The impact of REM sleep loss on human brain connectivity

Sleep deprivation is common and linked to negative emotional, metabolic, and cognitive outcomes. REM sleep predominates in late-night sleep, whereas NREM dominates early-night sleep. REM sleep is implicated in coordinating brain structures and processes (PGO waves, theta rhythms, synaptic plasticity), yet human whole-brain mechanisms by which REM affects resting-state functional connectivity are underexplored. This study asks: (1) which brain networks are associated with REM sleep when integrating resting-state fMRI across different REM distributions; and (2) whether lack of REM sleep (early vs late-night deprivation) alters REM-related brain networks compared with a full rested night, given REM’s late-night predominance. Using a split-night paradigm and between-subject design, healthy adults underwent early- or late-night half-sleep or full-night sleep. REM sleep was quantified by duration and proportion. Connectome-based predictive modeling (CPM) was applied to identify REM sleep-associated whole-brain functional connectivity and to test group differences in REM sleep connectomes.

Prior neuroimaging work shows insufficient sleep disrupts functional activity and connectivity in executive control, hippocampal-amygdala circuits, default mode, attention, and salience networks. REM-related physiology suggests widespread brain coordination: PGO waves propagate across the brain; REM theta supports hippocampus-dependent memory; REM promotes synaptic reinforcement/suppression in neocortex. REM has been linked to emotional memory and creativity, and sleep timing shapes REM distribution (greater in late night). CPM offers advantages over traditional regression by avoiding overfitting and improving generalizability, making it suitable to identify REM sleep connectomes. Existing studies report DMN vulnerability to sleep loss and thalamocortical state dependence, motivating investigation of REM loss impacts on large-scale connectivity.

Participants: 113 right-handed healthy adults from multiple universities in Beijing were randomly assigned to three groups: late-night sleep deprivation (sleep 23:00–03:30; n=41), early-night sleep deprivation (sleep 03:00–07:30; n=36), and full-night sleep (sleep 23:00–08:00; n=36). Inclusion criteria included 7–8 h habitual sleep, no psychiatric/neurological history, no drug use, and no MRI contraindications; females had regular cycles and no oral contraceptives. Compliance was monitored via 7-day sleep diary and actigraphy; participants abstained from drugs/alcohol/caffeine for 48 h pre-study. Ethical approval: IRB00001052-23141; written informed consent obtained.

Polysomnography (PSG): Somte mobile PSG (Grael, Compumedics) recorded two nights: adaptation and experimental. EEG electrodes at F3, F4, C3, C4, O1, O2 (10–20 system), bilateral EOG, submental EMG, ECG. Sleep staged in 30-s epochs per AASM by two independent blinded technicians. Exclusion rules: sleep latency >30 min or single awakening >30 min terminated the experiment; falling asleep during restriction resulted in termination.

Split-night procedure: Both deprivation groups had 4.5 h sleep opportunity. Late-deprivation: sleep 23:00–03:30, then awake 03:30–07:30 (predominantly NREM in first half; awake during second). Early-deprivation: awake 23:00–03:00, then sleep 03:00–07:30 (predominantly REM in latter half). Quiet wake activities permitted (reading paper materials, indoor walking, water).

rs-fMRI acquisition: 3.0-T GE Discovery MR750. Resting-state scans 8 min (TR=2 s; 240 time points), TE=30 ms, flip angle 90°, 33 axial slices, slice thickness 4.2 mm, no gap, in-plane resolution 3.5×3.5 mm², FOV 224×224×64 mm³. Structural T1: voxel 1×1×1 mm³, 192 slices (TR=6.7 ms; flip angle 12°; thickness 1.0 mm).

Preprocessing: SPM12 and DPABI in MATLAB. First 10 volumes discarded; realignment to mean image; co-registration to anatomy (affine); DARTEL normalization to MNI space; functional smoothing FWHM 6×6×6 mm. Motion exclusion: average displacement >3.0 mm or rotation >3.0° removed.

Network parcellation: Power template defining 227 cortical and subcortical ROIs, categorized into 10 networks: DMN, VIS, FPN, DAN, VAN, SAN, CON, AUD, SMN (sensorimotor/dorsal), SUB.

Connectivity and CPM: For each subject, mean time series per ROI extracted; Pearson correlations between all pairs computed and Fisher z-transformed to form connectivity matrices. CPM used REM sleep duration (selected due to high correlation with proportion) and whole-brain FC to predict individual REM duration. Positive and negative edge sums computed per subject as features. Leave-One-Out Cross-Validation (LOOCV) used, with linear models fitted to predict REM duration from edge sums. Network-level and regional characterizations assessed predictive edge counts and correlations to quantify feature importance and contributions. Group-level analyses compared REM connectome connectivity across full-night late segment (late-FS), early-deprivation, and late-deprivation.

