A Convolutional Neural Network for Seizure Prediction using Synthetic and Electrocorticographic Intracranial Data

The ability to predict the onset of seizures would represent a significant increase in quality of life for patients with epilepsy refractory to treatment. However, the goal of accurate real-time seizure prediction has not been realized despite decades of efforts. Here, we combine intracranial electrocorticographic recordings possessing high temporal resolution from a patient database with synthetically created data to train a convolutional neural network. Using a wavelet transform, the voltage time-series data from each electrode channel can be converted into a separate image layer. The approach takes advantage of the mature ecosystem of tools for machine learning using medical images. The synthetic data is generated using a computational model of connected artificial neurons based on the Hodgkin–Huxley coupled differential equations with tunable parameters. This model has been shown to have the ability to exhibit both healthy and ictal brain function regimes with physiologically interpretable variables. The results of this work may help lead to the development of algorithms that can provide more accurate forecasts of real-time seizure risk.