Few-Shot EEG Speech Imagery Classification with Hybrid Self-Supervised and Meta-Learning

Abstract

Electroencephalography (EEG)-based Speech Motor Imagery (SMI) classification holds considerable promise for enabling silent, non-invasive communication through brain-computer interfaces (BCIs), yet its practical deployment is critically hindered by pronounced session-to-session variability arising from the inherent non-stationarity of neural signals and by the burdensome calibration procedures that demand large volumes of labeled data for each new recording session. In this paper, we propose a novel hybrid framework that synergistically integrates self-supervised contrastive learning with EEG-specific augmentation strategies and Model-Agnostic Meta-Learning (MAML)-based meta-learning to enable robust cross-session few-shot adaptation for EEG-SMI classification. The self-supervised pre-training module employs temporal jittering, channel-wise dropout, Gaussian noise injection, band-specific frequency perturbation, and time segment masking to learn session-invariant representations from abundant unlabeled multi-session EEG data, while the meta-learning module, initialized with the SSL-pretrained encoder, learns an optimization trajectory that permits rapid adaptation to unseen sessions using only a handful of labeled samples. We conduct comprehensive experiments on three benchmark datasets under strict cross-session evaluation protocols: BCI Competition IV Dataset 2a (motor imagery), KaraOne (imagined speech), and CHISCO (speech imagery). The proposed framework achieves 10-shot classification accuracies of 80.1% on BCI Competition IV-2a (representing a 10.0% absolute improvement over the best baseline MAML at 70.1%), 61.5% on KaraOne, and 67.7% on CHISCO, with all improvements reaching statistical significance (p < 0.01, Wilcoxon signed-rank test). Ablation analyses confirm the complementary and synergistic nature of the SSL and meta-learning components. These results demonstrate that the proposed hybrid SSL and meta-learning framework substantially reduces calibration requirements while maintaining robust cross-session performance, advancing the practical feasibility of EEG-based speech imagery BCIs.

Keywords:

• Meta-learning,
• Self-supervised learning,
• Speech motor imagery