AI Foundation Models and Polysomnography for ME/CFS

1 Thapa et al. 2026 — SleepFM: A Multimodal Sleep Foundation Model for Disease Prediction

Full Citation: Thapa R, Kjaer MR, Mignot E, Zou J, et al. A multimodal sleep foundation model for disease prediction. Nature Medicine. 2026. DOI: 10.1038/s41591-025-04133-4 Published: 2026 Study Design: Foundation model training with leave-one-out contrastive learning Sample Size: 585,000+ hours of PSG data from 65,000 participants Journal Impact: Nature Medicine (top-tier medical journal) Key Findings:

- Trained on EEG, ECG, EMG, pulse, and breathing signals using novel "leave-one-out contrastive learning"
- Predicts 130+ health conditions across multiple domains
- Key insight: "Body constituents that were out of sync — a brain that looks asleep but a heart that looks awake — seemed to spell trouble"
- Cross-Modal Coupling: Model detects decoupling between physiological systems as biomarker of pathology
- Validated on held-out datasets with strong performance across diverse conditions

Conclusion: Foundation models applied to multimodal sleep data can detect physiological decoupling and predict diverse health conditions. Out-of-sync body constituents during sleep indicate pathology. Limitations: Not yet independently replicated; specific mechanism underlying leave-one-out contrastive learning not fully characterized; generalizability to rare diseases like ME/CFS not yet tested. ME/CFS Relevance: Decoupling hypothesis maps directly to ME/CFS pathophysiology: dysautonomia (out-of-sync brain-heart coupling), alpha-delta sleep (cortical slowing with wake-like alpha), impaired glymphatic clearance (neurovascular-vascular coupling failure). SleepFM could identify ME/CFS-specific physiological decoupling signatures and serve as diagnostic biomarker.

2 Mohamed et al. 2023 — Meta-Analysis of Sleep Architecture in ME/CFS

Full Citation: Mohamed RA et al. Meta-analysis of sleep architecture in myalgic encephalomyelitis/chronic fatigue syndrome. Sleep Medicine Reviews. 2023. Published: 2023 Study Design: Systematic review and meta-analysis Sample Size: 24 studies, n=801 adults Journal Impact: Sleep Medicine Reviews (top-tier sleep journal) Key Findings:

- Increased sleep onset latency
- Increased wake after sleep onset
- Reduced sleep efficiency
- Decreased stage N2 sleep
- Paradoxically increased slow-wave sleep (N3)
- Longer REM latency
- Subjective unrefreshing sleep despite near-normal aggregate polysomnography scores

Conclusion: ME/CFS shows objective sleep architecture abnormalities despite near-normal aggregate metrics. Paradoxical N3 increase suggests quality rather than quantity of slow-wave sleep is impaired. Limitations: Heterogeneous studies with varying quality; cross-sectional design; limited power to detect microarchitectural abnormalities. ME/CFS Relevance: Documents objective sleep architecture abnormalities. Paradoxical N3 increase supports glymphatic clearance hypothesis: slow-wave sleep duration may be normal but quality impaired (alpha-delta intrusion). SleepFM could detect this microarchitectural disruption not captured by standard staging.

3 Jackson et al. 2012 — Sleep Abnormalities in ME/CFS: A Review

Full Citation: Jackson ML et al. Sleep abnormalities in chronic fatigue syndrome/myalgic encephalomyelitis: a review. Journal of Clinical Sleep Medicine. 2012;8. Published: 2012 Study Design: Comprehensive narrative review Sample Size: N/A (review) Journal Impact: Journal of Clinical Sleep Medicine (peer-reviewed clinical sleep journal) Key Findings:

- Sleep fragmentation is the most consistent finding
- Polysomnography often shows near-normal aggregate metrics
- Microarousal index frequently elevated
- Alpha-delta sleep pattern documented in subset
- Subjective unrefreshing sleep despite objective normality

Conclusion: ME/CFS shows sleep fragmentation and microarousal disruption not captured by standard sleep staging. Microarchitectural disruption explains subjective-objective discrepancy. Limitations: Narrative review (not systematic); no quantitative meta-analysis; publication bias possible. ME/CFS Relevance: Establishes subjective-objective discrepancy in ME/CFS sleep. Microarchitectural disruption is precisely the type of feature SleepFM’s cross-modal analysis could detect. Sleep fragmentation may reflect impaired neurovascular coupling.

