Artificial Intelligence-Driven Multidimensional Phenotyping of Gut Metabolic States for Personalized Prebiotic, Probiotic, and Postbiotic Strategies

Abstract

This study presents multidimensional physicochemical phenotyping of human gut metabolic states using integrated multimodal profiling and machine learning. A total of 680 stool samples were analyzed using time-resolved optical, electrochemical, acoustic, magnetic, and spectral measurements implemented in a compact 3D-printed screening platform. Multivariate analysis explained 78.4% of structured variance, and cluster optimization identified seven primary clusters and 26 structurally retained subclusters representing proteolytic, saccharolytic, bile/lipid-rich, oxidative, diarrheal, pigment-linked, and normobiotic-like profiles. Clinical categories were linked to 21 subclusters after descriptor-based structure definition. A separate supervised layer assessed out-of-fold reproduction of fixed primary CL1-CL7 assignments, achieving 81.7% accuracy and a 0.799 macro F1-score in repeated cross-validation. Separately, cohort-internal diagonal category-to-subcluster distribution reached 84.1% (572/680 diagonal assignments, Wilson 95% CI: 81.2%–86.7%). 16S rRNA/laboratory-marker profiling supported the physicochemical structure, showing high diversity and low dysbiosis in the normobiotic-like phenotype, acidic high-diversity behavior in the saccharolytic phenotype, alkaline proteolytic behavior with elevated phenols and ammonia, bile-associated functional enrichment, and high dysbiosis with reduced diversity in the pigment-rich phenotype. SHAP and LIME attributed CL1-CL7 assignment to coordinated multimodal contributions. The method represents standardized extractable stool matrix profiling associated with microbiome-supported functional states, not a stand-alone diagnostic test.