文献类型：期刊文献

英文题名：Prior knowledge-guided and unsupervised domain adaptation enhanced fine-tuning for electroencephalogram classification

作者：Chen, Dingxin[1];Chen, Ning[1];Chen, Yibo[1];Zhu, Hongqing[1];Zhu, Zhiying[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, 130 Meilong Rd, Shanghai 200237, Peoples R China

年份：2026

卷号：181

外文期刊名：ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

收录：;EI(收录号：20262420905891);WOS:【SCI-EXPANDED(收录号:WOS:001800231700001)】；

基金：Acknowledgments This work was supported by the National Natural Science Foundation of China [grant number 61771196, 61872143] . We would like to express our gratitude to the authors of Jiang et al. (2024) for providing the code of LaBraM model, which has been instrumental in our research.

语种：英文

外文关键词：Fine-tuning; Unsupervised domain adaptation; Data augmentation; Foundation model

摘要：Due to the limited amount of pre-training data, electroencephalogram (EEG) foundation models often face challenges under cross-subject and cross-domain EEG distribution shifts. To effectively and efficiently fine-tune pre-trained foundation models for downstream tasks with limited labeled data, this paper presents a Parameter-Efficient Fine-Tuning (PEFT) method that incorporates a neuroscience-inspired selective fine-tuning strategy and an unsupervised domain adaptation strategy. Specifically, we hypothesize that feed-forward networks (FFNs) in higher Transformer layers contain more discriminative task-related representations. Therefore, only higher-layer FFNs are fine-tuned, while most backbone parameters remain frozen. In addition, to address the non-stationarity and cross-subject variability of EEG signals, we introduce a class-aware unsupervised domain adaptation strategy that promotes intra-class consistency and inter-class discrimination. Extensive experimental results on four public datasets demonstrate that, across the evaluated cognitive tasks (emotion recognition, abnormality detection, or event type classification), (i) the proposed model achieves certain improvements in classification performance with fewer training parameters than the correspondingly full fine-tuned foundation model; (ii) it demonstrates certain adaptation performance improvements compared with existing PEFT and domain adaptation methods; (iii) both the proposed neuroscience-inspired selective fine-tuning strategy and the class-aware domain adaptation strategy contribute to performance improvements. The experimental code and supplementary materials are available at https://github.com/wenjiu100/PKGFT-CAUDA.