A novel AI-driven EEG images emotion recognition generalized classification model for cross-subject analysis

• Published in: Advanced engineering informatics Vol. 68; p. 103744
• Main Authors: Li, Jingjing, Lee, Ching-Hung, Duan, Dingna, Zhou, Yanhong, Xie, Xueguang, Wan, Xianglong, Liu, Tiange, Li, Danyang, Yu, Hao, Hasan, W.Z.W., Song, Haiqing, Wen, Dong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2025
• Subjects: Cross-subject; Electroencephalography (EEG); Shared architecture; Siamese cluster transformer model with shared architecture; Similarity features; Shared architecture; Electroencephalography (EEG); Cross-subject; Siamese cluster transformer model with shared architecture; Similarity features
• ISSN: 1474-0346
• DOI: 10.1016/j.aei.2025.103744

AI algorithms can effectively perform encoding and decoding analysis on Electroencephalography (EEG) signals, which are widely used in emotion recognition due to their ability to reflect brain activity characteristics. However, the significant inter-individual variability in EEG signals complicates cross-subject analysis, thereby limiting their generalizability. In this work, we propose a novel AI-driven EEG general classification model called the Siamese Cluster Transformer Model with Shared Architecture (SCTMSA). The SCTMSA model includes three primary components: (1) A Siamese subnet with a shared convolutional structure for extracting similarity features. (2) A transformer module with dual-head self-attention to capture global dependencies. (3) Based on traditional hard-assignment k-means clustering, we introduce a learnable soft-assignment mechanism and dynamic centroid updating strategy to achieve adaptive low-dimensional compression of the feature space. We validated the proposed model on three public dataset and one private datasets, with results demonstrating that the SCTMSA model effectively captures characteristic similarities. Experimental analysis and interpretability results show that Gamma and Beta frequency bands are most relevant to emotion recognition, highlighting their significant role in representing cross-subject emotional brain activity. This research offers a new perspective on reducing individual differences in EEG and advancing emotion classification tasks.