MEF-net: A generalized hybrid multi-entropy feature fusion neural network model for cross-subject EEG-based emotion recognition

Cross-subject emotion recognition based on multi-channel Electroencephalography (EEG) signals remains a significant challenge due to complexity, non-linearity, high dimensionality, temporal variability, and individual differences. This study proposes a generalized hybrid multi-entropy feature fusion neural network, MEF-Net, which first integrates five distinct entropy-based features: Differential Entropy (DE), Kullback–Leibler Divergence Entropy (KL), Cross Entropy (CS), Conditional Entropy (CE), and Joint Entropy (JE), to represent emotional dynamics from various statistical and information-theoretic perspectives. Then, each feature is fed into a dedicated branch of an Enhanced Depthwise Parallel Convolutional Neural Network (EDPCNN), specifically designed to exploit the spatial-frequency characteristics of multi-channel EEG signals. Finally, an ensemble voting method is applied to aggregate the classification outputs from the five branches for enhancing cross-subject classification performance. Extensive evaluations across three benchmark EEG emotional datasets (DEAP, SEED, and MAHNOB-HCI) demonstrate that MEF-Net outperforms existing methods, achieving classification accuracies of up to 98.58%, 99.82%, and 98.10% for two-class, three-class, and four-class tasks, respectively. Furthermore, ablation experiments show an accuracy improvement of 7.35%, indicating that MEF-Net performs superior cross-subject emotion recognition by effectively analyzing the multi-dimensional characteristics of EEG signals through complementary entropy fusion and a parallel deep learning architecture. Consequently, the proposed method offers a generalized solution applicable to cross-subject emotion recognition and establishes a novel direction for analyzing EEG signals in the field.