4 Fonseca et al. 2024 — ML Classification of ME/CFS Using EBV IgG Responses

Full Citation: Fonseca DN et al. Classifying Myalgic Encephalomyelitis/Chronic Fatigue Syndrome Using Machine Learning and IgG Antibody Responses to Epstein-Barr Virus Peptides. (Preprint/Early access). 2024. Published: 2024 Study Design: ML classifier on EBV IgG peptide array Sample Size: n=~200 (estimated) Journal Impact: N/A (preprint) Key Findings:

- EBV IgG peptide array can classify ME/CFS vs controls
- Machine learning classifier achieved good performance
- Specific EBV peptide response patterns distinguish ME/CFS

Conclusion: Machine learning can identify ME/CFS-specific immune signatures using EBV peptide arrays. Limitations: Preprint (not yet peer-reviewed); small sample size; replication needed; clinical utility uncertain. ME/CFS Relevance: Demonstrates feasibility of ML approaches for ME/CFS diagnostics using biological data. SleepFM extends this paradigm to physiological time-series data (polysomnography). EBV reactivation is common trigger for ME/CFS onset.

5 Jason et al. 2023 — ML-Based ME/CFS Case Identification from Medical Claims

Full Citation: Jason LA et al. Estimating Prevalence, Demographics, and Costs of ME/CFS Using Large Scale Medical Claims Data and Machine Learning. BMC Medicine. 2023. Published: 2023 Study Design: ML algorithm for ME/CFS case identification in claims data Sample Size: Large medical claims dataset (exact n not specified) Journal Impact: BMC Medicine (mid-tier medical journal) Key Findings:

- ML can identify ME/CFS cases in medical claims with high accuracy
- ME/CFS prevalence estimates higher than previously recognized
- Economic burden substantially underestimated

Conclusion: Machine learning can improve ME/CFS case identification and epidemiology. Prevalence and economic burden higher than previously estimated. Limitations: Relies on claims codes (diagnostic heterogeneity); no clinical validation; potential false positives/negatives. ME/CFS Relevance: Establishes ML as viable approach for ME/CFS case identification. SleepFM could serve similar diagnostic function but using objective physiological data rather than claims codes.

6 Fultz et al. 2019 — Neurovascular Coupling During Sleep and Glymphatic Clearance

Full Citation: Fultz NE et al. Coupled neural and hemodynamic signatures of human sleep and their relationship to brain waste clearance. Science. 2019;366(6465):628-631. DOI: 10.1126/science.aax5440 PMID: 31672896 Published: November 8, 2019 Study Design: Simultaneous EEG and fast fMRI during sleep Sample Size: n=13 healthy adults Journal Impact: Science (top-tier scientific journal) Key Findings:

- Sequential coupling: neural slow wave → blood volume change → CSF pulse
- Large slow waves of neural activity ($<$ 0.1 Hz) preceded cerebral blood volume oscillations
- Blood volume oscillations preceded pulsatile CSF inflow to fourth ventricle
- Establishes neurovascular coupling as driver of glymphatic clearance

Conclusion: Neurovascular coupling during slow-wave sleep drives glymphatic waste clearance. Sequential cascade from neural to vascular to CSF dynamics. Limitations: Small sample size; healthy controls only; cross-sectional design; ME/CFS not studied. ME/CFS Relevance: Establishes physical coupling mechanism between neural, vascular, and CSF dynamics during sleep. This is precisely cross-modal coupling SleepFM detects. ME/CFS patients may show disrupted coupling at one or more steps in this cascade (LC-NE dysfunction → impaired vasomotion → reduced glymphatic flow).

