Transfer learning and clustering analysis of epileptic EEG signals on Riemannian manifold

Due to the nonlinearity and non-stationarity of EEG signals, the manifold becomes a new direction to analyze EEG signals. However, the problem of imbalanced epileptic seizure data and limited labeling data results in insufficient training data. The significant difference in feature distribution between cross-subject EEG signals also affects model generalization. A manifold transfer learning method based on divergence alignment (MTLDA) was proposed to overcome the few-shot problem. By constructing an objective function based on domain divergence, the transformation matrix is obtained to reduce the feature distribution difference between source and target domains. Additionally, for less labeled data on the manifold, a Riemannian manifold fuzzy clustering (RMFC) was proposed to realize unsupervised pattern recognition of the sample on the manifold. MTLDA and RMFC use Riemannian distance and Riemannian mean rather than the classical Euclidean distance and mean. Both methods do not require labeling information from subjects. The experiment results on two public datasets and one private dataset show that MTLDA and RMFC are effective and outperform other comparison algorithms.& COPY; 2023 Elsevier B.V. All rights reserved.

Automated summary

Due to the nonlinearity and non-stationarity of EEG signals, the manifold has become a new direction for analyzing EEG signals. A manifold transfer learning method based on divergence alignment (MTLDA) is proposed to overcome the problem of imbalanced epileptic seizure data and limited labeling data. Additionally, a Riemannian manifold fuzzy clustering (RMFC) method is proposed for unsupervised pattern recognition on the manifold. Both MTLDA and RMFC use Riemannian distance and mean, and show better performance compared to other algorithms.