Enhancing P300 Feature Extraction through Multi-Scale Separable Convolution

P300, one kind of Event-Related Potentials (ERP), plays a crucial role in Brain-Computer Interfaces (BCIs). P300 detection is a key step in P300-based BCIs. Convolutional Neural Networks (CNNs) have emerged as popular solutions in P300 detection due to their automatic feature extraction capabilities. From 2010 to 2021, various CNN architectures with different layers and convolutional methods have been developed for P300 feature extraction. Among them, SepConv1D, utilizing a single separable convolution layer with only 4 kernels (or 225 parameters), has achieved state-of-the-art performance. This study builds upon SepConv1D's success by proposing a novel approach: multi-scale separable convolution for P300 feature extraction. Unlike SepConv1D, which employs the same size for each convolutional kernel, our approach utilizes different sizes for each kernel. Experimental results demonstrate superior performance, showing improved feature saliency maps and higher AUC values in P300 detection.