Application Of Feature Selection In Neural Decoding Based On Local Field Signal

Invasive Brain-Machine Interface (BMI) builds a directly communication channel between the brain and external devices, which uses the recorded brain information as control signals. Neural Decoding is the key step in BMI researches. Compared with spike, Local Field Potential(LFP) is a low frequency signal and non-sensitive to time. It carries much more information from cluster of neurons instead of signal neuron and can decode the neural information effectively, which makes up the instability of spike-based BMI. But LFP data is so large and its features have high dimension, it is difficult to make traditional algorithms work and reach high accuracy with low computing-time expense. This paper researches on signal processing and analysis of LFP in invasive BMI, especially introduces feature selection and parallel computing methods to conquer these problems. Based on the characteristics of multi-frequency and multi-channel LFP signal, feature selection is used to decrease the dimensions and reduce redundancy. Then neural decoding is done with the selected features, and parallel computing is used to improve the algorithms performance.First, based on monkey grasp experiment, the thesis preprocesses LFP signals and extracts the features from the original data with short-time Fourier transform. The results in the thesis confirm LFP spectral power can represent the whole data as features and decode the reach-grasp task. Second, the thesis explores coherence between two LFP signals of different channels. Results note that LFP features between different channels and frequencies are strongly coherent, therefore large amount of redundancy exists among features’information. These findings strongly indicate the great importance of feature selection.In the thesis feature selection, the thesis studies four feature selection methods including Max-Relevance and Min-Redundancy based on mutual information (MRMR), spectral feature selection and improved spectral feature selection with minimum redundancy, Sparse logistic and Partial Least Square Regression. All the four algorithms choose approximately20features to obtain the highest or stabilized accuracy. It is observed that the accuracy has increased about10%by using improved spectral feature selection. The importance of each frequency band is evaluated by using all the features’weight before decoding, as well as the selected features according to the decoding accuracy. These results demonstrate each frequency band have different effects on decoding—high part of frequencies and low part are both the most significant one, while mid-frequencies has least relevance with behavior movement. Furthermore, utilizing parallel computing based on CUDA increases the process of feature extraction and decoding, which realizes high-performance computation of LFP BMI.