Integrating tongue coating microbiome, tongue coating metabolomics, and exhaled breath metabolomics via machine learning to identify chronic atrophic gastritis.

BACKGROUND: Chronic atrophic gastritis (CAG) is a precancerous gastric condition with limited non-invasive diagnostic options. Alterations in the oral microbiome and its localized metabolic profiles may provide early multi-omics signatures of disease progression. METHODS: We performed 16S rRNA gene sequencing of tongue coating samples, along with untargeted metabolomics of both tongue coating and exhaled breath condensate (EBC) in 139 patients, including 99 patients with CAG and 40 with chronic non-atrophic gastritis (CNAG) serving as a disease comparison group. The tongue coating microbiome, tongue coating metabolome, and systemic breath volatile profiles were comprehensively characterized. The clinical cohort was randomly partitioned into a training set and an independent validation test set in a 7:3 ratio. Unsupervised PCA algorithm was implemented for metabolic clustering, and LASSO regression was applied for cross-domain feature integration. RESULTS: CAG patients exhibited significant tongue coating microbiome dysbiosis, characterized by the enrichment of g__Arthrobacter, Veillonella, and Streptococcus, and depletion of Prevotella and Haemophilus compared to the CNAG disease comparison group. Metabolomic profiling of the tongue coating uncovered pronounced metabolic alterations, mapped via unsupervised PCA scores (PC1: 49.5%; PC2: 5.7%), identifying 56 differential tongue coating metabolites that were predominantly upregulated and significantly enriched in alanine, aspartate, and glutamate metabolism pathways. Concurrently, EBC metabolomics mapped via PCA coordinate configurations (PC1: 18.8%; PC2: 8.2%) detected 73 differential metabolites based on unadjusted Student's t-test (p < 0.05), with a majority (56/73) being significantly downregulated, reflecting systemic metabolic variations. To bridge these high-dimensional cross-domain interactions, an optimized 26-feature panel was filtered via LASSO regression. A generalized linear Logistic Regression classifier excelled in discriminating CAG from the CNAG disease comparison group, achieving a top-tier Area Under the Curve (AUC) of 0.8667 (95% CI: 0.748-0.954), an accuracy of 78.6%, and a sensitivity of 90.0% on the independent 7:3 test set, consistently outperforming complex non-linear ensemble tree algorithms. Calibration curve and decision curve analysis (DCA) further confirmed robust fit metrics (Brier Score = 0.1634) and substantial clinical net benefit on the held-out test cohort. CONCLUSION: This study establishes an advanced, regularized linear machine learning framework, demonstrating that a streamlined 26-feature panel of tongue coating and exhale profiles offers a transparent, non-invasive triage tool for chronic atrophic gastritis. These tightly synchronized local and systemic multi-source trajectories provide crucial insights into oral-gastric microbial interactions and host-microbiome co-metabolism in CAG.