Analysis of Brain Activity in Response to Visual Stimuli Using Electroencephalography: Comparative Study of Neural Network and Machine Learning Classifiers

Electroencephalography (EEG) presents a non-invasive and economical method for examining brain reactions to visual stimuli, demonstrating significant promise for applicable brain–computer interface (BCI) uses. This research examines neural responses triggered by both static and dynamic visual stimuli by utilizing EEG signals collected from 14 healthy subjects with a consumer-grade NeuroSky MindWave device. Fast Fourier Transform (FFT) and Discrete Wavelet Transform (DWT) used as extraction’s feature to capture both frequency and time–frequency-domain information. Four classifiers—Multilayer Perceptron with Backpropagation (MLP-BPNN), Kolmogorov–Arnold Network (KAN), Convolutional Neural Network (CNN), and Support Vector Machine (SVM)—were systematically compared. Performance of each classifier’s method was evaluated through accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC, with robustness ensured by 5-fold cross-validation. Statistical analysis employed one-way ANOVA to examine performance differences across classifiers. Results revealed that MLP-BPNN combined with DWT features achieved the best overall performance (accuracy = 92.14%, F1-score = 92.14%), while SVM obtained the highest recall (91.86%). KAN achieved competitive results but underperformed in minority class detection, whereas SVM exhibited high training accuracy but indications of overfitting. ANOVA confirmed significant differences among classifiers (p < 0.0001), with MLP-BPNN and SVM outperforming CNN, while no significant difference was observed between MLP-BPNN and KAN. These findings demonstrate the feasibility of reliable EEG-based visual stimulus classification using low-cost consumer devices. Additionally, they emphasize the essential function of statistical validation in evaluating classifiers. Future research should extend to larger, more diverse groups and investigate sophisticated deep learning models for improved generalization in BCI applications.