Neural decoding using local field potential based on partial least squares regression

Recent studies have shown that a promising cortical control signal in brain-machine interface is local field potential (LFP), of which low and high frequencies bands contains information about planning or executing dexterous movement. In this paper, we analyzed LFP signals recorded from primary motor cortex of rats as they performed a lever-pressing task. The decoding performance of partial least squares regression (PLSR) in LFP was evaluated by comparing with two traditional decoding algorithms, Wiener filtering (WF) and Kalman filtering (KF). The results demonstrated that PLSR not only had good performance as the other two methods, but also had particular predominance in avoiding over-fitting and computation complexity, due to its capability in dealing with the small sample capacity and high variable dimension that exist in LFP decoding.