School of Mechanical Engineering & Automation,Northeastern University,Shenyang 110819,China. Corresponding author: WANG Hong,E-mail: hongwang@mail. neu. edu. cn
Show less
文章历史+
Received
Published
2024-12-15
2026-03-15
Issue Date
2026-05-22
PDF (1358K)
摘要
针对认知负荷跨被试检测的难点,结合黎曼几何方法在脑电(electroencephalography,EEG)信号分析中的优势,提出基于均值最小距离(minimum distance to mean,MDM)和切空间的认知负荷跨被试检测方法.本研究基于60通道EEG信号的多源干扰任务(multi-source interference task,MSIT)数据,经数据预处理后提取事件相关数据段,并采用xDAWN空间滤波器优化EEG协方差特征表示.通过对比传统特征方法,验证所提方法在认知负荷检测任务中的有效性.结果表明,在1.2 s长度的数据窗口中,基于黎曼几何的方法在留一法(leave one subject out,LOSO)跨被试分类中,最优分类性能的受试者工作特征曲线下面积(area under the receiver operating characteristic curve,ROC AUC)达0.790 1;而在被试内(Within)分类中,ROC AUC最高为0.806 2.相较于深度学习模型,所提方法在保持较高分类性能的同时,显著提升了计算效率,为实时监测提供了可行方案.
Abstract
To address the challenge of cross-subject cognitive load detection, by combining the advantages of Riemannian geometry methods in electroencephalography (EEG) signal analysis, a method of cross-subject cognitive load detection based on minimum distance to mean (MDM) and tangent space was proposed. Based on 60-channel EEG signals from the multi-source interference task (MSIT) data, event-related data segments were extracted through data preprocessing, and the xDAWN spatial filter was employed to optimize the representation of the EEG covariance matrix features. Compared with traditional feature-based methods, the effectiveness of the proposed method in cognitive load detection tasks was verified. The results show that, for the data window with the length of 1.2 s, the Riemannian geometry-based method achieves an area under the receiver operating characteristic curve (ROC AUC) of 0.790 1 for the optimum classification performance in leave one subject out (LOSO) cross-subject classification and reaches a maximum ROC AUC of 0.806 2 within subject (Within) classification. Compared to deep learning models, the proposed method maintains high classification performance while significantly improving computational efficiency and providing a feasible solution for real-time monitoring.
以上方法通常假设数据服从欧几里得空间的分布特性,而脑电信号的协方差矩阵却具有非欧几里得性质,这种不匹配进一步限制了模型的性能[7].针对上述问题,黎曼几何方法为跨被试认知负荷检测提供了新的解决思路.黎曼几何方法能够有效处理脑电信号的协方差矩阵,这种矩阵定义在对称正定矩阵(symmetric positive definite,SPD)空间中[8],该空间构成一个黎曼流形而非传统的欧几里得空间.黎曼几何方法通过利用流形空间的内在几何结构,能够更加准确地刻画不同被试之间的脑电信号特性.近年来,基于黎曼几何的分析方法逐渐成为EEG信号处理的一个重要方向,尤其是在运动想象[9-10]和稳态视觉诱发电位[11-12]范式BCI识别的研究中.在黎曼几何框架下,最小均值距离法和切空间映射(tangent space mapping,TSM)是基于黎曼几何的信号处理和分类中常用的方法[13],尤其适用于EEG数据分析.两者虽然都利用了协方差矩阵,但在特征提取和分类策略上有很大不同.MDM方法基于黎曼流形空间中类别均值的几何距离进行分类.通过计算测试样本与各类别均值协方差矩阵之间的黎曼距离,MDM方法避免了对高维特征的显式建模,具有简单直观且计算效率高的特点,尤其适合小样本场景.切空间映射方法则通过将协方差矩阵映射到参考点(通常是流形的均值)处的切空间,将流形上的非线性问题转化为切空间的线性问题.这种方法不仅保留了黎曼流形的几何信息,还能够结合传统的线性分类器(如支持向量机),在复杂分类任务中展现出较强的泛化能力[14].
