Spatiotemporal wavelet-domain neuroimaging of chaotic EEG seizure signals in epilepsy diagnosis and prognosis with the use of graph convolutional LSTM networks

In the crucial arena of neurological care, pre-seizure, and seizure diagnosis stand as imperative focal points. While existing literature has probed this area, it demands sustained exploration given the intricate nature of seizures and the profound implications of prompt diagnosis on patient prognosis. Greater insights and novel advancements in the field of epilepsy diagnosis and prognosis can significantly bolster patient health and potentially redefine treatment management. Deep learning models like long short-term memory networks (LSTM) show promise for sequential data analysis. However, their application to electroencephalogram (EEG) signals for seizure detection reveals challenges, especially in imbalanced datasets. In response, we develop a hybrid graph neural network, integrating Convolutional Neural Networks (CNN) and LSTM through optimized skip connections. These connections, combined with our optimized graph structure, ensure no loss of crucial temporal data. The CNN layer efficiently extracts spatial features from samples, while LSTM emphasizes the EEG signal's temporal nuances. A unique facet of our proposed architecture is its optimized structure which is obtained based on Bayesian optimization. It does not merely refine network parameters but also systematically determines the optimal neuron count, layering, and overall architecture of our graph neural network. Alongside our deep learning methodology, we conduct a dynamical analysis elucidating the intrinsic chaotic patterns of seizure neural EEG signals. We demonstrate that the phase space analysis provides valuable insight for wavelet time-scale pre-processing for pre-seizure and seizure diagnosis. The numerical and empirical results validate the performance of our novel and breakthrough approach. Also, the results are compared with outcomes obtained using LSTM in different conditions.