A Study of Brain-Computer Interface Recognition Performance Crossing Action Observation Paradigms

Action observation-based brain-computer interface (AO-BCI) could induce visual motor imagery through biological motion while relying on its movement frequency to stimulate steady-state visual evoked potential. This hybrid BCI with dual-brain-region activation offers significant potential for stroke rehabilitation. Since varying AO paradigms are employed in the rehabilitation of different limb movements, a limited training dataset can compromise recognition performance. Thus, this study tried to investigate the BCI performance crossing different AO paradigms for the first time. Three AO paradigms, each containing four actions, were designed to establish an online BCI system. Task discriminant component analysis was utilized to analyze the online and offline EEG data. Three training schemes were developed to construct spatial filters including target session (TS) data, source session (SS) data, and a combination of both. Results indicated that the paradigm content significantly affected the recognition performance (F=7.65, p=0.0039). The recognition accuracies of the four actions for each AO paradigm were 71.86%, 89.71%, and 82.71%, respectively. Among the three training schemes, the combined TS and SS data approach notably enhanced recognition accuracy for the AO paradigm with poor performance using TS data alone (p=0.0319). This study demonstrated that EEG data from existing AO paradigms can be used to construct training sets for new paradigms. And combining a small amount of data from the new paradigm could improve the recognition performance. Future research should focus on developing data calibration methods specific to cross-AO paradigms to further enhance recognition accuracy. This work will provide valuable insights for advancing AO-BCI applications in rehabilitation.