Toward an Automatic Classification of the Different Stages of Sleep: Exploring Patterns of Neural Activity in the Subthalamic Nucleus

Abstract

Sleep disorders substantially impact quality of life, especially in patients with neurodegenerative diseases like Parkinson's disease. Recent advances in deep brain stimulation highlight the potential of closed-loop adaptive stimulation that utilizes neural feedback signals recorded directly from the stimulation electrodes. The subthalamic nucleus, a distinct structure located deep in the brain, plays a major role in processing cortical information and could be used to classify sleep stages. We recorded local field potentials in the subthalamic nucleus of two freely moving nonhuman primates across three nights. Our study examined subthalamic neuronal activity across different vigilance stages using spectral activity, multiscale entropy analysis, and an automatic classification. Results revealed distinct spectral patterns in subthalamic activity corresponding to sleep stages, with a high synchronization between subthalamic nucleus and EEG signals during deeper sleep stages. These deeper stages were associated also with reduced entropy, suggesting decreased neural activity complexity. An automated machine learning classifier based on subthalamic nucleus spectral activity distinguished wakefulness from sleep with high accuracy (94% for both animals). While the classifier performed well for deeper sleep stages, its accuracy was lower for lighter sleep stages. Our findings suggest that subthalamic nucleus activity can mirror cortical dynamics during sleep, supporting its potential use in developing closed-loop stimulation therapies for sleep disorders. This work provides a foundation for further studies in Parkinson's disease models to evaluate the translational relevance of subthalamic nucleus activity in clinical applications.

Keywords: Parkinson's disease; deep brain stimulation; nonhuman primates; sleep and wakefulness; subthalamic nucleus.

FIGURE 1

(A) Sagittal radiography showing electroencephalography electrodes (EEG electrode) implanted in contact with the dura mater and the deep brain stimulation electrode (DBS electrode) implanted in the subthalamic nucleus (STN). The guideline between the anterior (AC) and the posterior commissure (PC) is highlighted in white. (B) Histological reconstruction of deep brain stimulation electrode placement in M1 and M2. STN is highlighted in black on the brain slice drawing. (C) Examples of raw 30‐s EEG, STN local field potentials (LFPs) and EMG from M1 and M2 during wakefulness (AW), sleep stage 1 (N1), stage 2 (N2), stage 3 (N3), and paradoxical sleep (REM). (D–F) Example of a 12‐h night for M1, with raw hypnogram (D), and corresponding time frequency analysis of EEG (E), and STN LFPs (F).

FIGURE 2

Mean cross‐correlation between matching EEG and STN signals during wakefulness and the different stages of sleep for Monkey 1 (A). Mean ± 3SD of cross‐correlation between unmatched signals is shown in black. Violins plots represent the cross‐correlation peak between each stage of vigilance for M1 (B) and M2 (C). The comparison was made with a Kruskal–Wallis followed by Dunn's test (*p ≤ 0.001).

FIGURE 3

Comparison of mean power spectral density of STN LFPs between stages of vigilance in each frequency bands for M1 and M2 combined. Power was averaged in frequency bands delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), and beta (13–32 Hz). Comparisons were made with a Kruskal–Wallis followed by Dunn's test, only pertinent significant comparison and nonsignificant comparison are shown (*p < 0.001; ns p > 0.05).

FIGURE 4

Comparison of MSE factor for EEG and STN signal between the different stages of vigilance for M1 and M2 combined. Mean data are represented ± standard deviation and comparison are made with a Kruskal–Wallis test followed by a Dunn's test (*: p ≤ 0.001).

FIGURE 5

(A) Confusion matrix showing the classification of sleep versus wake using STN features for M1 and M2 when tested on data from one night of each monkey. (B) Confusion matrix showing the classification of stages of sleep using STN features for M1 and M2 when tested on data from one night of each monkey. The value in each square represents the rate of 30‐s segments that were predicted correctly compared with the ground truth class.