Towards Effective Deep Neural Network Approach for Multi-Trial P300-Based Character Recognition in Brain-Computer Interfaces

Brain-computer interfaces (BCIs) enable direct interaction between users and computers by decoding brain signals. This study addresses the challenges of detecting P300 event-related potentials in electroencephalograms (EEGs) and integrating these <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{P} 300$</tex> responses for character spelling, particularly within oddball paradigms characterized by uneven P300 distribution, low target probability, and poor signal-to-noise ratio (SNR). This work proposes a weighted ensemble spatiosequential convolutional neural network (WE-SPSQ-CNN) to improve classification accuracy and SNR by mitigating signal variability for character identification. We evaluated the proposed WE-SPSQ-CNN on dataset II from the BCI Competition III, achieving P300 classification accuracies of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 9. 7 \%}$</tex> for subject <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A$</tex> and 79.9 % for subject <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$B$</tex> across fifteen epochs. For character recognition, the model achieved average accuracies of 76.5 %, 87.5 %, and 94.5 % with five, ten, and fifteen repetitions, respectively. Our proposed model outperformed state-of-theart models in the five-repetition and delivered comparable performance in the ten and fifteen repetitions.