Efficient Detection of High-frequency Biomarker Signals of Epilepsy by a Transfer-learning-based Convolutional Neural Network

Abstract

High-frequency oscillation (HFO) is an important electrophysiological biomarker for estimating the epileptogenic zone in patients with epilepsy, but its clinical use is limited due to the high false-positive detection rate associated with conventional auto-detection methods based on one-dimensional spectral energy features. The purpose of this study was to apply a convolutional neural network (CNN)-based classifier to the candidate signals detected using conventional methods, and to extract HFOs more accurately and automatically. We adopted an image-based CNN because HFOs exhibit a localized power distribution in both time and frequency, which is utilized for the visual inspection of HFOs. To reduce the number of training datasets required for one patient, we employed transfer learning of an existing natural image classifier or the CNN HFO-classifier of another patient. We applied the proposed methods to the electrocorticography data of two patients with focal epilepsy who underwent pre-surgical examination. When the natural image discriminator AlexNet was transfer-learned to the HFO classifier, an accuracy of 93.0 ± 0.997% was achieved using 3000 training datasets. The false discovery rate (FDR) of HFO was 78.0% at the completion of the conventional method, which was significantly improved to 19.0 ± 4.42% after applying the CNN-based HFO classifier. When the HFO classifier trained with one patient was further relearned using the training datasets of another patient, the accuracy of determining HFOs in the latter patient was consistently above 91.0% (maximum 93.3 ± 0.967%) with the incorporation of 200 or more training datasets. These results suggest that the proposed method may provide an accurate, automatic, and personalized HFO classifier while liberating neurologists from the time-consuming manual detection of HFO signals for diagnosis.