Fig. 4

a Number of nearest healthy neighbors chosen in the CLOUD test prediction to find outliers in an international cohort (the HMP data set). We randomly selected a test dataset of 50 subjects and randomly selected a training dataset of 100 subjects 30 times in the full dataset of 200 subjects. We repeated the random selection of the training dataset 30 times. We did not identify outliers, excepting for extreme value of k in several random training datasets. This analysis demonstrates the robustness of the CLOUD test to the neighborhood size. The vertical bar represents 5% of the training dataset, the default neighborhood size for the test. b The same analysis was performed for the Global gut data set. c Principal coordinate plot of the Global gut dataset, from Unweighted UniFrac distance, demonstrating that the CLOUD test is robust to strong clustering effects with a reference group