Non-Invasive EEG-Based Classification of Mental States Using Valence-Arousal Mapping from the DEAP Dataset

Abstract

Understanding and classifying human mental states based on electroencephalogram (EEG) signals is essential for the advancement of intelligent and adaptive brain-computer interfaces (BCIs). Mental state classification like stress, relaxation, and arousal, is fundamental to enabling machines to understand and respond to human emotions and cognitive states. BCIs are systems that interpret brain activity to enable interaction with external devices. Their effectiveness depends on accurate interpretation of brain signals like EEG. The motivation behind using EEG to detect mental states in a non-invasive and automated way. This study proposes a non-invasive methodology for mental state classification by utilizing EEG data from the DEAP dataset, combined with a valence-arousal (V-A) emotional model for effective state representation. Preprocessed EEG signals (recorded by 10–10 and 10–20 electrode placement systems) were further segmented into alpha (8–13 Hz) and beta (13–30 Hz) bands and features were extracted based on the Welch method for power spectral density estimation. The data of 30 subjects were chosen and pre-processed an artifact removal in the EEG recordings. With the valence and arousal score of every trial, data was categorized into four emotional conditions. Characteristics based on alpha and beta-band activity were used to train machine learning classifiers to identify these emotional states. Using alpha (8–13 Hz) and beta (13–30 Hz) band features, the proposed system reached an average classification of 78.5% with a Random Forest (RF) classifier, which performed better than SVM (72.3%) and KNN (69.8%). The results illustrate that quadrant-based valence-arousal mapping is efficient in discriminating mental states, with alpha power associated with relaxation and beta power with stress. The results demonstrate that the quadrant-based V-A mapping provides a reliable and interpretable structure for mental state recognition, with alpha power correlating to relaxed states and beta power indicating heightened arousal or stress. Consequently, four mental states are classified into four quadrants on the graph. This work contributes to affective computing by offering a robust, interpretable, and non-invasive method for classifying mental states, with potential applications in personalized BCI systems, mental health assessment, neurofeedback, and emotion-aware computing.