本文设置了4种测试模型和4个基准模型,如表1所示,包括CSP结合SVM线性分类器[22]以及脑电领域的3种经典的深度学习模型(EEGInception[23],EEGNetv4[24]和ShallowFBCSPNet[25]),作为非黎曼几何方法的基准模型.黎曼几何方法主要为MDM分类器和基于切空间特征的SVM线性分类器.作为对比实验,以验证黎曼几何方法检测认知负荷的分类性能.所有实验统一使用受试者工作特征曲线下面积(area under the receiver operating characteristic curve,ROC AUC)指标来表征模型性能.考虑到类别不平衡问题,使用加权指标进行评估.此外,还通过模型的样本平均耗时对模型进行综合评价.为了验证跨被试分类的有效性,实验中使用MOABB(mother of all BCI benchmarks)框架[26]进行了3种模式的模型评估:跨被试评估使用留一被试法(leave one subject out,LOSO)和留部分被试法(leave part subjects out,LPSO),被试内(within subject,Within)评估使用五折交叉验证.在LPSO中设置20%的比例,即每次有10名被试作为测试集且不重复.此外还进行了几组消融实验:①通过对比MDM和切空间方法是否使用xDAWN滤波,测试xDAWN模块的作用;②xDAWN和CSP空间滤波算法可以配置滤波器数量参数,本文设置了常用的4(CSP和 X 的默认值)和8(模型名称分别标明为CSP8和X8),以比较不同滤波器数量的效果;③截取刺激点后的数据段进行检测,评估模型是否真正依赖任务相关特征进行决策;④基础分类器使用线性SVM,同时对比测试了使用径向基核函数(radial basis function,RBF)的SVM分类器.
ZhouYue-ying, GongPei-liang, WangPeng-pai, et al. Cross-operator cognitive workload recognition based on convolutional neural network and domain generalization[J]. Journal of Electronics & Information Technology, 2023, 45(8): 2796-2805.
[4]
PaasF, TuovinenJ E, TabbersH, et al. Cognitive load measurement as a means to advance cognitive load theory[M]. Routledge: Cognitive Load Theory, 2016: 63-71.
[5]
Jiménez-GuarnerosM, Gómez-GilP. Custom domain adaptation: a new method for cross-subject, EEG-based cognitive load recognition[J]. IEEE Signal Processing Letters, 2020, 27: 750-754.
[6]
RoyY, BanvilleH, AlbuquerqueI, et al. Deep learning-based electroencephalography analysis: a systematic review[J]. Journal of Neural Engineering, 2019, 16(5): 051001.
XiaoXiao-lin, XinFeng-ran, MeiJie, et al. A review of adaptive brain-computer interface research[J]. Journal of Electronics & Information Technology, 2023, 45(7): 2386-2394.
[9]
CongedoM, BarachantA, BhatiaR. Riemannian geometry for EEG-based brain-computer interfaces: a primer and a review[J]. Brain-Computer Interfaces, 2017, 4(3): 155-174.
[10]
ArsignyV, FillardP, PennecX, et al. Geometric means in a novel vector space structure on symmetric positive-definite matrices[J]. SIAM Journal on Matrix Analysis and Applications, 2007, 29(1): 328-347.
[11]
DingX T, YangL, LiC S. Study of MI-BCI classification method based on the Riemannian transform of personalized EEG spatiotemporal features[J]. Mathematical Biosciences and Engineering, 2023, 20(7): 12454-12471.
[12]
GaurP, PachoriR B, WangH, et al. A multi-class EEG-based BCI classification using multivariate empirical mode decomposition based filtering and Riemannian geometry[J]. Expert Systems with Applications, 2018, 95: 201-211.
[13]
MingZ Y, ZhangD Y, LiuS Y, et al. Exploring the feasibility of single-frequency multi-target SSVEP-based BCI for online control[C]//2023 17th International Conference on Complex Medical Engineering (CME). Suzhou, 2024: 1-5.
