An Interpretable Machine Learning Model Based on Metabolomics for Predicting Plaque Burden in Cryptogenic Stroke

ABSTRACT Cryptogenic stroke represents 25%–40% of ischemic strokes, with many cases harboring unrecognized large artery atherosclerosis (LAA) requiring specific secondary prevention. In this multicenter pilot study, we developed a metabolomics‐based machine learning model to identify LAA and predict plaque burden in cryptogenic stroke patients. Plasma metabolites from 572 acute ischemic stroke patients across three hospitals were analyzed using untargeted metabolomics. A two‐stage machine learning approach was developed: Model 1 distinguished cardioembolic (CE) from non‐CE stroke, and Model 2 separated LAA from small vessel occlusion (SVO). Models were integrated to directly predict LAA among all subtypes. Model 1 achieved exceptional performance distinguishing CE from non‐CE (AUC = 0.998, accuracy = 97.9%), with pyruvate and glutamine as key discriminators. Model 2 differentiated LAA from SVO with AUC = 0.949, identifying pyroglutamate and 2‐hydroxybutyrate as primary markers. The combined model predicted LAA with AUC = 0.821. Critically, cryptogenic stroke patients predicted as LAA showed significantly higher plaque burden than those predicted as CE/SVO (23.9% vs. 10.9%, p = 0.014), with increased neovascularization and ulcerative features. This metabolomics‐based approach accurately identifies LAA in cryptogenic stroke patients and predicts atherosclerotic plaque burden, offering a novel diagnostic tool to guide personalized antithrombotic therapy selection. Trial Registration: Chinese Clinical Trial Registry registration number: ChiCTR1800015956