BCIwiki (bciwiki.com) — The review was published in Sensors on August 11, 2026, with authors Muhammad Ahmed Abbasi, Hafza Faiza Abbasi, Muhammad Arsalan, Danish Khan, Andres Annuk, and Xiaojun Yu. Motor imagery (MI) classification serves as the backbone to brain-computer interfaces by strengthening the communication bridge between the human brain and external peripheral devices, with two decades of progress spanning medical, gaming, and robotic-control applications. The authors note that early MI classification relied primarily on classical signal processing techniques that were heavily affected by signal variation, while recent deep learning trends in sequence-oriented, attention-based, hybrid, and generative architectures have significantly improved MI classification efficiency and robustness.
The review systematically compares recurrent neural networks (RNNs), variational autoencoders (VAEs), generative adversarial networks (GANs), and transformers across multiple public MI datasets, analyzing the strengths and limitations of traditional models such as RNNs and LSTMs alongside emergent models such as VAEs and transformers in extracting intricate patterns from EEG data. It highlights persistent challenges including inter-subject variation, low signal-to-noise ratio (SNR), and obstacles in real-time signal classification. The study is available at DOI 10.3390/s26165097, published by MDPI.