Transfer Learning for Motor Imagery EEG Signals in Riemannian Manifold Tangent Space
Xu Hui1, He Hong1*, Zhang Huiming1, Zhang Li2
1(School of Health Science and Engineering, University of Shanghai for Science Technology, Shanghai 200093, China) 2(Department of Geriatric Neurology, Affiliated Brain Hospital of Nanjing Medical University, Nanjing 210029, China)
Abstract：Aiming at the few-shot problem of motor imagery EEG signals, most of the existing analysis methods are based on Riemannian manifolds design classifiers after alignment in the manifold, or classify in Euclidean space after tangent space projection. It is complicated to design classifier operation in manifold, and direct classification in tangent space will lead to errors due to the inconsistency of tangent space distribution between the source and target domain. Therefore, a transfer learning algorithm based on Riemannian manifold was proposed in this study to classify motor imagery EEG signals. By calculating the covariance matrices of the source and target domain and obtaining their Riemannian mean as the tangent points, the covariance matrices were respectively mapped into the tangent spaces, and then the two tangent spaces were aligned by using a set of common feature bases to complete transfer learning. Three motor imagery datasets with 7, 9 and 9 subjects were used for validation, and there were 300, 144 and 120 samples in the three datasets, respectively. The performance of the algorithm was evaluated using classification accuracy, data distribution map and statistical methods. The average classification accuracy of the proposed algorithm on the three data sets is 81.45%, 77.14% and 66.94% respectively, which increased by 25.05%, 17.69% and 10.98% higher compared with that of the model without transfer learning. The Mann-Whitney U test verified that the difference between the classification results under the two models was statistically significant. Compared with other four comparison algorithms, the performance of the proposed algorithm was significantly different, which showed the superiority of the proposed algorithm. The proposed algorithm can effectively reduce the distribution difference of data between different domains, improve the classification accuracy of cross-subject data, and achieve the expansion of few-shot data to a certain extent.
[1] Wan Zitong, Yang Rui, Huang Mengjie, et al. A review on transfer learning in EEG signal analysis [J]. Neurocomputing, 2021 (421):1-14. [2] Soleimani E, Nazerfard E. Cross-subject transfer learning in human activity recognition systems using generative adversarial networks [J]. Neurocomputing, 2021, (426):26-34. [3] Kang H, Nam Y, Choi S. Composite common spatial pattern for subject-to-subject transfer [J]. IEEE Signal Proc Let, 2009, 16(8):683-686. [4] Azab AM, Ahmadi H, Mihaylova L, et al. Dynamic time warping-based transfer learning for improving common spatial patterns in brain-computer interface [J]. J Neural Eng, 2020,17(1):016061. [5] Barachant A,Congedo M, Jutten C, et al. Classification of covariance matrices using a Riemannian-Based kernel for BCI applications [J]. Neurocomputing, 2013, 112(10):172-178. [6] Zhou Jian, Chu Shujie, Li Xin, et al. An EEG emotion recognition method based on transfer learning and echo state network for HilCPS [J]. Microprocess Microsy, 2020, (1):103381. [7] Fernando B, Habrard A, Sebban M, et al. Unsupervised visual domain adaptation using subspace alignment [J]. IEEE ICCV, 2013:2960-2967. [8] Chiang KJ, Wei CS, Nakanishi M, et al. Cross-subject transfer learning improves the practicality of real-world applications of brain-computer interfaces [C]//2019 9th International IEEE/EMBS Conference on Neural Engineering (NER). San Francisc: IEEE, 2019: 424-427. [9] 谢小峰. 基于协方差特征的EEG解码及其在运动想象脑机接口系统的应用研究 [D]. 广州:华南理工大学,2018. [10] Zanini P, Congedo M, Jutten C, et al. Transfer learning:a Riemannian geometry framework with applications to brain computer interfaces [J]. IEEE Trans Biomed Eng, 2018, 65(5):1107-1116. [11] Rodrigues P, Jutten C, Congedo M. Riemannian procrustes analysis:transfer learning for brain-computer interfaces [J]. IEEE Trans Biomed Eng, 2019, 66(8):2390-2401. [12] Gong Boqing, Shi Yuan, Sha Fei, et al. Geodesic flow kernel for unsupervised domain adaptation [J]. 2012 IEEE CVPR, 2012: 2066-2073. [13] Barachant A, Bonnet S, Congedo M, et al. Multiclass brain-computer interface classification by Riemannian geometry [J]. IEEE Trans Biomed Eng, 2012, 59(4):920-928. [14] Shariat A, Zarei A, Karvigh S A, et al. Automatic detection of epileptic seizures using Riemannian geometry from scalp EEG recordings [J]. Med Biol Eng Comput, 2021, 1(15):1431-1445. [15] Gupta AK, Nagar DK, Matrix Variate Distributions [M]// London:Chapman and Hall/CRC, 2018. [16] Smith ST. Covariance, subspace and intrinsic Cramer-rao bounds [J]. IEEE Trans Signal Process2005, 53(5):1610-1630. [17] Forstner W, Moonen B, Geodesy. The challenge of the 3rd millennium [J]. Berlin, Heidelberg: Springer, 2003, :299-309. [18] Jeuris B, Vandebril R, Vandereycken B, et al. A survey and comparison of contemporary algorithms for computing the matrix geometric mean [J]. Electron Trans Numer Ana, 2012 (39):379-402. [19] Blankertz B, Tomioka R, Lemm S, et al. Optimizing spatial filters for robust EEG single-trial analysis [J]. IEEE Signal Process Mag, 2007, 25(1):41-56. [20] He He, Wu Dongrui. Transfer learning for brain-computer interfaces: a Euclidean space data alignment approach. [J]. IEEE Trans Biomed Eng, 2019,67(2):339-410. [21] 郑捷文,张悦舟,兰珂,等. 基于心率变异性分析的睡眠分期算法研究和验证 [J]. 中国生物医学工程学报, 2020, 39(4): 432-439. [22] 鲁柯柯,祁霞,张建保,等. 睡眠呼吸暂停综合征患者的睡眠分期算法研究 [J]. 中国生物医学工程学报, 2022, 41(3): 273-281.
