for EEG decoding
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.108228.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Neuroscientific investigations have revealed the presence of regional pathway connections among functional brain areas, as well as the asymmetrical structural characteristics of the left and right hemispheres. These connections, along with their potential coupling relationships and strengths, significantly influence the representation of neuronal signals as recorded by electroencephalography (EEG) within the cerebral cortex. Therefore, there is an urgent necessity to develop data-driven approaches that can effectively decode latent feature representations from EEG data. In this regard, the current study presents a functional-structural adaptive decoupling learning framework (FS-AD), which is informed by cognitive insights into the functional structure of the brain and integrates local-global spatial representations to decode EEG patterns across various states. To accomplish this, we initially implemented a one-dimensional separable convolution module and designed a local-domain attention interaction layer to extract inter-channel interaction information for each region, thereby enabling the capture of fully connected regional pathways. Following this, we developed a global-local kernel-level fusion decoder (GKFD) to amalgamate multiple local-domain features and decode them through a global-domain connection layer. Furthermore, a cross-domain adaptive fusion decoder (CAFD) was meticulously crafted to dynamically identify the fully connected optimal cross-domain pathway and decode it via a local-domain connection layer. The primary aim of FS-AD is to excavate the connectivity patterns of different brain states to enhance the efficiency of EEG decoding. The results indicate that the proposed FS-AD learning system significantly surpasses existing competitive methods in EEG decoding tasks related to various brain states, including fatigue, emotion, and motor imagery. Importantly, this study elucidates the variations in coupling strength among brain regional pathway connections and their representation of brain activity, while also investigating the optimal regional pathways under distinct brain states. This study contributes to the advancement of universal brain decoding methodologies.
Journal
IF:
6.3
Papers:
7.8K
Citations:
3.0W