pipette
ENEnglish

Shared-Encoder Residual Task-Bridging Network for Motor Execution to Motor Imagery EEG Transfer in Brain-Computer Interfaces

P. Wang, Y. Zhu, T. Xie, W. Shen, Y. Zhou, X. Zhang

PreprintUso en el mundo realCódigo disponible

En palabras de los autores

Motor-imagery brain-computer interfaces (MI-BCIs) hold promising application prospects in the fields of limb motor rehabilitation and human-computer interaction. However, the reliability of MI calibration data is limited by the difficulty of verifying participants' engagement during motor imagerywhereas motor execution (ME) provides a more readily supervised source of motor-related information.To investigate the potential of ME-derived knowledge for assisting MI decoding, we propose a Shared-Encoder Residual Task-Bridge Network (SE-RTBNet) to achieve accurate intra-subject cross-task knowledge transfer from ME to MI. The proposed framework consists of two major components: a shared encoder and a residual task-bridge module. The shared encoder extracts stable general motor representations from multi-band EEG signals and explores the shared feature patterns underlying ME and MI tasks. The residual task-bridge module is equipped with a performance-aware routing mechanism. It adaptively compensates for task-specific discrepancies between the two tasks, enabling precise cross-task feature alignment and knowledge transfer. To validate the effectiveness of the proposed model, experiments are conducted on three public datasets. The proposed method achieves average classification accuracies of 69.39%, 74.61%, and 98.94%, respectively, and outperforms existing transfer-learning algorithms in overall performance. Our model provides an effective strategy for reducing dependence on extensive MI calibration data through intra-subject cross-task knowledge transfer.The source code of the proposed SE-RTBNet is publicly available at https://github.com/Daisyzyx2002/SE-RTBNet

Resultado principalEl resumen no menciona limitaciones.

Apareció: jueves, 19 de noviembre. bioRxiv. Preprint, todavía sin revisión por pares.

DOI: 10.64898/2026.11.17.767217