Feature Interpretability in Motor Imagery Brain Computer Interfaces: A Meta-Analysis Across Connectivity, Spatial Filtering, and Riemannian Methods

Introduction: Brain–computer interfaces (BCIs) translate brain activity into commands, enabling applications in communication, control, and neurorehabilitation. A major challenge in noninvasive BCIs is balancing classification performance with interpretability, as many approaches prioritize accuracy while overlooking the neural mechanisms underlying their predictions. Methods: In this study, we conduct a meta-analysis of feature interpretability across widely used methods in motor imagery (MI)-based BCIs, including power spectral density, common spatial patterns (CSP), Riemannian geometry, and functional connectivity. Specifically, we explore how network topology and spatial organization contribute to MI decoding by investigating brain network lateralization. Results: Through evaluations on multiple EEG-based BCI datasets, our results confirm the superior classification performance of CSP and Riemannian methods. However, network lateralization provides stronger neurophysiological plausibility, revealing robust lateralization patterns in sensorimotor and frontal regions contralateral to imagined movements. Discussion: These findings underscore the potential of connectivity-based features as a complementary tool for enhancing interpretability, supporting the development of more transparent and clinically relevant MI-based BCIs.