[14]
ChevallierS, KalungaE K, BarthélemyQ, et al. Review of Riemannian distances and divergences, applied to SSVEP-based BCI[J]. Neuroinformatics, 2021, 19(1): 93-106.
XuHui, HeHong, ZhangHui-min, et al. Transfer learning for motor imagery EEG signals in Riemannian manifold tangent space[J]. Chinese Journal of Biomedical Engineering, 2023, 42(6): 659-667.
[17]
ZhaoC L, DongE Z, TongJ G, et al. Machine learning classification of Riemannian tangent spaces based on MI-BCI[C]//2023 IEEE International Conference on Mechatronics and Automation (ICMA). Harbin, 2023: 807-812.
[18]
GuiguiN, MiolaneN, PennecX. Introduction to Riemannian geometry and geometric statistics: from basic theory to implementation with geomstats[J]. Foundations and Trends in Machine Learning, 2023, 16(3): 329-493.
[19]
RaninenE. Oracle approximating shrinkage covariance matrix estimators for complex elliptical distributions[C]//2024 32nd European Signal Processing Conference (EUSIPCO). Lyon, 2024: 2717-2721.
[20]
ArsignyV, FillardP, PennecX, et al. Log-Euclidean metrics for fast and simple calculus on diffusion tensors[J]. Magnetic Resonance in Medicine, 2006, 56(2): 411-421.
[21]
DzianokP, KublikE. PEARL-Neuro database: EEG, fMRI, health and lifestyle data of middle-aged people at risk of dementia[J]. Scientific Data, 2024, 11: 276.
[22]
CuiY J, XieS Y, XieX Z, et al. LDER: a classification framework based on ERP enhancement in RSVP task[J]. Journal of Neural Engineering, 2023, 20(3): 036029.
[23]
WangH T, WangZ H, SunY, et al. A cascade xDAWN EEGNet structure for unified visual-evoked related potential detection[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2024, 32: 2270-2280.
[24]
WeiQ G, DingX J. Intra- and inter-subject common spatial pattern for reducing calibration effort in MI-based BCI[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023, 31: 904-916.
LiuChong, ZhaoHai-bin, LiChun-sheng, et al. CSP/SVM-based EEG classification of imagined hand movements[J]. Journal of Northeastern University (Natural Science), 2010, 31(8): 1098-1101.
[27]
Santamaria-VazquezE, Martinez-CagigalV, Vaquerizo-VillarF, et al. EEG-inception: a novel deep convolutional neural network for assistive ERP-based brain-computer interfaces[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2020, 28(12): 2773-2782.
[28]
LawhernV J, SolonA J, WaytowichN R, et al. EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces[J]. Journal of Neural Engineering, 2018, 15(5): 056013.
[29]
SchirrmeisterR T, SpringenbergJ T, FiedererL D J, et al. Deep learning with convolutional neural networks for EEG decoding and visualization[J]. Human Brain Mapping, 2017, 38(11): 5391-5420.
[30]
JayaramV, BarachantA. MOABB: trustworthy algorithm benchmarking for BCIs[J]. Journal of Neural Engineering, 2018, 15(6): 066011.
[31]
KhanamF, HossainA B M A, AhmadM. Electroencephalogram-based cognitive load level classification using wavelet decomposition and support vector machine[J]. Brain-Computer Interfaces, 2023, 10(1): 1-15.
[32]
XiaL K, FengY, GuoZ H, et al. MuLHiTA: a novel multiclass classification framework with multibranch LSTM and hierarchical temporal attention for early detection of mental stress[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023, 34(12): 9657-9670.
[33]
Al-ShargieF, TangT B, KiguchiM. Stress assessment based on decision fusion of EEG and fNIRS signals[J]. IEEE Access, 2017, 5: 19889-19896.
[34]
AvvaruS, ProvenzaN R, WidgeA S, et al. Decoding human cognitive control using functional connectivity of local field potentials[C]//2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society(EMBC). Guadalajara, 2021: 451-454.