Calibration-free Plug-and-Play EEG-based BCIs

Calibration-free plug-and-play operation is critical for real-world EEG-based brain-computer interfaces (BCIs), yet existing methods demand subject-specific calibration or batch processing of target data. To overcome this, we propose Local Online Transfer Learning (LOTL), a novel online transfer learning method enabling effective and immediate response without calibration. Specifically, for each domain (subject), LOTL generates a global hyperplane and multiple local hyperplanes by considering the nonlinear separation problem of samples. During online processing, LOTL attempts to update the current model by retaining the new model close to the current source models and the target model while imposing a margin separation on the latest samples. This is achieved by minimizing the differences between the current and new global and local hyperplanes of the source and target domains, making the learned global and local hyperplanes of the source and target domains transferable. Crucially, LOTL works in a one-pass online manner, using each arriving sample only once without storing any historical samples. We derive a closed-form solution for effective model updates and establish theoretical guarantees mathematically, including mistake bound and convergence analyses. Extensive experiments on four EEG datasets (motor imagery and emotion analysis) and a real-world BCI system provide encouraging evidence for the effectiveness of the proposed method, suggesting the potential of LOTL to facilitate the deployment of BCI applications under conditions where the source and target domains share a consistent channel configuration.