Zero-calibration motor imagery brain-computer interfaces (MI-BCIs) play an important role in clinical rehabilitation, which endeavors to generalize the electroencephalogram (EEG) decoding methods to unseen subjects. Existing models typically involve domain generalization (DG) at a single level and often only focus on the generalization of data marginal and conditional distributions, which severely constrains MI-EEG classification performance during zero-calibration.

To this end, we propose a novel S tratified DG framework via D eep M ixture Di stribution R epresentation (SDG-DMDR) learning. The proposed SDG-DMDR model specifically enhances DG across multiple subjects at both the sample level and deep representation level through centroid alignment and mixture distribution learning techniques. Leveraging the feature extractor and classifier, the proposed model engages in joint optimization by incorporating weighted domain-generalized loss, center losses, and classification loss, which aims to enhance classification performance and generalization capabilities simultaneously.

Empirical experiments on two publicly available MI-EEG datasets have revealed the feasibility and effectiveness of the proposed SDG-DMDR model. We achieved average accuracy and Cohen's kappa value of 68.75 and 0.577 in BCIIV-2a dataset, as well as 84.07 and 0.678 in BCIIV-2b dataset. Results analyses and ablation studies have also verified the superiority of the SDG-DMDR model with parameters insensitivity.

We provide a novel option to build zero-calibration MI-BCIs based on stratified DG framework.