A data-driven universal gut microbiome health assessment: a machine learning framework trained on large metagenomic data.

The gut microbiota is essential to maintain host physiology, and its disruption (dysbiosis) is associated with a wide range of diseases. Machine learning (ML) offers a powerful tool to model species-level microbiome profiles, but classifiers that reliably separate healthy from diseased individuals across independent cohorts are still lacking. In this study, we developed a ML classifiers trained on 7,452 publicly available stool metagenomes spanning 32 studies and 12 diseases, designed to distinguish healthy individuals (absence of a clinically diagnosed disease) from non-healthy individuals (presence of a clinically diagnosed disease) based on species-level gut microbiome profiles. We trained 16 supervised models combining four algorithms (RF, SVM-LIN, SVM-RBF, and LR-ElasticNet) combined with all feature sets and three feature-selection algorithms. Performance was assessed by F1 score and ROC-AUC on held-out test data and externally validated on 642 samples from six independent cohorts, including previously unseen diseases. On the test set, all models achieved F1 scores of 78-86% and ROC-AUC values of 89-95%. An SVM-RBF model using permutation-based feature selection performed best (F1 = 86.6%, ROC-AUC = 95.5%; healthy F1 = 86.6%, non-healthy F1 = 88.7%). Importantly, external validation confirmed the generalizability of the full-feature SVM-RBF model (overall F1 = 70.6%; ROC-AUC = 84.7%), including unseen disease types such as