Detection of high-frequency biomarker signals of epilepsy by combined deep-learning feature selection and linear discrimination analysis

Abstract

Interictal high-frequency oscillations (HFOs) are potential biomarkers to identify epileptogenic brain regions of epilepsy patients and track disease activity. We have previously developed a Convolutional Neural Network (CNN)-based automated HFO detection method from electrocorticogram data. However, the clinical use of HFO information with the proposed method might be limited due to the need for extensive training datasets to achieve sufficient accuracy. This study therefore aimed to improve the accuracy of the classifier even with the small training dataset. We adopted a hybrid system where features selected by the CNN model are further transferred to a state-of-the-art classifier. We experimented with 6 pre-trained CNN models and 6 classifiers on our dataset. Results suggest transferring the feature information from the trained VGG19 model to the Linear Discriminant Analysis classifier provides the best result; an accuracy of 87% was achieved for 50 training images. Our proposed method could contribute to clinical research by reducing the size of annotated datasets required for the personalized and accurate HFO detection.