Bayesian Inference on Brain-Computer Interfaces via GLASS
提出GLASS贝叶斯模型，通过稀疏时变效应和潜通道分解处理EEG信号低信噪比和空间相关性问题，提升P300脑机接口对渐冻症患者的分类性能，并识别关键脑电通道。
Brain-computer interfaces (BCIs), particularly the P300 BCI, facilitate direct communication between the brain and computers. The fundamental statistical problem in P300 BCIs lies in classifying target and non-target stimuli based on electroencephalogram (EEG) signals. However, the low signal-to-noise ratio (SNR) and complex spatial/temporal correlations of EEG signals present challenges in modeling and computation, especially for individuals with severe physical disabilities—BCI’s primary users. To address these challenges, we introduce a novel Gaussian Latent channel model with Sparse time-varying effects (GLASS) under a Bayesian framework. GLASS is built upon a constrained multinomial logistic regression particularly designed for the imbalanced target and non-target stimuli. The novel latent channel decomposition efficiently alleviates strong spatial correlations between EEG channels, while the soft-thresholded Gaussian process (STGP) prior ensures sparse and smooth time-varying effects. We demonstrate GLASS substantially improves BCI’s performance in participants with amyotrophic lateral sclerosis (ALS) and identifies important EEG channels (PO8, Oz, PO7, and Pz) in parietal and occipital regions that align with existing literature. For broader accessibility, we develop an efficient gradient-based variational inference (GBVI) algorithm for posterior computation and provide a user-friendly Python module available at https://github.com/BangyaoZhao/GLASS.