Overview of Biomarker Development
1 Why Biomarkers Are Needed
The absence of validated biomarkers has been one of the most significant obstacles to ME/CFS recognition, research, and treatment:
- Diagnostic uncertainty: Without objective markers, diagnosis relies entirely on clinical criteria and exclusion of other conditions
- Stigmatization: Lack of measurable abnormalities has contributed to the perception of ME/CFS as a psychosomatic condition
- Research challenges: Heterogeneous patient populations (due to imprecise diagnosis) may obscure findings
- Treatment development: Drug development requires objective endpoints for clinical trials
- Disability assessment: Social security and insurance determinations benefit from objective evidence
- Subgroup identification: Biomarkers may identify pathophysiologically distinct subgroups requiring different treatments
2 Types of Biomarkers
Different biomarker types serve different purposes:
2.1 Diagnostic Biomarkers
Markers that distinguish ME/CFS from healthy individuals and from patients with other fatiguing conditions:
- High sensitivity (few false negatives)
- High specificity (few false positives)
- Practical for clinical use (accessible, affordable)
- Reproducible across laboratories
2.2 Prognostic Biomarkers
Markers that predict disease course or outcome:
- Likelihood of spontaneous improvement
- Risk of progression to more severe illness
- Long-term functional outcomes
2.3 Treatment Response Biomarkers
Markers that predict or monitor response to specific treatments:
- Baseline markers predicting treatment response
- Dynamic markers reflecting treatment effects
- Stratification markers for personalized treatment selection
2.4 Mechanistic Biomarkers
Markers that reflect underlying pathophysiology:
- May not be diagnostic but inform disease mechanisms
- Guide development of targeted therapies
- Enable subgroup classification
3 Challenges in ME/CFS Biomarker Research
Multiple factors have complicated biomarker identification:
- Case definition heterogeneity: Different diagnostic criteria capture overlapping but distinct populations
- Disease heterogeneity: ME/CFS likely encompasses multiple distinct conditions with different pathophysiology
- Illness duration effects: Biomarkers may differ between early and chronic illness
- Severity effects: Severely affected patients (often excluded from studies) may differ from ambulatory patients
- Sex differences: The NIH study demonstrated distinct abnormalities in men and women
- Comorbidities: Overlapping conditions (POTS, MCAS, fibromyalgia) may confound findings
- Small sample sizes: Many studies underpowered to detect moderate effect sizes
- Lack of replication: Few findings have been consistently replicated across laboratories
4 Convergent Biology Despite Heterogeneity
Despite the challenges listed above, emerging evidence suggests that heterogeneous upstream triggers converge on common downstream pathological pathways, with direct implications for biomarker strategy:
miRNA convergence: Cheema et al. 2023 in PLOS ONE (Cheema et al. 2023) found that circulating miRNAs, although different between individual patients, target the same specific gene cluster (\(p < 0.002\)) converging on exercise hyperemia, angiogenesis, antioxidant defenses, and mitochondrial fission. This suggests heterogeneous molecular perturbations funnel into common downstream pathology measurable through functional pathway biomarkers rather than single-molecule markers.
Metabolic convergence: Naviaux et al. 2016 in PNAS (Naviaux et al. 2016) showed that despite heterogeneous triggers, the cellular metabolic response was homogeneous—a dauer-like hypometabolic state affecting 20 pathways with 80% of diagnostic metabolites decreased and 94–96% diagnostic accuracy.
