커리큘럼 학습을 통한 원시 단일 채널 EEG의 자동 수면 단계 분류 성능 향상

Abstract

This paper presents a curriculum learning method to improve sleep scoring performance of the deep neural network on raw single-channel EEG. Although deep neural networks have shown remarkable results on automatic sleep scoring, the remaining problem is that N1 and REM stages are hard to be distinguished due to their similar frequency characteristics. To solve this problem, we propose curriculum learning method for automatic sleep scoring; a deep neural network is trained in three phases which are in order of sleep scoring difficulty. The network is trained with three sleep stages (Wake, Non-REM, REM), four sleep stages (Wake, Light Sleep, Deep Sleep, REM), and five sleep stages (Wake, N1, N2, N3, REM) in Phase 1, Phase 2, and Phase 3, respectively. We compared the sleep scoring performance w/ and w/o curriculum learning for three deep neural networks from prior works and two different EEG channels. As the effect of curriculum learning, the overall performance has been enhanced maximum 1.7%p, 2.1%p, 0.022 in ACC, MF1, Kappa, respectively. Especially, per-class F1-scores of N1 and REM have been increased maximum 2.4%p and 2.8%p, respectively.