Decoding Forelimb Muscle Activitv With Local Field Potentials From Mouse Motor Cortex

Brain-machine interfaces(BMIs)normally use neuroelectrophysiological signals to decode motion parameters,while studies on electromyography(EMG)as decoding objects are relatively few,mainly focused on humans and monkeys.However,it is a big challenge to carry out studies in mice,since electrophysiological signals in mice are difficult to record and only a few electrodes are available for neural decoding.In conclusion,this paper is the first step towards online BMIs.This paper focuses on the relationship between the local field potential(LFP)of the caudal forelimb areas(CFA)or the rostral forelimb areas(RFA)of mice and the EMG of the biceps brachii muscle of the forelimb,which is of great significance for basic neuroscience and for the design of neural prosthesis.In this paper,we recorded four-channel local field potential signals from the CFA and RFA of mice during the lever-press task,and recorded the EMG of the biceps brachii muscle of the forelimb simultaneously.We analyzed the correlation between different frequency bands of LFP and EMG.And the highest correlation coefficient was obtained in the frequency band between ?1 and high-?1(30-100hz),indicating that the mid-frequency bands have more contribution(weights)in EMG decoding.It was also found that when the EMG signal reached the peak,the LFP signal was nearly to the end,and the change of LFP signal was slightly earlier than that of EMG.The LFP features were taken as the input,and the envelope of the EMG signal was taken as the output of the model.The decoding models were established by using linear model like kalman filter,as well as nonlinear model like general regression neural network(GRNN)and recurrent neural network(RNN),to decode the EMG signals of the forelimbs during the lever-press task in freely moving mice.Even though the number of neural signal channels is small,these three algorithms can successfully decode the four-channel LFP signal and obtain the predicted waveform similar to the real EMG signals.Among the three,RNN has the best decoding performances,with an average CC value of 0.78 and an MSE value of 0.02.Similar to the results of other decoding studies,it is shown that low dimensional neural signals can still successfully decode EMG signals.On account of LFP signal changes slightly ahead of EMG,time-delay parameters were added to improve the decoding models.There was an obvious improve in decoding accuracy on GRNN,but not on Kalman filter or RNN.In addition,on the three decoders,it was found that the delay time of RFA achieved its the optimal decoding performance about 62.5±29.2ms longer than RFA did,which suggesting that RFA could make the motor planning earlier than CFA.Furthermore,we compared the mean power spectra of the LFP in the two cortex regions and the mean EMG signals,and found that RFA was active about 75ms earlier than CFA.The results were in line with the large body of evidence which showed that RFA played a major role in motor planning rather than CFA.