SleepFM Cross-Modal Decoupling Validation in ME/CFS
1 Background and Rationale
SleepFM (Thapa et al., 2026, Nature Medicine) is a multimodal AI foundation model trained on 585,000+ hours of polysomnography from 65,000 participants that predicts 130+ future health conditions from one night’s sleep data (Thapa et al. 2026). The model’s core finding — that cross-modal physiological decoupling (“a brain that looks asleep but a heart that looks awake”) is the strongest disease predictor — maps onto the constellation of sleep architecture abnormalities documented in ME/CFS (Section Post-Exertional Malaise May Involve Inflammation-Induced Routing Disruption of Brain Clearance). This external validation across 130+ conditions (C-index >0.80 for many) provides a rationale for investigating whether the decoupling observed in ME/CFS is pathophysiologically significant, though SleepFM’s observational design cannot distinguish causal mechanisms from epiphenomena in any specific condition. No study has applied SleepFM or any equivalent multimodal sleep foundation model to ME/CFS polysomnography data. This proposed study would be the first to test whether cross-modal decoupling signatures can distinguish ME/CFS from healthy controls and from clinically relevant comparator conditions (fibromyalgia, Long COVID brain fog, idiopathic hypersomnia, major depression).
2 Study Design
Design: Retrospective case-control analysis of existing polysomnography datasets, followed by prospective validation. Phase 1 — Retrospective Discovery:
Cohort: Aggregate existing ME/CFS PSG datasets: Mohamed et al. 2023 meta-analysis dataset (24 studies, n=801 adults) (Mohamed et al. 2023), Bateman Horne Center clinical PSG archive, NIH deep phenotyping study PSG subset (Walitt et al. 2024), and the Stanford Sleep Medicine Center cohort (35,000 patients, 1999–2024, with linked electronic health records) (Thapa et al. 2026).
Analysis: Apply SleepFM or a comparable leave-one-out contrastive learning architecture to raw PSG signals. Extract multimodal reconstruction error (decoupling score) for each recording. Compare decoupling scores between ME/CFS patients and age-/sex-matched controls.
Primary endpoint: AUC for ME/CFS vs. healthy control discrimination using decoupling score alone, independent of standard sleep staging metrics.
Secondary endpoints: (a) Decoupling score correlation with symptom severity (fatigue, unrefreshing sleep, cognitive dysfunction). (b) Comparison of decoupling patterns across ME/CFS subtypes (sleep-predominant, brain fog-predominant, PEM-predominant). (c) Decoupling scores in ME/CFS vs. fibromyalgia vs. Long COVID brain fog to assess disease specificity. Phase 2 — Prospective Validation:
Cohort: De novo recruitment of 100 ME/CFS (Fukuda + CCC criteria), 100 age-/sex-matched healthy controls, 50 fibromyalgia (ACR criteria), 50 Long COVID with brain fog, and 50 idiopathic hypersomnia.
Procedure: One-night attended PSG with standard clinical montage (EEG, EOG, EMG, ECG, respiratory inductance plethysmography, pulse oximetry). Morning symptom questionnaires. SleepFM decoupling analysis.
Primary endpoint: Sensitivity >0.85, specificity >0.90 for ME/CFS vs healthy controls using a pre-registered decoupling threshold from Phase 1.
Secondary endpoints: Subtype classification accuracy. Treatment-response prediction (baseline decoupling score predicts response to 4-week trazodone 25–50 mg at bedtime). Phase 3 — Longitudinal PEM Prediction:
Subset: 50 ME/CFS patients from Phase 2 undergo 7-night home PSG (simplified montage: EEG lead, ECG patch, respiratory band, pulse oximeter) with daily symptom diaries and activity monitoring.
Primary endpoint: Within-subject worsening of decoupling score (>25% increase from individual baseline over 3 nights) predicts PEM onset within 48 hours with positive predictive value >0.75.
Secondary endpoint: Day 2 post-CPET decoupling score vs. Day 1 pre-CPET score.
3 Feasibility Assessment
Strengths:
Most data already exist (Mohamed 2023 meta-analysis, Stanford cohort, Bateman Horne Center archive) — Phase 1 is computationally inexpensive and requires no new data collection.
Analysis methods are published (leave-one-out contrastive learning, multimodal reconstruction error).
Comparator conditions are clinically relevant and address differential diagnosis needs.
Home monitoring extension (Phase 3) uses consumer-accessible hardware. Limitations:
SleepFM model weights and architecture are not publicly available as of 2026; the study requires either collaboration with the Zou/Mignot lab or independent reimplementation of the leave-one-out contrastive learning architecture.
PSG montages, hardware, and scoring conventions vary between historical datasets, introducing systematic variance that may mask biological signal.
Severe/bedbound ME/CFS patients cannot attend sleep laboratories; Phase 3 home monitoring partially addresses this but with reduced signal quality.
The training dataset (clinical PSG referrals) may differ systematically from ME/CFS research cohorts in demographic and comorbidity profiles.
4 Expected Impact
Scientific Impact: First quantitative test of the cross-modal decoupling hypothesis in ME/CFS. If confirmed, decoupling becomes a candidate mechanistic biomarker bridging autonomic, sleep, and glymphatic research domains. The subtype analysis may identify distinct decoupling endotypes that explain heterogeneous treatment responses. Clinical Impact: An objective sleep-based diagnostic biomarker would reduce the 5-year average diagnostic delay for ME/CFS. PSG is a standardised, reimbursed clinical procedure available in any accredited sleep laboratory. A validated software-based biomarker eliminates reagent costs and inter-laboratory assay variability. Patient Community Impact: Provides an objective measure of the unrefreshing sleep that patients consistently report but that conventional sleep staging dismisses as “normal PSG.” Quantifies the physiological toll of PEM through sleep-based metrics.