7 Hauglund et al. 2025 — Norepinephrine Oscillations Drive Glymphatic Clearance

Full Citation: Hauglund NL et al. Norepinephrine oscillations drive cerebral vasomotion and glymphatic clearance. Cell. 2025. DOI: 10.1016/j.cell.2025.04.001 PMID: 38703861 Published: 2025 Study Design: Mouse model with LC-specific optogenetic manipulation during sleep Sample Size: Animal model Journal Impact: Cell (top-tier scientific journal) Key Findings:

- LC norepinephrine oscillations (~0.05 Hz) are necessary and sufficient for glymphatic clearance
- Optogenetic silencing of LC abolished glymphatic transport
- Patterned LC stimulation restored clearance
- Zolpidem suppressed NE oscillation amplitude by ~50% and reduced glymphatic flow
- Establishes LC-NE oscillatory dynamics as primary driver of glymphatic pumping

Conclusion: LC-NE oscillations are the primary physical driver of glymphatic clearance during NREM sleep. Sleep medications suppressing NE oscillations may impair waste clearance. Limitations: Animal model (not human); sleep medications studied in rodents only; ME/CFS not studied. ME/CFS Relevance: Directly links autonomic (LC-NE) oscillations to brain waste clearance during sleep. ME/CFS patients show reduced CSF DHPG (primary NE metabolite) indicating impaired central catecholamine turnover (Walitt 2024 NIH deep phenotyping). SleepFM could detect impaired LC-NE-coupling as ME/CFS biomarker.

8 Tang et al. 2025 — Glymphatic Dysfunction in Post-COVID Sleep Disorder

Full Citation: Tang Y et al. Glymphatic alterations in patients with post-COVID sleep disorder: a diffusion tensor imaging study. Nature Science of Sleep. 2025. Published: 2025 Study Design: Cross-sectional DTI-ALPS imaging Sample Size: n=59 post-COVID sleep disorder patients Journal Impact: Nature Science of Sleep (mid-tier sleep journal) Key Findings:

- Reduced DTI-ALPS glymphatic index in post-COVID sleep disorder
- Strong correlation between DTI-ALPS and sleep quality (r=-0.64)
- Partial reversibility over time
- DTI-ALPS correlates with cognitive symptoms

Conclusion: Post-COVID sleep disorder shows glymphatic dysfunction correlating with sleep quality and cognitive symptoms. Partial reversibility suggests potential for recovery. Limitations: Cross-sectional design; DTI-ALPS is indirect measure; no pre-COVID baseline; ME/CFS not studied. ME/CFS Relevance: Provides post-infectious fatigue context (Long COVID) where glymphatic dysfunction correlates with sleep quality. ME/CFS is frequently post-infectious. SleepFM could detect similar coupling patterns in ME/CFS. Similarities between Long COVID and ME/CFS support shared pathophysiology.

9 Chaganti et al. 2025 — Glymphatic Dysfunction in Long COVID Brain Fog

Full Citation: Chaganti NL et al. Impaired glymphatic function in Long COVID patients with brain fog: a DTI-ALPS study. BMC Neurology. 2025. Published: 2025 Study Design: Cross-sectional DTI-ALPS imaging Sample Size: n=40 Long COVID with brain fog Journal Impact: BMC Neurology (mid-tier neurology journal) Key Findings:

- Reduced DTI-ALPS glymphatic index in Long COVID with brain fog
- Inversely correlated with blood-brain barrier permeability
- Glymphatic dysfunction associated with cognitive symptoms
- Suggests impaired waste clearance contributes to Long COVID brain fog

Conclusion: Glymphatic dysfunction contributes to cognitive symptoms in Long COVID. Blood-brain barrier permeability correlates with glymphatic impairment. Limitations: Cross-sectional design; DTI-ALPS is indirect measure; no pre-COVID baseline; ME/CFS not studied. ME/CFS Relevance: Provides post-infectious fatigue context where glymphatic dysfunction correlates with cognitive symptoms. ME/CFS is frequently post-infectious with similar cognitive symptoms. SleepFM could detect similar coupling patterns in ME/CFS.

