An MAML-Based Lightweight Neural Network With Domain Adaptation for Cross-Subject and Generalized EEG-Based Motor Imagery Classification

EEG-based motor imagery (MI) classification remains constrained by severe inter-subject variability, domain shift, and the computational cost of deep models. This study introduces a novel MAML-based lightweight neural network with a dual-branch domain adaptation pipeline that jointly meta-optimizes task-specific and domain-invariant features. The architectural novelty lies in combining depthwise–separable convolutions, SE-attention, and GRL-based domain alignment inside a MAML loop, enabling fast adaptation with 40% fewer parameters than conventional CNN pipelines. Evaluated on BCI-IV 2a and PhysioNet, the model achieves 81.3% cross-subject accuracy, 76.3% cross-task accuracy, and 79.6% four-class performance, outperforming state-of-the-art CNN, ATCNet, WST-CNN, and DB-ATCNet baselines. The lightweight design (0.82M parameters, 0.17G MACs) enables real-time deployment on Jetson Nano at 38 fps, confirming its suitability for portable and edge BCI applications.