FB-RCSP-RA: A Filter-Bank Regularized Common Spatial Pattern Framework with Per-Band Riemannian Alignment for Cross-Subject Motor Imagery EEG Decoding

Read full text Summarise with VIRAT
Authors: FOUAD CHOUAG
Cited by: 0
References: 27

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Cross-subject motor imagery (MI) electroencephalography (EEG) decoding remains an open challenge in brain-computer interface (BCI) research, with inter-subject covariance shift being the principal obstacle to zero-calibration deployment. Existing Common Spatial Pattern (CSP) extensions address individual limitations — spectral heterogeneity, covariance ill-conditioning, or domain shift — but no unified framework has simultaneously integrated all three within a principled cross-subject pipeline. This paper proposes FB-RCSP-RA, a novel Filter-Bank Regularized CSP with Per-Band Riemannian Alignment that unifies: (1) filter-bank decomposition across nine overlapping sub-bands to capture the full mu and beta oscillatory spectrum; (2) per-band Ledoit-Wolf shrinkage regularization to ensure well-conditioned covariance estimates; and (3) per-band Riemannian alignment at the population Riemannian mean to normalize inter-subject covariance domain shift before spatial filter estimation. The framework is evaluated under a strict cross-subject Leave-One-Subject-Out (LOSO) protocol on BCI Competition IV Dataset 2a (nine subjects, four MI classes) with Euclidean Alignment preprocessing. FB-RCSP-RA achieves the highest mean LOSO accuracy of 42.98% and the highest within-subject 5-fold CV accuracy of 71.18% among all five evaluated methods, outperforming CSP (38.46%), ACMCSP (38.93%), RCSP (38.89%), and Riemannian MDM (40.70%). The consistent directional advantage across subjects (6/9 per comparison) and medium-to-large effect sizes (Cohen's d = 0.203 to 0.540) suggest genuine performance benefits. The pipeline operates below 40 ms per trial on standard CPU hardware without GPU requirements. All source code is publicly available at https://github.com/fouadchouag/FB-RCSP-RA.

Fields of Study & Subjects

EEG and Brain-Computer Interfaces Muscle activation and electromyography studies Emotion and Mood Recognition

Publication details

Journal / source
Research Square
Publication date
2 April 2026
Language
EN
Document type
Preprint