• REM timing within full-night sleep: Late-night sleep (03:30–08:00) had significantly greater REM duration and proportion than early-night sleep (23:00–03:30) in the full-sleep group (duration: n=33, t=8.26, Cohen d=1.40, P=1.93e−09; proportion: n=33, t=8.76, d=1.55, P=5.15e−10).
• Effects of deprivation relative to full-night segments: • Early-deprivation vs late-FS: REM duration decreased (t=−2.02, d=−0.24, P=0.046) and REM proportion decreased (t=−3.82, d=0.67, P=2.81e−04). • Late-deprivation vs early-FS: REM duration decreased (t=−2.27, d=−0.29, P=0.026); REM proportion not significantly different (t=−0.99, d=−0.12, P=0.32).
• Early- vs late-deprivation: Early-deprivation exhibited more favorable REM pattern than late-deprivation (duration: n=77, t=6.17, d=0.70, P=3.19e−08; proportion: n=77, t=6.80, d=0.78, P=2.20e−09).
• CPM prediction: The positive network significantly predicted REM duration (n=110, r=0.20, one-tailed P=0.017).
• REM sleep connectome characterization: • Dominant predictive edges within DMN (DMN–DMN) and between DMN–VIS, with notable contributions also from CON–AUD and SUB–VIS. • At the large-scale network level, DMN, VIS, and SUB networks contributed most to predictive edges (~51%); prediction contributions were broadly distributed across networks, with DMN, SMN, VIS, and AUD important. • Regional level: Thalamus, calcarine cortex, lingual gyrus, and superior temporal gyrus (STG) showed highest edge counts and contributions; regional edge and correlation measures were highly similar (n=227, r=0.79, P=6.17e−50).
• Group differences in REM connectome under deprivation: • DMN–DMN connectivity differed across groups (F(2,107)=8.10, P=5.29e−04). Late-deprivation < late-FS (n=69, t=3.55, d=0.43, P=7.12e−04) and late-deprivation < early-deprivation (n=77, t=3.18, d=0.36, P=0.002). • VIS–SUB connectivity differed across groups (F(2,107)=6.19, P=0.003). Late-FS > early-deprivation (n=74, t=2.80, d=0.43, P=0.007) and late-FS > late-deprivation (n=69, t=2.80, d=0.43, P=0.007). • Specific DMN–DMN edges with group differences included mPFC–paracingulate (F(2,107)=4.01, P=0.029) and SFG–PCC (F(2,107)=3.11, P=0.048). VIS–SUB notable edges included thalamus–lingual gyrus (F(2,107)=7.68, P=6.61e−04) and thalamus–calcarine (F(2,107)=6.34, P=0.003).

REM sleep loss alters resting-state functional connectivity, with REM connectome features concentrated within and between DMN, VIS, CON, and SUB networks. The thalamus demonstrated the highest centrality and contribution, aligning with its role in sensory relay and sleep stage regulation, and suggesting a pathway by which REM loss may affect emotion and cognition via thalamocortical circuits. The DMN, central to spontaneous cognition, emotion, and social processing, showed particular vulnerability: late-night sleep deprivation produced the greatest reduction in DMN–DMN connectivity, indicating that depriving sleep during the REM-rich late-night period is more detrimental than early-night deprivation. These network-level changes provide neural signatures of REM sleep loss and help explain how insufficient REM may contribute to cognitive-emotional dysfunctions linked to psychiatric risk. The broad distribution of predictive contributions across networks and the prominence of visual and auditory cortices further indicate widespread network reorganization under REM loss.

Insufficient REM sleep disrupts dynamic reorganization of resting-state functional brain networks, most notably within the DMN, with significant contributions from thalamic and sensory cortices. Using CPM, the study identified a REM sleep connectome that predicts individual REM duration and reveals consistent group-level alterations under split-night deprivation, with late-night deprivation exerting stronger detrimental effects. Future work should couple real-time REM measures with fMRI to capture dynamic REM-related connectivity and examine total REM deprivation to refine brain–behavior predictive models and clinical translation.

The study focused on brain connectivity related to REM sleep without linking changes to behavioral outcomes (e.g., memory, cognition). NREM deprivation effects were not analyzed in the CPM framework.