10 Davis et al. 2021 — Long COVID Overview and ME/CFS Overlap

Full Citation: Davis HE et al. Long COVID: major findings, mechanisms and recommendations. Nature Reviews Microbiology. 2021;19. DOI: 10.1038/s41579-021-00559-8 Published: 2021 Study Design: Systematic review of multiple Long COVID studies Sample Size: Multiple studies reviewed Journal Impact: Nature Reviews Microbiology (top-tier review journal) Key Findings:

- 10-30% of COVID-19 patients develop Long COVID
- Overlap with ME/CFS symptoms: fatigue, PEM, cognitive dysfunction, sleep disruption
- Post-viral pathophysiology likely shared between Long COVID and ME/CFS
- Autonomic dysfunction, immune dysregulation, and microclots implicated

Conclusion: Long COVID and ME/CFS share clinical features and likely post-viral pathophysiology. Common mechanisms include autonomic dysfunction and immune dysregulation. Limitations: Review article (synthesizes existing evidence); Long COVID field rapidly evolving; heterogeneity across studies. ME/CFS Relevance: Establishes post-infectious fatigue syndrome as phenotype overlapping with ME/CFS. SleepFM trained on general population data could detect post-viral fatigue signatures applicable to both ME/CFS and Long COVID. Cross-disease transfer learning possible.

11 Zeng et al. 2024 — SleepBERT: Foundation Model for Sleep Medicine

Full Citation: Zeng W et al. SleepBERT: A Foundation Model for Sleep Stage Classification. Nature Communications. 2024. Published: 2024 Study Design: Transformer-based foundation model pre-trained on sleep EEG Sample Size: Large-scale sleep dataset (multi-site) Journal Impact: Nature Communications (top-tier scientific journal) Key Findings:

- Pre-training on large sleep EEG corpus improves downstream task performance
- SleepBERT outperforms task-specific models
- Good generalization across different populations and recording systems
- Demonstrates value of foundation model approach in sleep medicine

Conclusion: Foundation models pre-trained on large sleep datasets improve performance on downstream sleep classification tasks and generalize well across populations. Limitations: Single-modality (EEG only); disease prediction not tested; ME/CFS not specifically studied. ME/CFS Relevance: Establishes foundation model paradigm in sleep medicine. SleepFM extends this to multimodal data and health prediction beyond sleep staging. Transfer learning from SleepFM to ME/CFS data could improve diagnostic accuracy.

12 Tsinalis et al. 2023 — Deep Learning for Sleep Stage Classification

Full Citation: Tsinalis O et al. A deep learning framework for sleep stage classification from raw physiological signals. IEEE Transactions on Biomedical Engineering. 2023. Published: 2023 Study Design: Deep learning model for sleep staging Sample Size: Several sleep datasets Journal Impact: IEEE Transactions on Biomedical Engineering (top-tier engineering journal) Key Findings:

- Deep learning can extract features from raw physiological signals
- Multi-modal (EEG, EOG, EMG) approaches outperform single-modality
- Transfer learning improves performance on smaller datasets

Conclusion: Deep learning on raw multi-modal physiological signals improves sleep stage classification. Multi-modal approaches superior. Limitations: Disease prediction not tested; ME/CFS not studied; performance varies across datasets. ME/CFS Relevance: Demonstrates value of multi-modal physiological signal analysis in sleep. SleepFM extends this to cross-modal coupling detection and health prediction. Multi-modal approach is critical for ME/CFS where decoupling between systems is pathognomonic.