Background: Food allergy (FA) is one of the most common chronic diseases in childhood, with a continuously increasing prevalence. However, there is a lack of non-invasive and accurate methods for early screening. Methods: This study utilized a prospective birth cohort of 1579 mother-infant pairs, followed up until the infants were 12 months old to diagnose FA. Among these, 182 infants with complete fecal samples were included for multi-omics analysis, and the Binary Reflection Learning Optimizer (bREPLO) meta-heuristic algorithm was introduced for high-dimensional metabolomic feature screening. Further targeted tryptophan metabolomics validation was performed, and 7 machine learning models were systematically compared to construct a risk prediction model for FA. Results: Children with FA exhibited gut microbiota dysbiosis. Untargeted metabolomics indicated that the tryptophan pathway was the significantly altered pathway . A stable metabolic panel centered on indolepyruvate (IPYA) was screened out by the bREPLO algorithm. Targeted metabolomics further verified that among 30 tryptophan metabolites, only the top three Indolelactic acid (ILA), L-5-Hydroxytryptophan (5-HTP), and IPYA accounted for a cumulative contribution of 53.8%, indicating a highly focused characteristic. Conclusion: The metabolite-based decision tree model constructed in this study has the triple advantages of non-invasive sampling, clinical interpretability , and intervenable targets, providing a scientific tool with clinical translation potential for the early screening and precise diagnosis of FA.