Pipeline-optimized machine learning for chronic fatigue syndrome diagnosis: A lightweight, interpretable model using blood biochemical and metabolomic data.

Observed Findings

• A 10-biomarker model achieved 93.9% diagnostic accuracy with ROC-AUC of 0.979 when discriminating CFS from both healthy controls and overlapping conditions
• Six biomarkers (urea, total protein, glucose, total bilirubin, leucine, vitamin D) showed causal associations with CFS in Mendelian randomization analysis
• Elevated glucose and leucine levels were associated with exacerbated CFS symptoms in SHAP and correlation analyses
• Stratified bootstrap sampling with case-control matching on covariates improved model robustness compared to traditional random sampling
• Decision curve analysis demonstrated clinical utility across wide decision threshold range (0.01–0.98)

Inferred Conclusions

• Systematic pipeline optimization from control selection through biomarker identification produces lightweight, interpretable diagnostic models superior to previous ad hoc approaches
• Objective blood biomarkers can effectively discriminate ME/CFS from both healthy individuals and symptomatically overlapping conditions
• Combining multiple interpretability frameworks (SHAP, Mendelian randomization, decision curve analysis) provides actionable mechanistic insights for personalized risk assessment
• The ClinMetML framework approach ensures reproducibility and provides a standardized methodology for future ME/CFS biomarker studies

Remaining Questions