<![CDATA[Motor imagery (MI) electroencephalography (EEG) is a foundational paradigm for non-invasive brain–computer interfaces (BCIs). However, its practical adoption is constrained by time-consuming per-user calibration and limited cross-subject generalization. This study evaluates a calibration-light MI-BCI framework that combines self-supervised masked EEG pretraining with a lightweight Conformer fine-tuning model. Experiments were conducted on BCI Competition IV Dataset 2b using only the labeled sessions 01T–03T, with artifact-annotated trials removed according to the official 1023 markers. Three deployment-relevant settings were examined: within-subject evaluation (01T–02T → 03T), strict leave-one-subject-out (LOSO) evaluation, and few-shot adaptation with k = 1/5/10 trials per class from the held-out subject’s screening sessions. Full within-subject benchmarking included CSP+LDA, EEGNet, DeepConvNet, ShallowFBCSPNet, supervised Conformer, and SSL+Conformer, while the subject-independent and few-shot analyses focused on CSP+LDA, EEGNet, supervised Conformer, and SSL+Conformer. In the fully calibrated setting, the best mean accuracy was obtained by ShallowFBCSPNet (62.23% ± 14.16%), whereas SSL+Conformer achieved 54.85% ± 11.15% and slightly outperformed the supervised Conformer (53.56% ± 8.81%). Under strict LOSO, EEGNet achieved the highest mean accuracy (52.92% ± 8.25%), while SSL+Conformer reached 51.56% ± 7.18%. In few-shot adaptation, SSL+Conformer achieved the highest mean accuracy at k = 10 (52.84% ± 7.64%) among the core calibration-light methods. The proposed model had a size of 0.1329 MB, a median CPU latency of 0.8777 ms/trial, and LOSO calibration values of ECE = 0.0630 and Brier = 0.4995. These results indicate that masked EEG pretraining provides a competitive lightweight baseline and is most useful when a modest amount of target-subject calibration data is available.]]>
Copyrights © 2025
