Purpose: This research systematically compares the performance of K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machine (SVM) in recognizing emotional and cognitive states from EEG data in a virtual reality (VR) environment. It aims to identify the model with the highest accuracy for each participant. Methods: EEG data were collected from four channels (TP9, AF7, AF8, TP10) with a data range of 0.0 - 1682.815 µV and a sampling rate of 2 Hz. The sampling rate is shallow compared to the standard EEG datasets. Features extracted included statistical measures (mean, standard deviation, skewness, kurtosis) and Hjorth parameters (activity, mobility, complexity), classifier (SVM, RF, KNN). Each classifier’s performance was evaluated using accuracy, indicating the proportion of correctly classified instances. Result: RF achieved the highest average accuracy but showed more significant variability. SVM demonstrated a high median accuracy with consistent performance, as indicated by a narrow interquartile range (IQR) and few outliers. KNN exhibited the lowest median accuracy and highest variability, suggesting sensitivity to data characteristics and parameters. These findings highlight RF’s potential for consistent performance with careful tuning and SVM’s reliability. Novelty: The research’s novelty lies in its personalized performance analysis, evaluating each model’s accuracy individually for participants. This tailored approach reveals the best-performing model for each person, emphasizing customized machine-learning applications in VR-EEG systems. The study’s detailed, participant-specific evaluation enhances emotion and cognitive state recognition precision, advancing individualized VR therapeutic interventions and cognitive research methodologies.
Read full abstract