Generative Enhancement of Noninvasive Datasets for Motor Brain-Computer Interface by Synthesizing Task-Relevant Neural Signals.

Despite the increasing adoption of deep neural networks (DNNs) in brain-computer interfaces (BCIs), developing high-degree-of-freedom (DOF) systems capable of decoding continuous movements, such as limb kinematics, remains a significant challenge. This difficulty stems from limited availability of task-specific neural features within individual neural signal datasets. To overcome this, we proposed a generative adversarial network (GAN) framework to enrich training features within neural signal datasets. Specifically, we synthesized artificial neural signal waveforms of the primary motor cortex (M1) from functionally related cortical regions, thereby enhancing neural datasets for improved motor kinematics decoding via DNN. Using magnetoencephalography (MEG) recordings during goal-directed arm-reaching tasks, our results showed that enhancing individual datasets with GAN-synthesized M1 signals significantly improved decoding performance by about 10% ( ${p} \lt 0.05$ ). Such improved performance is sustained even in the absence of real M1 signals. We further generalized the proposed enhancement to the motor imagery BCI competition dataset to improve classification accuracy. Our results highlight the potential of signal-generative networks to improve and augment motor BCIs to achieve freely intended movements.