Machine learning-based prediction of ICU admission and mortality in Crimean–Congo hemorrhagic fever by using wide-range targeted metabolomics

PURPOSE: This study aimed to evaluate the potential of amino-acid profiles to predict disease progression in patients with Crimean-Congo Hemorrhagic Fever (CCHF) and to identify metabolic biomarkers associated with clinical outcomes and survival. METHODS: Of the 115 confirmed CCHF patients, 18 required intensive care unit (ICU) admission and 16 died. Notably, 15 of the deaths occurred among ICU patients, whereas only one death occurred outside the ICU. For each patient, 32 amino acid concentrations were used as input for machine-learning (ML) models. RESULTS: Among the classification models evaluated for predicting ICU admission, XGBOOST and LASSO achieved the highest performance, each with an AUC of 0.958. Arginine and glutamic acid consistently emerged as the most predictive features across all models, followed by 1-methyl-L-histidine, tryptophan, and tyrosine, which appeared among the top variables in four of the five best-performing models. In survival analysis, the mean concordance index (and integrated Brier score) was 0.973 (0.10) for Survival LASSO, 0.971 (0.11) for RFSRC, and 0.942 (0.12) for Survival XGBOOST. In survival models, the top five amino acids contributing to predictions were ornithine, gamma-aminobutyric acid, ethanolamine, arginine, and histidine. CONCLUSION: ML models based on amino-acid profiles can accurately predict disease progression in CCHF, supporting early risk stratification and providing insights into the metabolic mechanisms underlying disease severity.