The performance of predictive models for three optimally identified parsimonious sets of predictors on EA treatment. Predictors include: 10 markers reflecting pathological symptoms of EAE, 8 microbial species and 8 microbial metabolites in fecal samples that were altered by EA. The random forest algorithm incorporated into a repeated double cross validation framework with unbiased variable selection was applied to effectively determine a parsimonious set of features from pathological symptoms markers of EAE mice ( n = 24), gut microbiota ( n = 65), and microbial metabolites ( n = 132) that could predict EA from EAE. In the figure, each swim lane represents one mice sample. For each sample, class probabilities were computed from 100 double cross-validations. Class probabilities are color coded by class and presented per repetition (smaller dots) and averaged over all repetitions (larger dots). Misclassified samples are circled. Predictive accuracy was calculated as a number of correctly predicted samples/total number of measured samples.
