Brain–computer interfaces (BCIs) based on steady-state visually evoked potentials (SSVEPs) are inexpensive and do not require user training. However, the highly personalized reaction to visual stimulation is an obstacle to the wider application of this technique, as it can be ineffective, tiring, or even harmful at certain frequencies. In our experimental study, we proposed a new approach to the selection of optimal frequencies of photostimulation. By using a custom photostimulation device, we covered a frequency range from 5 to 25 Hz with 1 Hz increments, recording the subjects’ brainwave activity (EEG) and analyzing the signal-to-noise ratio (SNR) changes at the corresponding frequencies. The proposed set of SNR-based coefficients and the discomfort index, determined by the ratio of theta and beta rhythms in the EEG signal, enables the automation of obtaining the recommended stimulation frequencies for use in SSVEP-based BCIs.

Brain-computer interfaces (BCIs) provide users with alternative communication channels for interacting with peripheral devices of various purposes and constitute a significant area of research in medicine and neurorehabilitation. Multiple paradigms exist for designing such systems, each with distinct characteristics. One such paradigm is based on Steady-State Visually Evoked Potentials (SSVEP). Its key advantage lies in its minimal user training requirements for BCI operation. However, it is also marked by individual variability in users' responses to photostimulation, which introduces specific challenges. Current research suggests that addressing this variability could be a significant step toward enhancing the accuracy and efficiency of SSVEP-based BCIs. With the progress in machine learning (ML), various methods are increasingly employed to analyze biomedical signals. Novel models and architectures are being proposed for domain-specific applications. This trend extends to BCIs, where ML is applied to classify input data represented as electroencephalogram (EEG) signal epochs. This study aims to investigate the impact of incorporating individual user responses on the classification performance of ML models used in BCI design. Several models previously applied to MI, SSVEP, P300, and other paradigms were tested such as ATCNet, DeepConvNet, EEGNet, EEGNeX, EEG-TCNet, MBEEG_SENet, ShallowConvNet and TCNet_Fusion. The open-access datasets “12JFPM” and “TsinghuaBCI_Benchmark” were utilized, providing experimental data from 45 subjects' in total. Using the EEG signal-to-noise ratio (SNR) as an indicator of response intensity, a frequency-specific compatibility coefficient was determined for each photostimulation frequency. Individual subject profiles were then created, each containing 4 frequencies with the highest and lowest compatibility coefficients, enabling the evaluation of model performance under two contrasting conditions. The results demonstrate that accounting for individual responses positively impact the performance metrics of leading ML models, such as classification accuracy and Cohen's kappa coefficient, particularly for ATCNet and EEG-TCNet. The paired Student's t-test confirmed the statistical significance of these improvements for ATCNet and EEG-TCNet (p < 0,01) and for MBEEG_SENet and TCNet_Fusion (p < 0,05). The averaged values along the main diagonal of the confusion matrices for these models exhibit a consistent positive trend.