EEG features and suitable decoding algorithm of RSVP-based brain-computer interface in continuous scenes

Brain-computer interface (BCI) based on rapid serial visual presentation (RSVP) hold significant value for achieving robust target detection through the integration of human and machine. RSVP in continuous scenes presents video materials and is thus much closer to real-world applications, which greatly exceeds traditional discrete-scene RSVP in terms of practicality. However, the similarities and differences in electroencephalography (EEG) features between continuous and discrete scenes have not yet been clearly clarified. And there is a lack of research on decoding algorithms that are more suitable for continuous scenes, which seriously hinders the development of continuous-scene target detection. To solve these problems, this study designed a comparative experiment based on RSVP paradigm in continuous and discrete scenes. Event-related potential (ERP), event-related spectral perturbation (ERSP), and inter-trial coherence (ITC) were used to investigate EEG features induced by distinct scenes. Further, this study used sliding hierarchical discriminant component analysis (sHDCA), shrinkage discriminative canonical pattern matching (SKDCPM) and attention-based temporal convolutional network (ATCNet) to implement target/non-target trial classification. Consequently, continuous scenes exhibited fewer induced ERP components, a shorter latency of P300, and reduced neural oscillation activities in alpha and beta1 bands over the occipital region within 0~0.2s. As for classification, traditional machine learning algorithms obtained significantly lower accuracy in continuous scenes. While ATCNet achieved the best and same level of accuracy in both scenes, indicating its suitability for decoding continuous-scene RSVP. The results contributed to develop more practical RSVP-BCI target detection systems.