Decoding multi-channel EEG signals of upper limb motor imagery in riemannian space

Decoding motor imagery (MI) from electroencephalography (EEG) is essen tial for brain-computer interface (BCI) applications. This study investigates the use of movement-related cortical potentials (MRCPs) enhanced with the xDAWN spatial f ilter and classified using Riemannian geometry-based methods: Minimum Distance to Mean (MDM) and Tangent Space Mapping combined with Linear Discriminant Anal ysis (TSM-LDA). Within-subject and cross-subject scenarios were evaluated for both binary (movement versus rest) and multiclass tasks, examining the effects of analysis window length (100 ms–1 s) and frequency bands (0.3–3 Hz and 7–30 Hz). Results indicate that within-subject binary decoding achieved up to 71% accuracy, with peak discriminative activity occurring approximately 1.2–1.3 seconds after cue onset, while cross-subject binary decoding reached around 67% accuracy with peak latency approxi mately 1.3–1.4 seconds. Multiclass decoding achieved 22–24% accuracy within-subject, dropping to approximately 20% in the cross-subject evaluation. Longer temporal win dows consistently improved performance, and xDAWN filtering enhanced the signal-to noise ratio. Confusion matrix analysis revealed that most misclassifications occurred between similar movement types. These findings demonstrate the potential of MRCP based Riemannian decoding for neuroprosthetic control and highlight the importance of advanced feature extraction methods to enable robust, calibration-free BCI systems.