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

Area of Science:

• Neurology
• Metabolomics
• Machine Learning

Background:

• Cryptogenic stroke accounts for a significant portion of ischemic strokes.
• Many cryptogenic strokes are linked to unrecognized large artery atherosclerosis (LAA), necessitating targeted secondary prevention strategies.

Purpose of the Study:

• To develop and validate a metabolomics-based machine learning model for identifying LAA in cryptogenic stroke patients.
• To predict atherosclerotic plaque burden in these patients to guide treatment decisions.

Main Methods:

• Utilized untargeted plasma metabolomics on 572 acute ischemic stroke patients.
• Developed a two-stage machine learning model: Stage 1 distinguished cardioembolic (CE) from non-CE stroke; Stage 2 separated LAA from small vessel occlusion (SVO).
• Integrated models to directly predict LAA and assessed plaque burden in predicted LAA versus other subtypes.

Main Results:

• Model 1 achieved high accuracy (97.9%) in differentiating CE from non-CE stroke (AUC=0.998).
• Model 2 effectively distinguished LAA from SVO (AUC=0.949), with pyroglutamate and 2-hydroxybutyrate as key markers.
• The combined model predicted LAA with an AUC of 0.821. Patients predicted with LAA exhibited significantly higher plaque burden (23.9% vs. 10.9%) and more aggressive plaque features.

Conclusions:

• A metabolomics-driven machine learning approach can accurately identify LAA in cryptogenic stroke.
• The model predicts atherosclerotic plaque burden, offering a novel diagnostic tool.
• This approach can personalize antithrombotic therapy selection for cryptogenic stroke patients.