Summary During sleep, the brain alternates between rapid eye movement (REM) and non-REM (NREM) sleep, with recurring REM sleep episodes forming the ultradian sleep cycle, a hallmark of mammalian sleep. However, the mechanisms regulating the ultradian timing of REM sleep remain elusive, underscoring the need for reliable physiological predictors. Here, we developed a machine learning framework to identify features in the electroencephalogram (EEG), electromyogram (EMG), and sleep history that predict the timing of REM sleep in mice. A bidirectional long short-term memory (BiLSTM) network, trained to classify sleep-wake states, embedded EEG and EMG signals into a low-dimensional latent space, revealing slow, ramping dynamics between consecutive REM sleep episodes. Using these latent EEG/EMG features together with preceding sleep history, we predicted the time until the next REM episode. Feature importance analysis identified heart rate in the EMG, combined with sleep history, as a key predictor of REM sleep timing. Under heightened homeostatic pressure for REM sleep, heart rate decayed faster and reached lower levels, whereas pharmacologically elevating heart rate reduced REM sleep. Our findings position heart rate as a physiological marker of REM sleep timing and suggest a close link between cardiovascular regulation and the mechanisms controlling the ultradian sleep cycle.
Hao et al. (Sun,) studied this question.