A Deep Residual Neural Network Based Framework for Epileptogenesis Detection in a Rodent Model with Single-Channel EEG Recordings

Epilepsy is one of the most common neurological disorders affecting patients across all ages. During the progression of the disease, termed epileptogenesis (EPG), patients may not yet show any clinical manifestation. The EPG phase can range from weeks to years and patients with epilepsy are usually diagnosed by the occurrence of a spontaneous seizure followed by electroencephalography (EEG) monitoring in the hospital. However, the more seizures they have, the less effective the treatment will be. Detecting the development of epilepsy before the first spontaneous s eizure may a llow for e arlier intervention and better treatment outcome. Here we propose a framework based on deep residual neural networks to identify the EPG phase based on EEG recordings in a rodent model where the epilepsy is induced by perforant pathway stimulation (PPS). A deep convolutional neural network is trained to distinguish EEG data recorded before (baseline period, BL) and after (epileptogenesis period, EPG) the EPG is triggered. The proposed model takes the Fast Fourier Transform (FFT) of the preprocessed five-second long EEG segments as input. During testing, we apply a prediction aggregation across multiple consecutive segments to accumulate information over a longer time period. When classifying a continuous stretch of one hour of data, our model achieves 83 % sensitivity and 83 % specificity. Further analysis suggests interpretable features in the FFT transformed data that contribute to the distinction of the two phases.