DEVELOPMENT AND VALIDATION OF SPATIAL PATTERNS BASED LEARNING AND UNSUPERVISED LEARNING APPROACHES WITH RIEMANNIAN GEOMETRY FOR MOTOR IMAGERY BRAIN COMPUTER INTERFACES

Abstract

The objective of this research is to classify hand movement using EEG signals through two novel approaches, the Bhattacharya-distance based Common Spatial Patterns and unsupervised learning based approach with pseudo labelling on a Riemannian manifold. The latter approach employs a methodology that integrates temporal blocking, frequency band analysis, and Riemannian distance for classification. Specifically, the approach involves the use of pseudo labeling based on Riemannian mean distance to enhance the accuracy of motor imagery (MI) classification tasks. Both the proposed methods were rigorously tested using three distinct datasets: the Clinical BCI (2020), BCI Competition IV 2a. The proposed methods demonstrated remarkable cross-subject classification accuracy, outperforming traditional techniques. The key findings indicate that the integration of unsupervised learning approaches with Riemannian Geometry with temporal and frequency domain analysis provides a robust framework for MI classification. The high cross-subject accuracy achieved suggests that this method holds significant potential for practical applications, particularly for patients suffering from injuries or hemiparetic conditions. The ability to accurately classify motor imagery across different subjects can facilitate the development of more reliable and effective Brain-Computer Interface (BCI) systems for clinical use. In conclusion, this research contributes to the field of BCI by presenting novel, accurate, and cross-subject adaptable methods for motor imagery classification. Future work will focus on refining this approach and exploring its applications in real-world clinical settings, aiming to enhance the quality of life for patients with motor impairments.