Fig. 4.

Forest plots of each prediction performance metric (R-squared – Panel A, Spearman correlation – Panel B) for each time point (6 weeks (n = 158), 12 months (n = 282)) across all 36 metabolites and 4 machine learning models. 95% credible interval and predictive posterior means were generated using Bayesian modelling of the evaluation statistic (Methods) after 100 repeats of 5-fold nested cross validation. Red vertical lines indicate a value of 0 for the evaluation metric (equivalent to null model). Metabolites were classified as predictable if the null value did not lie within the estimated 95% credible interval. For most metabolites, predictive performance was not significantly better than null models