Recently, decoding that predicts behavior from an electroencephalogram(EEG) has been many reported, and decoding an electromyogram(ECG) from an EEG is one of them. However, there are many waveforms in the brain such as noise from experiment environment, and the EEG signal is not only directly related to the EMG but also includes the waveform reflecting another role. Filtering of Wavelength has been mainly used as a method for classifying such as EEG, but in recent years, a denoised method using deep learning has also been used for waveform analysis. The purpose of this study is to evaluate decoding accuracy of denoised EMG using deep learning and to classify the ratio of EEG reflected in muscle activity and EEG reflected in different activity.
We recorded primary motor cortex(M1) EEG and 4 or 5 regions of EMG in the rat brain. Next, the denoising EMG using Latent Factor Analysis via Dynamical Systems(LFADS), a deep learning method reported using EMG analysis. After denoising, EMG and EEG were coded each other to evaluate their accuracy.
Denoised EMG using deep learning showed higher decoding accuracy than just filtering. Moreover, the reflection ratio of EEG to electromyogram was evaluated by mutual decoding of EMG and EEG. This study suggested　that the usefulness of noise removal by machine learning and the EEG classification can be made more accurate.