A Robust and High Accurate Method for Hand Kinematics Decoding from Neural Populations

Offline decoding of movement trajectories from invasive brain-machine interface (iBMI) is a crucial issue of achieving cortical movement control. Scientists are dedicated to improving decoding speed and accuracy to assist patients in better controlling neuroprosthetics. However, previous studies treated channels as normal sequential inputs, merely considering time as a dimension representing channel information quantity. So, this inevitably leads underutilization of temporal information. Herein, a QRNN network integrated with a temporal attention module was proposed to decode movement kinematics from neural populations. It improves the performance by 3.45% compared to the state-of-the-art (SOTA) method. Moreover, this approach only incurs a increase of parameter less than 0.1% compared to the QRNN with same hyperparameter configuration. An information-theoretic analysis was performed to discuss the efficacy of the temporal attention module in neural decoding performance.