Automated Sleep-Spindle Detection : Performance Comparison Analysis and By-Subject Optimization

Abstract

Sleep spindle is brain activity mainly observed in stage 2 (N2) of human sleep cycles by electroencephalography (EEG). Because of its role in brain development and learning, biological and clinical interest in sleep spindle are rapidly growing and several automated methods of spindle detection have been developed. In this research, we evaluated performance of automated spindle algorithms in objective ways applying published automated spindle algorithms to information of sleep EEG and previously detected sleep spindles in polysomnographic data from middle-aged subjects with sleep disorders. We also defined expert group consensus and established the gold-standard data set. After evaluating performance of automated spindle algorithms using published parameter, we optimized automated spindle algorithms by trying variation of published parameter for each automated spindle algorithm to improve performance. In performance comparison analysis of automated algorithms, it had an opposite characteristic that when precision was high, recall was low and when recall was high, precision was low. The limitation of the worst performance in specific automated algorithm and severe variation in subjects had to be reconsidered. We need to try possible combinations of 2 or more different algorithms and apply deep learning method to improve automated spindle detection.