Small Sample Epilepsy Detection Method Based on Convolutional Prototype Learning

Abstract

Abstract Background: Several scalp EEG epilepsy detection methods based on machine learning have achieved good detection accuracy. However, in clinical applications, different EEG acquisition equipment and experience of neurologists make the quality and style of EEG signals different, which makes previous epilepsy detection models cannot be widely used. The establishment of epilepsy detection model for a certain hospital usually depends on a large number of EEG samples, but there are usually few EEG samples from a certain hospital. Methods: To solve this problem, we proposed a small sample epilepsy detection method based on convolutional prototype learning (CPL) in this paper. CPL consists of convolutional neural network (CNN) and prototype learning. CNN is used as an adaptive feature extraction algorithm, and prototype learning is used as a small sample classification algorithm. Results: In the experiment, we select 20, 40, 60, 80, 100 and 120 samples to train and save 6 CPL-based detection models. The 6 models are used to classify the test samples, and the accuracy are 75.97%, 83.24%, 85.67%, 88.27%, 91.09% and 94.43% respectively. Conclusions: The CPL can realize automatic feature extraction of EEG signals, and solve the problem of insufficient training samples in epilepsy detection.