Reducing Computational Costs in Responsive Neurostimulation via Seizure Analyses Based on Three-Dimensional Convolutional Neural Networks

Responsive neurostimulation is considered an effective treatment for patients with drug-resistant epilepsy. This system, implanted in the crania, continuously monitors brain activity though recording electrodes, and delivers targeted electrical stimulation to suppress seizures when detected. However, the hardware resources of the implantable device and the number of connectable electrode channels are limited. Therefore, this study aimed to systematically reduce the number of recording channels and the computational cost of seizure detection by analyzing the spatial and temporal kernels in convolutional neural networks trained on three-dimensional electro-corticogram (ECoG) data, which preserves both the grid electrode spatial configuration and ECoG time-series information.