From structural design to model compression: Research progress on lightweight algorithms for motor imagery brain–computer interfaces

Abstract Motor imagery brain–computer interfaces (MI‐BCIs) enable direct brain‐to‐device communication by decoding movement intentions from EEG signals. Although deep learning models have significantly improved decoding accuracy, their high computational complexity and large parameter counts hinder deployment on resource‐constrained devices. This paper systematically reviews lightweight strategies for MI‐BCI decoding algorithms, termed “lightweight motor imagery (LightMI).” It focuses on two main aspects: convolutional neural network (CNN) structural optimization and model compression techniques. At the structural level, it examines efficient convolutional architectures, including depthwise separable, multi‐branch, and dilated convolutions, as well as the integration of attention mechanisms. For model compression, it covers pruning, quantization, knowledge distillation (KD), and low‐rank decomposition. Representative lightweight models are compared in terms of parameter count, computational cost, and classification accuracy. Results indicate that lightweight CNNs show strong potential for MI‐BCIs. However, existing methods still face limitations, such as loss of critical spatiotemporal features in pruning, performance degradation in low‐bit quantization, strong dependence on teacher models in KD, and poor architectural adaptability in low‐rank decomposition. Manual design is also inefficient and hardware‐specific. Future research should prioritize automated neural architecture search, differentiable search, and hardware‐software co‐design to enable practical wearable MI‐BCI systems.