Identification Post-Stroke of Motor Imagery and Asynchrony of Channel Pairs using Multiple RNN

Abstract

Stroke can cause disability, so the patient needs rehabilitation, and it is necessary to measure its effectiveness. The electroencephalogram (EEG) can capture the electrical activity in the brain, which can be real-time in post-stroke rehabilitation monitoring. EEG signal consists of several variables, including motor imagery and asynchronous from the symmetric channel. Both are features of post-stroke patients that are frequently used from previous studies, among other variables. EEG signals recorded from many channels can enrich the information on activity in the brain, including stroke. Motor imagery as a variable that reflects the stroke also be dominant in specific channels. Likewise, the asymmetry of the channel pair is too. Each channel or channel pair has its characteristics that are useful in identification. Meanwhile, a suitable method for identifying interconnected signals in time sequences is Recurrent Neural Networks (RNN). Therefore, to maintain the connectivity and take advantage of the EEG signal from multichannel, this paper proposed the Multiple RNN method in which each channel was processed by one network connected by a fusion function. The two variables - motor imagery variable and asynchronous of the symmetric channel pair are obtained from the Wavelet transform. The motor imagery feature involves FC5 and FC6 channels, while the asynchronous channel involves the AF3-AF4, F7-F8, F3-F4, FC5-FC6, T7-T8, P7-P8, and O1-O2 channel pairs. Both variables were obtained from the EEG signal using Wavelet at 1–7 Hz for asynchronous channel pairs and 8 – 30 Hz for motor imagery. The results showed that the Multiple RNN provided an accuracy of 88.04%, which increased by 8% compared to a Single RNN which obtained an accuracy of 80.09%. The results also showed the importance of choosing a learning rate to get the best accuracy.