Emerging large-scale validation panels: The EpiSwitch CFS test (Hunter et al. 2025, J Transl Med) (Hunter et al. 2025) achieved 92% sensitivity and 98% specificity using a 200-marker 3D genomic profiling model in severe ME/CFS (discovery cohort performance, \(n=47\) ME/CFS vs $ 61$ controls; commercial platform, Oxford BioDynamics; no independent external validation published as of 2026 — standard discovery-cohort overfitting caveats apply), and the BioQuest project (OMF, launched 2024) (Open Medicine Foundation 2024) aims to analyze approximately 1,000 samples across metabolomics, proteomics, and cytokines with subgrouping as a primary aim. [The practical implication of these convergence patterns is that biomarker strategies should target convergent downstream pathways (energy metabolism, vascular function, immune dysregulation) rather than seeking a single upstream marker. As argued by Birch], [Younger (Birch, Younger, et al. 2025), multi-marker panels stratifying patients by endotype are more likely to succeed than individual biomarkers applied to the entire population.],
5 Machine Learning Approaches to EBV-Based Diagnosis
Fonseca et al. (2024) applied a Super Learner ensemble algorithm to IgG antibody responses against 3,054 EBV peptide variants, identifying a 26-antibody classifier that distinguished infection-triggered ME/CFS patients from healthy controls with 100% accuracy in the training dataset and 90% in the test dataset However, the classifier failed to reach target accuracy (85%) when applied to all ME/CFS patients or to those with non-infectious or unknown disease triggers. Despite some EBV peptides showing sequence homology with human proteins (potential molecular mimicry), no significant correlation emerged between antibody importance in the classifier and cross-reactivity potential with host proteins. The trigger-specificity of this classifier has two implications: diagnostically, it may prove useful as a confirmatory tool specifically for patients reporting infection-linked onset, supporting the clinical assessment of post-infectious ME/CFS. Mechanistically, it reinforces the subtyping paradigm—infection-triggered and non-infection-triggered ME/CFS appear immunologically distinct, and biomarker panels developed in one subgroup should not be assumed to generalize to the other.
6 AI-Based Approaches Using Sleep Physiology
Machine learning applied to sleep data represents a distinct biomarker category that exploits physiological time-series rather than molecular analytes. The conceptual basis is that polysomnography (PSG) captures multiple physiological systems simultaneously — brain activity (EEG), cardiac function (ECG), muscle tone (EMG), respiration, and pulse oximetry — and that the relationships between these systems, not just each in isolation, carry diagnostic information.
6.1 Multimodal Foundation Models for Sleep-Based Disease Prediction
SleepFM (Thapa et al., 2026), published in Nature Medicine, is the first large-scale multimodal sleep foundation model trained on 585,000+ hours of PSG data from 65,000 participants (Thapa et al. 2026). Using a novel leave-one-out contrastive learning approach, the model predicts the future onset of 130+ health conditions from one night of sleep data, achieving C-indices \(>\) 0.80 for Parkinson’s disease (0.89), dementia (0.85), hypertensive heart disease (0.84), myocardial infarction (0.81), prostate cancer (0.89), breast cancer (0.87), and all-cause mortality (0.84). The model’s architectural innovation is potentially relevant to ME/CFS biomarker strategy, pending application to ME/CFS data. By training to reconstruct a hidden modality from the remaining signals, SleepFM learns a latent representation of physiological coupling integrity — how well brain, heart, muscle, and respiratory systems coordinate. The key finding is that decoupling between systems (e.g., “a brain that looks asleep but a heart that looks awake”) is a stronger disease predictor than any single modality (Thapa et al. 2026).
6.2 Other AI Approaches to Sleep-Based ME/CFS Classification
The broader AI-in-sleep-medicine literature provides a methodological framework that predates and contextualises SleepFM. SleepBERT (Zeng et al., 2024), a transformer-based foundation model pre-trained on large-scale sleep EEG, demonstrated that pre-training on sleep data corpora substantially improves downstream task performance compared to task-specific models (Zeng et al. 2024). Tsinalis et al. (2023) showed that multi-modal deep learning approaches (EEG + EOG + EMG) systematically outperform single-modality approaches for sleep staging — a finding that generalises to the diagnostic domain: combining physiological signals should outperform any single channel for disease classification (Tsinalis et al. 2023). In ME/CFS specifically, machine learning has been applied to administrative (Jason et al., 2023 (Jason et al. 2023) — ML case identification from medical claims) and molecular data (Fonseca et al., 2024 — EBV IgG peptide array classification, Section Overview of Biomarker Development), demonstrating feasibility but limited by the input data type. Claims-based ML suffers from diagnostic coding inconsistency; molecular ML requires blood draws and specialised assays. Physiological ML from sleep data offers a complementary third route: objective, multi-system, standardised, and collected during a naturally occurring physiological state rather than an artificial challenge.