A Novel Multi-scale 3D CNN with Deep Neural Network for Epileptic Seizure Detection

Accurate and timely detection of seizure activity during continuous EEG monitoring either in epilepsy monitoring unit or in neuro-intensive care unit is crucial for both physician and paramedical personnel. However, it is laborious work and required special training for epilepsy and EEG interpretation. In order to detect automatically seizure, we propose a Multi-scale 3D-CNN with Deep Neural Network (DNN) model for non-patient-specific seizure detection. We considered spectral, spatial and temporal features. The EEG signals are transformed to frequency domain using Short Time Fourier Transform (STFT) to extract spectral features. The spectral features are mapped to 2D images to preserve the position of the electrodes. The proposed model is composed of 3D-CNN and bidirectional Gated Recurrent Unit (GRU) to extract spatial and temporal features from the 2D mapped images. We evaluated the proposed model using CHB-MIT and Seoul National University Hospital (SNUH) Scalp EEG database. Our proposed model achieves the sensitivity of 89.4% and 97% and a false positive rate of 0.5/hours and 0.6/hours on the CHB-MIT database, and the SNUH database, respectively.