Multi-Networks of Motor Imagery and Asymmetric Channel Pair for EEG Stroke Scale Classification

Stroke disrupts the blood supply and nutrients to the brain, causing dysfunction or disability that reduces the patient's quality of life. Rehabilitation is required to restore function and independence. So, monitoring instruments such as an electroencephalogram (EEG) are required. EEG records the electrical brain activity, which needs feature extraction before classification. It is done to obtain variables that distinguish the stroke scale captured through EEG signals. Some of these are the ability to imagine movement or motor imagery and the difference in waves from symmetric channel pairs. This paper proposes a stroke scale classification method based on EEG signals using multi-networks of two-dimensional convolutional neural networks. Wavelet filters separate EEG signals into some frequency bands called waves. The multi-network capability allows CNN to map EEG signal variables into class features simultaneously. The 2D-CNN, with its precision in connecting inter-channel and inter-temporal signals, ensures reliable classification. The frequency bands of EEG signals represent the variables of motor imagery and the asymmetry of channel pairs. The experimental results show that motor imagery and asymmetric channel pairs based on EEG signals are characteristics of stroke levels. The proposed method, a multi-network 2D CNN, achieves an accuracy of 90.19%, 9% higher than that of a single network. In addition, multi-network 2D CNN also speeds up computation, which is very beneficial in EEG signals. The experiment also tested the fact that stroke information is almost identical in all channels. Motor imagery variables are slightly more accurate than the asymmetry of symmetric channel pairs.