[4Xin1] Poster session 2

[4Xin1-14] Improvement of deep-learning-based epileptic EEG feature detection by data augmentation

〇Masaki Sawada¹, Nawara Broti¹, Yutaro Takayama², Masaki Iwasaki³, Yumie Ono⁴ (1.Electrical Engineering Program, Graduate School of Science and Technology, Meiji University, 2.Department of Neurosurgery, Yokohama City University Hospital, 3.Department of Neurosurgery, National Center Hospital, National Center of Neurology and Psychiatry, 4.Department of Electronics and Bioinformatics, School of Science and Technology, Meiji University)

Keywords:Data augmentation, High-frequency Oscillation, Convolutional Neural Network

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 rate with conventional detectors based on one-dimensional spectral energy features. We have previously developed a convolutional neural network-based HFO classifier using a spectral image of the candidate signal in the time-frequency domain. However, there still remains a limitation of high resource cost for data annotation by experienced neurologists. In this study, we applied data augmentation to increase the number of training data for 16 times by image flipping, inverting contrast, and adding salt and pepper noise. The F1 score of 0.85 was achieved with approximately 250 image which is a quarter of the previous data requirement. These results demonstrate the usefulness of data augmentation for clinical biomarker signals for an epilepsy diagnosis.