Fig. 4

Predictive modeling of CDI recurrence achieved the highest accuracy using metabolomic data. The performance of predictive models was assessed using leave-one-out cross-validation (N = 26 at pre-CDI treatment, N = 48 at week 1, and N = 40 at week 2). Data sources input to models were (1) clinical variables associated with recurrence in prior studies (age, previous PPI use, antibiotic treatment regimen, and CDI diagnostic test used), (2) untargeted gut metabolomics, (3) amplicon sequencing variants (ASVs) of the gut microbiome, (4) gut short-chain fatty acids (SCFAs), (5) data sources 1–4 combined. Performance of A logistic regression with lasso and B random forests, which predict binary labels (recurrence/no recurrence), were assessed with the area-under-the-curve (AUC) metric. C Cox regression, which predicts survival time, was assessed with the concordance index (CI). Models achieving median ≥ 0.70 AUC or CI scores (adequate performance) are denoted with red dashed rectangles. The “All Data” models with ≥ 0.70 AUC or CI were found to select only metabolomic features