Motor imagery electroencephalogram classification based on Riemannian geometry and deep learning framework

Brain-Computer Interface (BCI) technology, as a direct connection between the human brain and computers, offers a novel approach for individuals with physical movement impairments. However, practical applications face challenges in accurately classifying motor imagery intentions across different sessions and individuals. To address this issue, this paper presents a novel approach based on Riemannian geometry and deep learning framework. Firstly, leveraging the advantages of Riemannian geometry, we manipulate the covariance matrix to improve the distribution of features in the Riemannian manifold. To tackle the issue of overfitting caused by high-dimensional features, we introduce the Sparse Marginalized Denoising Autoencoder (SmDAE) algorithm. Through experiments conducted on real BCI datasets, our proposed method achieved a classification accuracy of 78.6%, an improvement of 10.5% over the baseline accuracy. Additionally, with an AUC value of 0.74, our method outperformed all compared methods. This research provides a new solution for addressing cross-session and cross-subject challenges in motor imagery classification for BCI.