Single-Channel P300 Decomposition Using Detector-Kernel Networks

Event-related potentials (ERPs), an electroencephalogram (EEG) component, are elicited by various sensory, cognitive, and motor events. P300, one of the ERPs, is widely utilized for brain-computer interfaces (BCIs) for its robustness. However, within a single trial, it is still difficult to accurately extract and detect the P300 component. This paper presents an approach to extract time-locked components such as P300 by single-channel signal decomposition. Our proposed method is based on detector-kernel networks, which generate decomposed signals by convolving short waves called kernels to outputs of the detector network. We modify the networks to be suitable for decomposing time-locked components by generating decomposed signals as amplitude-modulated kernels. To validate our decomposition method, we conducted an EEG classification experiment in a P300 speller. Our method improved the accuracy of the classification, showing the effectiveness of the proposed method.