The Riemannian geometry of user learning in MI-BCI: A Cybathlon longitudinal study

Abstract

Motor imagery (MI) brain-computer interfaces (BCIs) are promising assistive technologies, however, their development is hindered by low electroencephalography (EEG) signal quality and an incomplete understanding of neural modulation during training. Traditional performance-based metrics provide limited insight into the mechanisms of skill acquisition. We hypothesize that Riemannian geometry offers a robust framework for analyzing structural and physiological EEG patterns associated with learning. This study analyzes longitudinal EEG data collected during a Cybathlon pilot to investigate how Riemannian features evolve throughout MI-BCI training. Novel metrics are introduced by combining geodesic distances on the Riemannian manifold with cosine similarity between tangent-space vectors, enabling the quantification of neural trajectories during training. The results show that the proposed Riemannian features capture structured longitudinal changes in the covariance representations. In addition, the extracted geometric features revealed recurring patterns that were compatible with two dominant geometric configurations, describing different levels of organization and consistency in the Riemannian feature space. These findings highlight the value of geometric EEG metrics for characterizing cortical adaptation during MI-BCI training and for guiding the development of adaptive training strategies in MI-based BCI systems.