Arousal classification of music emotion based on multi-segment dynamic rhythmic EEG features

To analyze the neural mechanisms of music emotion recognition, this paper proposed a multi-temporal dynamic rhythmic feature fusion method. This study aimed to improve high/low arousal classification in a music video watching task based on the DEAP dataset.To capture arousal's dynamic neuronal features, we devided multiple 5-second time windows (2s overlapped) on Electroencephalogram(EEG) time series. Then, based on the Reassigned Smoothed Pseudo Wigner-Ville Distribution (RSPWVD), we extracted four dynamic rhythmic features: dominant band statistics, band power slope, dynamic band energy, and dominant band transition patterns. These features faciliated to capture EEG rhythm changes over time from various angles, overcoming limitations of traditional static features. For classification, we built a two-layer heterogeneous decision model. The first layer used Support Vector Machine and Random Forest as base classifiers. The second layer employed a Multi-Layer Perceptron as a meta-classifier, merging the base classifiers' outputs. This method achieved an average accuracy of 72% in arousal classification. By ranking electrode classification accuracy, we found that prefrontal and temporal brain regions significantly contribute to arousal classification. Our multi-temporal dynamic rhythmic feature fusion strategy shows clear advantages in capturing arousal response's temporal evolution. It validates the effectiveness of our features and two-layer model for tracking arousal's temporal patterns, offering a "feature extraction-decision mechanism" dual innovation framework for neural arousal recognition.