Advancing Assistive BCIs: Ensemble Learning for Robust RGB-Evoked EEG Classification in Complex Visual Paradigms

This study explores a real-time Brain-Computer Interface (BCI) system using RGB-evoked EEG signals to assist individuals with communication disabilities. The study utilized two datasets: Dataset A, comprising data from 21 subjects, in which one color (red, green, or blue) was displayed on a screen at a time while recording EEG signals, and Dataset B, involving 23 subjects, where all three colors were simultaneously displayed on the screen, with a cross indicating the color to focus on. Dataset B thus enables the investigation of increased data complexity, better suited for BCI communication. General, individual-based, and transfer-optimized neural network models were trained for both datasets to assess performance differences across model types. The mean accuracy for the best models based on Dataset A reached 95.5%, utilizing the Sequential Ensemble Learning technique. In Dataset B, the heightened complexity of the protocol led to the top-performing models achieving a mean accuracy of 75.8%, notably lower than that of Dataset A. To address this, the effect of dataset size on Dataset B was explored, with one subject providing five times more data than others. This improved subject-specific performance from 70.6% to 91.3%, indicating that a larger dataset is necessary for neural networks to excel with complex data protocols. Finally, a simplified protocol similar to Dataset B, better suited for real-world application, was tested on the same subject, increasing performance from 91.3+% to 98.8%.