Deep learning methods for diagnosis of epilepsy from EEG using convolutional neural networks

Abstract

Spike like waveforms, which are different from normal background waveforms, are usually discovered in electroencephalogram(EEG) of epilepsy patients. Diagnosing epilepsy by using spikes can be tedious and requires doctors with special training. Therefore, we aim to develop algorithms for automated spike detection to assist doctors in decision making and help patients in areas with few specialized doctors. Over the years, scientists have tried different methods for spike detection, however, there is still huge space for accuracy improvement. Among them, deep learning has shown huge potential in driving research work to a tremendous leap forward. In this project, we aim specifically in optimizing deep learning convolutional neural network(CNN) architecture to improve the accuracy of spike detection. Moreover, we compared the different performance between 1D and 2D CNN models, and further discovered the relations between spike number and epilepsy patients diagnosing. After training and testing on EEG signal of 93 patients and 63 healthy subjects, the best model has achieved an accuracy of 99.97% in spike detection and an accuracy of 90.5% in patient diagnosis. The model has proven its capability in detecting abnormal data pattern in the premise that no spike definition input and human intervention were given.