[4E3-GS-2] Machine learning

[4E3-GS-2-03] Imaging of time-series EGG intended to imitate visual information processing and detection of abnormal waves by EfficientGAN

〇Riku Sugiura¹, Tsuyosi Ishizone¹, Kazuyuki Nakamura¹ (1. Meiji University)

Keywords:GAN, Anomaly detection, EGG

In the field of electroencephalography (EEG), which is used to diagnose epileptic seizures, the demand for the development of an EEG reading AI is increasing due to the decrease in the number of EEG physicians today. In this study, we developed an AI for detecting abnormal EEG waveforms to support EEG physicians. We used time-series EEG data published in "Open Source EGG Resources" and converted into 2-channel fill-in 1images at 6-second intervals to imitate the visual information processing mechanism of an EEG reader. As a research method, classification by anomaly score of the data calculated by EfficientGAN was employed. Even when training was conducted with only a small amount of normal data, differences of anomaly score were confirmed from the results of adaptation to both normal cases and abnormal cases. This study suggests the possibility of detecting abnormal waveforms by learning with a small amount of normal EEG data using EfficientGAN.