Epilepsy Detection Based On Deep Active Learning

EEG plays an important role in epilepsy detection.Experts can judge whether patients have epilepsy by analyzing EEG.Due to the disadvantages of artificial judgment,many epilepsy detection algorithms based on machine learning and deep learning have been proposed and achieved good results.However,the premise of excellent performance of machine learning and deep learning is the need for a large number of labeled training data.In practical application,most of EEG data are unlabeled data,so it is difficult to obtain a large number of labeled training sets.Therefore,this paper proposes an epilepsy detection framework based on deep active learning,which aims to use active learning to reduce the demand of epilepsy detection model for labeled training set.In order to solve this problem,the work of this paper is as followsFirstly,multi-channel EEG signals are transformed into multi-dimensional spectrograms to preserve the temporal,frequency-domain and spatial structure of EEG signals.The spectrogram was input into the following four models: convolutional neural networks,long short-term memory,maxcnn and mix,and then a better model was selected for active learning according to the experimental results.Secondly,in the framework of deep active learning,the initial training set is constructed with some labeled samples and the model is initialized.Then,the combination of uncertainty measure and diversity measure is used as the selection strategy,and "high value sample" is given priority to expert annotation and model training.Finally,the experiment is carried out on the Neonatal EEG public data set.Experimental results show that compared with the epilepsy detection based on deep learning,the proposed epilepsy detection framework based on deep active learning can reduce the labeling cost by about 25%.