Cross-Subject Generalization and Few-Calibration Learning for Fine Motor Imagery BCIs: Evidence Boundaries, Evaluation Recommendations, and a Research Agenda
Background: Fine motor imagery (FMI) is a demanding paradigm for scalp EEG brain–computer interfaces, yet most cross-subject and few-calibration methods have only been validated on coarse-grained motor imagery, so their transfer to FMI is unproven. Methods: This targeted, evidence-graded review separates direct FMI studies from candidate evidence drawn from upper-limb motor imagery, mixed motor execution/imagery, and coarse-grained tasks. We formalize cross-subject, cross-session, few-shot, zero-calibration, inductive, and transductive settings; assess five candidate method families (data-level alignment, domain adaptation, meta-learning, self-supervised pretraining, and subject-conditioned representation learning); and propose a reproducible leave-one-subject-out evaluation framework with repeated calibration draws, per-subject uncertainty, and explicit hardware metadata. Results: The review shows that the evidence does not permit a ranking of method families for FMI, that no FMI-specific cross-subject benchmark is available in the searched literature although direct individual-finger decoding evidence exists, and that class confusion is a substantial contributor to FMI decoding difficulty, while its dominance over subject shift remains an untested hypothesis. Conclusions: We identify the evidence gaps requiring task-matched, multi-dataset validation before reliable FMI claims can be made and provide a staged research agenda.
Authors
- Yueqi Zhang (ORCID: https://orcid.org/0009-0006-7195-230X)
- Jiaming Chen
- Dan Wang
Institutions
- Beijing University of Technology (CN)
Publication Details
- Journal
- Bioengineering
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/bioengineering13111390
- Primary Topic
- EEG and Brain-Computer Interfaces
- Type
- article
- Field-Weighted Citation Impact
- 0.00