Decoding upper limb kinematic parameters from motor cortical activity using a deep learning algorithm

The current neural interface technology opens a new window for collecting multi-site streams of the firing activities of hundreds of neurons simultaneously. It provokes a need to equip with the means to harness as much information as possible from such huge amounts of neural data. Deep learning algorithms capable of representing latent components in large data may offer a means to extract useful information related to specific behavioral functions from these neural data. In this study, we aimed to decode movement-related information from motor cortical neuronal ensembles in a primate while the animal moved the arm and hand to perform an eight-target center-out task. The previous studies addressed the problem by decoding the velocity parameter to reconstruct arm-movement trajectories. However, as velocity can be decomposed into speed and direction, it may be advantageous to decode each parameter independently. Thus, we decoded speed and direction of the hand separately with the long short-term memory network (LSTM) to from the ensemble of one hundred fifty-eight primary motor cortical neurons. A comparison of the suggested LSTM decoder with traditional decoders directly predicting the velocity parameter using the linear Kalman filter or LSTM demonstrated a significant increase in the performance of reconstructing 2D hand trajectory. Our results may add accumulating evidence to the employment of deep learning algorithms for intracortical brain-machine interfaces and suggest that speed and direction can be decoded independently.