Multi-source deep domain adaptation ensemble framework for cross-dataset motor imagery EEG transfer learning

• DOI: 10.1088/1361-6579/ad4e95
• Publication Date: 2024-05-21T22:43:52Z
• Authors: Minmin Miao; Zhong Yang; Zhenzhen Sheng; Baoguo Xu; Wenbin Zhang; Xinmin Cheng
• Keywords: Transfer of learning; Domain Adaptation; Motor Imagery

Abstract Objective . Electroencephalography (EEG) is an important kind of bioelectric signal for measuring physiological activities the brain, and motor imagery (MI) EEG has significant clinical application prospects. Convolutional neural network become a mainstream algorithm MI classification, however lack subject-specific data considerably restricts its decoding accuracy generalization performance. To address this challenge, novel transfer learning (TL) framework using auxiliary dataset to improve classification performance target subject proposed in paper. Approach We developed multi-source deep domain adaptation ensemble (MSDDAEF) cross-dataset decoding. The MSDDAEF comprises three main components: model pre-training, adaptation, ensemble. Moreover, each component, different designs were examined verify robustness MSDDAEF. Main results Bidirectional validation experiments performed on two large public datasets (openBMI GIST). highest average reaches 74.28% when openBMI serves as GIST source dataset. While 69.85% In addition, surpasses several well-established studies state-of-the-art algorithms. Significance study show that TL feasible left/right-hand decoding, further indicate promising solution addressing variability.