Brain responses to different action observation paradigms and assessing transferable cross-paradigm decoding.

backgroundAction observation-based brain-computer interface (AO-BCI) is a novel technology for stroke rehabilitation. Since various AO paradigms are employed in the rehabilitation of different limb movements, a limited training dataset can compromise recognition accuracy. Thus, this study aimed to analyze the brain responses to various AO paradigms systematically and assess the BCI recognition performance across AO paradigms for the first time.

methodsOn the basis of a frame rate reduction method, three AO paradigms, each containing four actions, were designed: the right-hand task-driven (RHTD) paradigm, the right-hand non-task-driven (RHNTD) paradigm, and the left-hand task-driven (LHTD) paradigm. EEG data were collected from 19 participants under the visual feedback condition and the soft robotic glove feedback condition. Task discriminative component analysis (TDCA) was utilized to perform online target detection. EEG responses were analyzed. Moreover, four training schemes, including target session (TS) data, source session data, a combination and transfer with least squares transformation, were developed to construct spatial filters in TDCA to assess the transferability of cross-paradigm decoding in AO-BCI.

resultsThe three designed AO paradigms elicited steady state motion visual evoke potentials (SSMVEP) in the occipital cortex and induced event-related desynchronization (ERD) in the sensorimotor cortex. The RHTD paradigm induced stronger ERD than the other two paradigms. In addition to EEG responses, the recognition accuracy of the SSMVEP under the RHNTD paradigm (85.20%) was significantly greater than that under the RHTD paradigm (72.37%) and the LHTD paradigm (77.50%). Furthermore, when combing EEG from the mirrored stimulus and TS data, the recognition accuracy increased to 77.50% (RHTD) and 80% (LHTD) respectively. However, when EEG data from observing the AO paradigm, whether task-driven or not, were directly used for training or combed with TS data, the recognition accuracy decreased.

conclusionThis study revealed that the content of AO paradigms affects the amplitude of the evoked SSMVEP and ERD. Furthermore, EEG data elicited by the mirrored stimulus can be directly utilized in cross-paradigm training to increase recognition accuracy, whereas data from observing paradigms involving task-driven or not require further development of calibration methods to be effectively applied in cross-paradigm decoding.