Leveraging Explainable Automated Machine Learning (AutoML) and Metabolomics for Robust Diagnosis and Pathophysiological Insights in Myalgic Encephalomyelitis/Chronic Fatigue Syndrome (ME/CFS).
Yagin, Fatma Hilal, Colak, Cemil, Al-Hashem, Fahaid et al. · Diagnostics (Basel, Switzerland) · 2025 · DOI
Quick Summary
Researchers used advanced computer learning and blood tests to find chemical markers that could help diagnose ME/CFS. They analyzed 888 different chemicals in blood samples from 106 ME/CFS patients and 91 healthy people, and found a pattern of chemical imbalances that correctly identified ME/CFS patients 87% of the time. The key chemical imbalances involved energy production in cells, inflammation, and gut health—areas that have long been suspected in ME/CFS.
Why It Matters
This research addresses a critical clinical need—ME/CFS currently lacks objective diagnostic biomarkers, forcing clinicians to rely on symptom-based diagnosis. A validated metabolic signature could improve diagnostic accuracy and speed of diagnosis, reducing the diagnostic odyssey many patients experience. Additionally, identifying dysregulated metabolic pathways provides mechanistic insights that could inform the development of targeted treatments.
Observed Findings
TPOT achieved 92.1% AUC, 87.3% accuracy, 85.8% sensitivity, and 89.0% specificity in distinguishing ME/CFS from controls.
Key dysregulated metabolites included succinic acid and pyruvic acid (mitochondrial energy metabolism), prostaglandin D2 and 11,12-EET (inflammation), glycocholic acid (gut-brain axis), and PC(35:2)a (cell membrane integrity).
PLS-DA analysis showed statistically significant but moderate discrimination between ME/CFS and control groups.
Permutation testing and cross-validation confirmed the robustness of identified features.
Inferred Conclusions
Plasma metabolomic profiles contain sufficient biological information to support accurate ME/CFS diagnosis using machine learning.
ME/CFS pathophysiology involves dysregulation across multiple interconnected systems: energy metabolism, inflammatory signaling, gut-brain communication, and cell membrane function.
Explainable AI methods can extract clinically meaningful insights from complex omics data, supporting both diagnostic and therapeutic discovery.
Remaining Questions
Will these findings validate in independent, prospectively collected patient cohorts from different populations and geographic regions?
Do these metabolic alterations persist longitudinally in individual patients, or do they fluctuate with disease activity or clinical phenotype?
What This Study Does Not Prove
This study does not prove these metabolites *cause* ME/CFS or that they are specific to ME/CFS alone; they may be altered in other conditions. The cross-sectional design cannot establish temporal relationships or causality. The findings require validation in independent cohorts and real-world clinical settings before the model could be used in clinical practice.