Multi-Scale Feature Extraction to Improve P300 Detection in Brain–Computer Interfaces

P300 detection is a difficult task in brain–computer interface (BCI) systems due to the low signal-to-noise ratio (SNR). In BCI systems, P300 waves are generated in electroencephalogram (EEG) signals using various oddball paradigms. Convolutional neural networks (CNNs) have previously shown excellent results for P300 detection compared to different machine learning models. However, current CNN architectures limit P300 detection accuracy because these models usually only extract single-scale features. Aiming to enhance P300 detection accuracy, an inception module-based CNN architecture, namely Inception-CNN, is introduced. Inception-CNN effectively learns discriminative features from both spatial and temporal information to reduce overfitting and computational complexity. Furthermore, it can extract multi-scale features, which effectively improves P300 detection accuracy and increases character spelling accuracy. To analyze the effect of the inception layer, two additional models are proposed: Inception-CNN-S, which uses the inception layer with a spatial convolution layer, and Inception-CNN-T, which uses the inception layer with a temporal convolution layer. The proposed model was evaluated on dataset II of BCI Competition III and dataset IIb of BCI Competition II. The experimental results show that Inception-CNN provides a promising solution for improving the accuracy of P300 detection, with F1 scores of 47.14%, 55.28%, and 78.94% for dataset II of BCI Competition III (Subject A and Subject B) and dataset IIb of BCI Competition II, respectively.