Time-varying Analysis And Decoding Of Neural Spike Activity For Motor Brain Machine Interfaces

Brain machine interfaces (BMIs) build a direct pathway for the communication between the brain and the external instruments, aiming to help restoring motor function for motor-impaired patients. The decoding algorithm is the core of BMIs, which translates the neural activity into behavior. Previous methods usually assume a static functional map between neural firing and movement. However, recent work indicates the significant time variance in neural activities during short experiments, which greatly reduces the decoding performance. Based on the invasive BMIs on rats and non-human primates, we analyze the time-varying neural encoding characteristics and integerate it into the design of the non-stationary decoding algorithms, which makes the prediction more accuracy and more stable.We establish the invasive BMI platform on rats performning lever-pressing task and non-human primates performing2D tracking task. Neural activities are collected from primary motor cortex and syncrized with the movement. As the firing patterns have been qualitatively proved to be time-varying, we develope a time-varying general regression neural network (GRNN) as a black box to decode the neural firings. The model is designed with a dynamic pattern layer, which can take in the new patterns and forget the old ones in time. Therefore it can follow the change of firing patterns. Futhermore, we analysis the tuning characteristic of single unit activity, and design a time-varying decoding algorithm as a grey box that involes more physiological information. First, we estimate the time-variant tuning curves in a data-driven way and find both functionally "inhibitory" and the traditional "excitatory" function neurons. Then, we find worthy noticing changes in both information amount and member of the important neuron subset. Also, we propose a random walk model to predict time-variant tuning curves. Based on these time-varying neural tuning analyses, we propose a dual Monte Carlo point process filter (MCPP). The grey box method enables the estimation on the dynamic tuning parameters. When applied on both simulated neural signal and in vivo BMI data, the proposed adaptive method performs better than the one with static tuning model.This study first evaluates the non-stationary neural activity of M1in rats and monkey quantitatively. The innovations are,(i) desiging a time-varying GRNN with a dynamic pattern layer. When applied to pressure prediction in rats’data, the mean prediction error decreases;(ii) establishing a parameteric way to predict time-varing neural activity based on the linear-nonlinear-posisson model, which captures multiple neural tuning propeties;(iii) proposing a dual MCPP method with updated tuning models. When applied to trajractory estimation of monkeys, the normalized mean square error decreases over5%. Overall, our results raise a promising way to design a long-term-performing model for BMI decoder.