A metabolomics and machine learning-based predictive model for AKI-to-CKD progression risk

Progression from acute kidney injury (AKI) to chronic kidney disease (CKD) increases mortality risk, yet reliable tools for early risk stratification remain limited. We performed untargeted liquid chromatography-mass spectrometry (LC–MS)-based plasma metabolomics in 74 participants, including 44 AKI-nonCKD patients with renal function recovery, 19 AKI-to-CKD progressors, and 11 healthy controls. A two-step differential analysis identified 104 consensus metabolites showing concordant directional changes between AKI-nonCKD and AKI-to-CKD groups. These metabolites were subsequently used as a predefined candidate metabolite pool for predictive modeling, with downstream Boruta screening and LASSO modeling performed within a nested cross-validation framework. A candidate 5-metabolite signature consisting of 2-Furoate, 1H-Pyrrole-2-carboxaldehyde, LPE(18:0), Cyclopentylamine, and Glycerophosphocholine was selected based on repeated LASSO retention frequency and model parsimony analysis. The 5-metabolite model showed comparable discrimination with a numerically higher AUC than the 4-metabolite model (0.981 vs. 0.957), whereas adding a sixth metabolite did not further improve discrimination. In nested cross-validation, the clinical-only model showed limited discrimination, while the metabolite-only and integrated LASSO models achieved substantially higher pooled AUCs in repeated nested cross-validation, with pooled AUCs of 0.999 and 0.999, respectively. Label-permutation negative control analysis reduced model performance to near-chance level, supporting that the observed discrimination was not generated by the resampling procedure alone. These findings suggest that plasma metabolomic signatures may provide candidate metabolomic predictors for early AKI-to-CKD risk stratification, although external validation in multi-center cohorts is required before clinical application.