Miniaturization of Epileptic Electrocorticogram Detector Using 3D-Convolutional Neural Network

Epilepsy is a neurological disorder characterized by sudden and recurrent seizures caused by abnormal electrical activity in the brain. Responsive Neurostimulation (RNS) offers a promising treatment option for patients with drug-resistant epilepsy. Responsive Neurostimulation (RNS) is an implantable device that employs a closed-loop system. It continuously monitors brain activity through electro-corticogram (ECoG) recordings. When the system detects seizure activity, it delivers direct electrical stimulation to the brain to suppress the seizure. Seizure detection algorithms require patient-specific optimization, leading to increased interest in deep learning approaches in recent years. While deeper network architectures generally improve detection accuracy, their implementation in implantable devices is constrained by limited hardware resources and the restricted number of electrode channels available for ECoG monitoring. To ensure the practical feasibility of RNS, it is crucial to systematically minimize both the computational costs of patient-specific deep learning models and the number of connected ECoG electrodes. This study systematically reduced the number of electrode channels and computational costs in seizure detection models by analyzing the spatiotemporal kernels learned by the first convolutional layer of a 3D convolutional neural network (3D CNN) trained on 3D ECoG data. This approach capitalizes on the networks ability to learn spatial relationships between grid electrodes and the temporal dynamics of ECoG signals. The performance comparison between the downsized seizure detection CNN model and the original CNN model revealed that, for at least some patients, it is possible to maintain inference performance while reducing the model size.