Parallel Hybrid 2D Convolutional Neural Networks and Recurrent Neural Networks in EEG Post-Stroke Identification Based on EEG Signal

Electroencephalogram (EEG) captures brain activity that can used for diagnosis and monitoring stroke scale. EEG signals contain spatial and temporal information to analyze signals with stroke characteristics. This research proposed a parallel hybrid approach using 2D Convolutional Neural Networks (2D CNN) and Recurrent Neural Networks. The Wavelet method obtains Delta, Theta, and Alpha wave representations of the EEG signals, representing the characteristic post-stroke frequencies (1 – 13 Hz). This approach, in parallel, uses 2D CNN for spatial feature extraction from each waveform, while RNN-GRU is used for temporal analysis within each signal segment. Experiments were conducted using an EEG dataset from 75 subjects, evaluating five techniques for weight correction. The results show that the parallel hybrid 2D CNN-GRU model achieves an 89.96-90.96% accuracy. However, the weight correction technique and learning rate influence performance but are insignificant. The SGD optimizer with a learning rate of 0.01 achieved the highest accuracy, although it required a longer training time.