Benefits of Being Bayesian: Motor Imagery Electroencephalogram Classification

Abstract

Brain-computer interfaces (BCI) based on electroencephalogram (EEG) pattern recognition have demonstrated clinical value in motor rehabilitation and neuroprosthetics, yet accurate classification of EEG signals remains challenging due to poor signal-to-noise ratio and intrinsic nonstationarity. Bayesian machine learning offers a principled framework for uncertainty quantification that may address these limitations, though prior comparisons with frequentist methods have been limited to informal evaluations on single datasets. This thesis presents a large-scale meta-analysis of Bayesian versus frequentist machine learning for motor imagery EEG classification across 20 benchmark datasets. Six top-performing frequentist pipelines spanning spatial filtering, Riemannian geometry, and deep learning approaches were benchmarked against pairwise Bayesian variants. Pooled effects were estimated using a three-level meta-analytic model accounting for heterogeneity both within and between datasets. Results yielded statistically significant improvements in calibration metrics — expected calibration error and maximum calibration error — for Bayesian models, while discrimination metrics including AUROC and MCC showed no significant difference. Computational cost analysis further suggests that Bayesian methods are economically viable for future BCI research. These findings motivate a hypothesis that the primary advantage of Bayesian learning lies in long-term adaptability through sequential model updating and transferability across subjects, with probabilistic resolution and reliability proposed as key indicators for sustaining high-performing BCI systems.