Probing epileptic disorders with lightweight neural network and EEG's intrinsic geometry

The nonlinear dynamical systems can be stabilized on attractors in chaotic states, where the attractors depicted by dynamical trajectories may take on specific geometries. Electroencephalogram (EEG) signals are typically chaotic signals that have various nonlinear dynamic characteristics. Intrinsic geometry of EEG signals could contribute to tracking the recurrence of seizures and probing epileptic disorders, but it is ignored in most deep network-based seizure detection algorithms. Therefore, this paper presents an automatic detection framework called Recursive State-Space Neural Network (RSSNN) to infer the EEG geometry from single-channel signals and identify different epileptic patterns with a fast computational speed. RSSNN consists of a mathematical mapping module and a deep learning model. The former reconstructs EEG geometry in a high-dimensional state-space and maps it to a two-dimensional graph. The latter is a newly designed lightweight (0.68 MB) fully convolutional network that decodes geometry into brain states. We validated RSSNN on a public EEG dataset collected from epileptic patients with seizure and seizure-free conditions and healthy volunteers. A sliding window with a one-second length is utilized to verify the performance of RSSNN at the segment level. Moreover, the voting strategy is adopted to obtain the final prediction at the subject level. In the testing phase, RSSNN obtains an overall 99.79% accuracy at the EEG segment level and reaches 100% accuracy at the subject level. Notably, it takes less than 25 ms to predict one subject. This study proves the potential of EEG's intrinsic geometry as a seizure indicator for real-time monitoring by combining it with a lightweight neural network. It enriches the deep learning-based seizure prediction methodology in nonlinear dynamics.