The smart gut report shows if a genus is high, low, or normal. These limits are defined as 0.5% and 99.5%-tiles respectively (the middle 99% are normal). They were determined from distributions of genus measured from a healthy cohort. These limits are defined in their plos paper.
The problem with their approach, and this leads to the false alarms, is that many of the distributions are very close to exponential with very small lambdas – many of them have none bacteria tested for. Another way to say it is that for a given genus, many folks have none of the bacteria. Therefore, if you would look at the smart gut report for the healthy cohort, they would report a low value of this bacteria.
A concrete example, look at Butyrivbrio crossotus. 88% of the healthy cohort has none of this bacteria. I used this in excel =COUNTIF(Table1[Butyrivibrio crossotus],”<0.000001")/COUNT(Table1[Butyrivibrio crossotus]). How do you define the percentiles of “normal” when 88% of the healthy cohort do not have this bacteria? How can you say the healthy cohort is “healthy” if you ran the smart gut report on them and 88% of them reported “low” for B. Crossotus?
From a machine learning point of view, they are doing kernel estimation and then creating a classifier. But, the kernels from which the data are generated contain many zero values so the classifier choice leads to a biased result giving the false alarms (this assumes the healthy cohort is in fact healthy – health was self-reported).
See paper here: http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0176555
Data here: https://doi.org/10.1371/journal.pone.0176555.s003
All in all, I think ubiome is doing a good job of pushing the science. I wish more folks would get involved looking at this data critically because I think it benefits everyone.