Real-time Automatic Detection of Definite-REM Sleep

Sleep is an important physiological activity for humans, and its quality and quantity are closely associated with both health and disease. However, its fundamental mechanisms of sleep remain unclear, and research from various perspectives is ongoing. Sleep is classified into five stages (WAKE, REM, N1, N2, N3), of which REM is known to be strongly associated with dreaming and memory. Various studies, including experiments on human participants, are currently underway. REM sleep with rapid eye movement is known as definite-REM sleep. In REM-sleep experiments, some interventional experiments have stimulated participants with sound or other stimuli during definite-REM sleep intervals. In current interventional experiments, sleep experts usually monitor the sleep-state using EEG, EOG, and EMG in real-time and manually detect the definite-REM sleep interval based on the characteristics of these biosignals. However, this method is highly demanding, requiring significant expertise and placing a considerable burden on sleep experts. Therefore, we developed a novel system architecture and two deep-learning models (definite-REM detection (D-REM) and weighted-definite-REM detection (WD-REM) models) for automatically detecting definite-REM intervals in real-time. In experiments using real sleep data, WD-REM achieved the best performance (accuracy, precision, and false-positive rate of 93.9%, 83.0%, and 3.31%, respectively, for 1-second detection interval) and can detect definite-REM sleep in real-time with high accuracy. Furthermore, the introduction of dynamic thresholds further enhanced the performance of the model, resulting in accuracy, precision, and false-positive rate of 94.0%, 83.5%, and 3.17%, respectively. Thus, the proposed system architecture and model are applicable to REM sleep intervention experiments, such as those involving stimulations, and are expected to contribute to the advancement of sleep research.