Dual-layer hand gestures decoding with wireless epidural braincomputer interface in a tetraplegia

Spinal cord injury disrupts the neural connections between the brain and limbs, resulting in tetraplegia. Brain-computer interface (BCI) hold promise for enabling voluntary limb movements in tetraplegic patients, yet achieving fine motor control of the hand remains a challenge. Invasive BCI based on intracortical electrode arrays have demonstrated real-time multi-gesture decoding. However, their long-term safety is a major barrier in clinical applications. In this study, a tetraplegic patient was implanted with our recently developed wireless minimally invasive BCI, which records epidural field potential from eight electrodes over the sensorimotor cortex to decode continuous hand movement intentions. Natural hand movements can be decomposed into dual layers: the high level movement states and the low level finger kinematics. Accordingly, we propose a dual-layer decoding algorithm for multi-gesture BCI decoding. The upper layer infers the movement state using a hidden Markov model, while the lower layer decodes finger motion variables through a mixture of experts and filters them with a state specific linear system. This approach enables the real-time decoding of six hand gestures, outperforming classical decoders and recurrent neural networks. The proposed dual-layer framework achieves multi-gesture decoding solely from epidural EEG signals, paving the way for the development of flexible and robust BCI control of hand movement.