The time window from stroke onset is critical for the treatment decision. However, in unknown onset stroke, it is often difficult to determine the exact onset time because of the lack of assessment methods, which can result in controversial and random treatment decisions. Previous studies have shown that serum biomarkers, in addition to imaging assessment, are useful for determining the stroke onset time. However, as yet there are no specific biomarkers or corresponding methodologies that are accurate and effective for determining the onset time of unknown onset stroke. Herein, we describe our novel advanced metabolites-based machine learning method (termed extreme gradient boost [XGBoost]) combined with recursive feature elimination, which accurately screened five metabolites from 1124 metabolites detected in serum. These metabolites were capable of both detecting acute ischemic stroke and backtracking the acute ischemic stroke onset time. To further investigate the pathological mechanisms of acute ischemic stroke, we also examined characteristic metabolites in different brain regions, and found two metabolites that could distinguish the core infarct area from the ischemic penumbra. Although this study is based on animal experiments, our machine learning framework and selected metabolites may provide a basis for clinical stroke evaluation and treatment.
Keywords: Machine learning; Metabolites; Onset time; Stroke; XGBoost.
Copyright © 2022 The Authors. Published by Elsevier Masson SAS.. All rights reserved.