Research on emotion recognition using sparse EEG channels and cross-subject modeling based on CNN-KAN-F2CA model

• Published in: PloS one Vol. 20; no. 5; p. e0322583
• Main Authors: Xiong, Fan, Fan, Mengzhao, Yang, Xu, Wang, Chenxiao, Zhou, Jinli
• Format: Journal Article
• Language: English
• Published: San Francisco Public Library of Science 27.05.2025
• Subjects: Accuracy; Artificial intelligence; Artificial neural networks; Biology and Life Sciences; Brain research; Channels; Classification; Color imagery; Computer and Information Sciences; Datasets; Deep learning; EEG; Electroencephalography; Emotion recognition; Emotions; Engineering and Technology; Fast Fourier transformations; Feature extraction; Fourier transforms; Frequency dependence; Frequency domain analysis; Machine learning; Medicine and Health Sciences; Methods; Neural networks; Physical Sciences; Physiology; Power spectral density; Research and Analysis Methods; Social Sciences; Spatial data; Support vector machines; Wavelet transforms; China
• DOI: 10.1371/journal.pone.0322583
• ISSN: 1932-6203

Emotion recognition plays a significant role in artificial intelligence and human-computer interaction. Electroencephalography (EEG) signals, due to their ability to directly reflect brain activity, have become an essential tool in emotion recognition research. However, the low dimensionality of sparse EEG channel data presents a key challenge in extracting effective features. This paper proposes a sparse channel EEG-based emotion recognition method using the CNN-KAN- F 2 C A network to address the challenges of limited feature extraction and cross-subject variability in emotion recognition. Through a feature mapping strategy, this method maps features such as Differential Entropy (DE), Power Spectral Density (PSD), and Emotion Valence Index (EVI) - Asymmetry Index (ASI) to pseudo-RGB images, effectively integrating both frequency-domain and spatial information from sparse channels, providing multi-dimensional input for CNN feature extraction. By combining the KAN module with a fast Fourier transform-based F 2 C A attention mechanism, the model can effectively fuse frequency-domain and spatial features for accurate classification of complex emotional signals. Experimental results show that the CNN-KAN- F 2 C A model performs comparably to multi-channel models while only using four EEG channels. Through training based on short-time segments, the model effectively reduces the impact of individual differences, significantly improving generalization ability in cross-subject emotion recognition tasks. Extensive experiments on the SEED and DEAP datasets demonstrate the proposed method’s superior performance in emotion classification tasks. In the merged dataset experiments, the accuracy of the SEED three-class task reached 97.985%, while the accuracy for the DEAP four-class task was 91.718%. In the subject-dependent experiment, the average accuracy for the SEED three-class task was 97.45%, and for the DEAP four-class task, it was 89.16%.