Predictive Metabolomics and Machine Learning Classification of Hepatotoxic Traditional Medicine Exposures Using Urine Biomarker Signatures in a Nigerian Population Cohort

Hepatotoxicity from traditional medicine use is a clinically significant but chronically underdiagnosed condition in Nigerian healthcare settings, where patients frequently do not disclose herbal medicine use to clinical providers. Urine metabolomics offers a non-invasive approach to identifying biomarker signatures of herb-induced hepatotoxicity that could be translated into diagnostic screening tools. This study developed and validated a predictive metabolomics and machine learning classification model for hepatotoxic traditional medicine exposure using urine biomarker signatures in a prospective Nigerian population cohort. A total of 480 participants were enrolled across three clinical sites in Enugu, Kano, and Port Harcourt, categorised as confirmed hepatotoxic TM users, non-hepatotoxic TM users, and non-TM users based on clinical, biochemical, and detailed exposure history. Urine samples were analysed by untargeted UHPLC-HRMS metabolomics. After multivariate statistical processing and batch correction, metabolite features were subjected to random forest and gradient boosting classification model development with nested cross-validation. A panel of twelve urinary biomarker metabolites including pyrrolizidine alkaloid metabolites, modified bile acid conjugates, and methylhistidine species accurately classified hepatotoxic TM exposure with area under ROC curve of 0.91 in the external validation cohort. Three metabolites previously unreported as hepatotoxicity biomarkers were structurally characterised by MS2 fragmentation and NMR. The validated metabolomics classifier provides an original diagnostic contribution with potential for integration into liver disease screening algorithms applicable to Nigerian clinical and community settings. Keywords: metabolomics, hepatotoxicity, traditional medicine, machine learning, urine biomarkers.