Low-Latency Asynchronous BCI by Detecting the Motion Onset VEP Induced by the Action Observation Stimulus in Senior

Brain–computer interface (BCI) based on a novel action observation (AO) paradigm can simultaneously induce steady-state motion visual evoked potential (SSMVEP) from the occipital area and sensory motor rhythm from the sensorimotor area. This unique double effect has great potentials in stroke rehabilitation. On the other hand, the asynchronous BCIs allow users to switch freely between intentional control (IC) and nonintentional control (NC) states. However, achieving lower latency and higher detecting accuracy are still a challenge. Unlike previous asynchronous SSMVEP BCI studies, which distinguish the IC and NC states by analyzing SSMVEP, the current study identified IC from NC by detecting the motion onset visual evoked potential (mVEP) induced by the designed AO stimulus. As a direct result, significantly lower latency was achieved, while high overall classification accuracy was maintained. The EEG data in 18 seniors responding to the designed AO stimuli were analyzed, temporally and spatially. Principal component analysis (PCA) followed by linear discriminant analysis (LDA) method (
$P+L$
), the multifeatures from discrete wavelet transform (DWT) followed by support vector machine (SVM) method, and the discriminative canonical pattern matching (DCPM) method were investigated to classify the mVEP period and the non-mVEP period in single trial, and the
$P+L$
method was utilized to detect the mVEP. The performance of detecting mVEP in asynchronous mode was compared with the performance of detecting the SSMVEP utilizing the convolutional neural network (CNN). Results showed that the mVEPs elicited by the AO stimulus had a positive deflection (P1) and a negative deflection (N2) within 300 ms of the motion onset. Furthermore, the
$P+L$
method achieved the highest classification accuracy, i.e., 79.688% ± 7.866%, and no significant difference in the classification accuracy between the senior group and young group. The latency to detect the mVEP utilizing the
$P+L$
method was only 266.1 ± 28.6 ms, which was significantly lower than the latency to detect the SSMVEP by utilizing CNN. Overall, the current study provides an approach to significantly improve the feasibility of low latency asynchronous BCI.