BCIwiki (bciwiki.com) — Adding primary motor cortex (M1) signals synthesized by a generative adversarial network (GAN) to training data improved arm-movement decoding from magnetoencephalography (MEG) by about 10% (p < 0.05), Kim H and Kim JS reported in IEEE Transactions on Neural Networks and Learning Systems on October 1, 2026 (PMID 42821728, DOI 10.1109/tnnls.2026.3736056).
Deep neural networks are increasingly used in brain-computer interfaces (BCIs), but building high-degree-of-freedom systems that decode continuous movements such as limb kinematics remains difficult, the authors wrote, partly because individual datasets contain limited task-specific neural features. Their GAN framework generates artificial M1 waveforms from functionally related cortical regions to enrich those features before decoding with a deep neural network.
Using MEG recordings taken during goal-directed arm-reaching tasks, the team found that augmenting individual datasets with synthesized M1 signals significantly improved decoding, and the gain was sustained even when real M1 signals were absent. They also applied the method to a motor imagery BCI competition dataset to improve classification accuracy.
The authors said signal-generative networks could help improve and augment motor BCIs toward freely intended movements.