Abstract
Filter Bank Common Spatial Patterns (FBCSP) is widely used as a classical baseline in motor imagery BCI research. Despite its popularity, many published implementations suffer from methodological shortcomings that limit both reproducibility and practical relevance. Typical issues include information leakage caused by feature selection on the full dataset, superficial use of ICA without consistent artifact removal, and evaluation protocols that ignore class imbalance or subject variability. In this work, we present a fully reproducible FBCSP pipeline designed to address these limitations. The approach employs nested stratified cross-validation to ensure unbiased performance estimation, integrates ICA with automatic component rejection to improve signal quality, and reports results using chance-corrected metrics such as Cohen's Kappa. All processing steps are explicitly defined and implemented using standard, open-source tools. Evaluation on the BCI Competition IV-2a dataset demonstrates a mean classification accuracy of 81.8 percent, corresponding to a Cohen's Kappa of 0.757. This represents an improvement of 7.5 percentage points over the original FBCSP baseline. The proposed pipeline provides a robust and transparent reference point for future studies, particularly in neurorehabilitationoriented BCI system development, and all code is released publicly to support independent verification and reuse.