基于跨参考多机制证据轨迹网络的脑电专注状态识别

EEG-Based Attention State Recognition Using a Cross-Reference Multi-Mechanism Evidence Trajectory Network

摘要: 针对短时窗脑电专注状态识别易受频谱估计波动、参考方式差异及重叠窗口信息泄露的问题，本文提出跨参考多机制证据轨迹网络(Cross-Reference Multi-Mechanism Evidence Trajectory Network, CRMET-Net)。该方法分别在原始参考视图与公共平均参考(common average reference, CAR)视图中分别提取相位–幅值耦合(phase-amplitude coupling, PAC)和复相位锁定特征，构建跨参考互补表征，并结合增益不变多锥窗稀疏复图轨迹(Gain-Invariant Multitaper Complex Graph Trajectory, GIMCGT)，同时表征短窗频谱特性与脑区连接关系。训练阶段通过内层交叉验证生成各基分类器的折外预测，并将不同基分类器之间的一致、互补与冲突信息编码为证据序列，利用因果时间卷积网络建模其动态变化。在MEMA放松–专注二分类任务中，CRMET-Net在跨Trial五折交叉验证下取得90.06%的准确率、90.06%的平衡准确率和90.04%的宏平均F1；在留一被试交叉验证(leave-one-subject-out cross-validation, LOSO)下，三项指标分别达到71.46%、71.35%和71.37%。实验结果表明，CRMET-Net具有较好的跨Trial识别性能，并在完全未见被试上表现出一定的迁移潜力。

Abstract: To address the challenges associated with identifying focused states from short-window EEG—specifically susceptibility to spectral estimation fluctuations, variations in referencing schemes, and information leakage across overlapping windows—this paper proposes the Cross-Reference Multi-Mechanism Evidence Trajectory Network (CRMET-Net). This method extracts phase-amplitude coupling (PAC) and complex phase-locking features from both raw reference and common average reference (CAR) views to construct complementary cross-reference representations. It integrates these with Gain-Invariant Multitaper Complex Graph Trajectory (GIMCGT) features to simultaneously characterize short-window spectral properties and connectivity relationships between brain regions. During training, inner-loop cross-validation generates out-of-fold predictions for base classifiers; information regarding consensus, complementarity, and conflict among these classifiers is encoded into evidence sequences, the dynamics of which are modeled using a causal temporal convolutional network. In the MEMA relaxation-versus-focus binary classification task, CRMET-Net achieved an accuracy of 90.06%, a balanced accuracy of 90.06%, and a macro-averaged F1-score of 90.04% under cross-trial five-fold cross-validation; under leave-one-subject-out cross-validation (LOSO), these three metrics reached 71.46%, 71.35%, and 71.37%, respectively. Experimental results demonstrate that CRMET-Net exhibits robust cross-trial recognition performance and shows potential for transferability to entirely unseen subjects.