Abstract: Aiming at traditional linear discriminant analysis’s limitation of insufficiently exploring hidden information for EEG classification, this study proposes a latent variable-based algorithm. It introduces latent variables, integrates low-rank regularization and structural consistency constraints, and uses Mu rhythm PSD features from two public datasets for validation. The method achieves higher subject-dependent accuracy and more stable cross-subject performance, providing a reference for enhancing MI-BCI effectiveness in complex environments.