Gut Microbiome Alzheimer's Tests Fail to Cross Cohorts, Reanalysis Finds

A reanalysis of four public gut-microbiome datasets found that which lab or country a stool sample came from explained more of the bacterial variation than whether the donor had Alzheimer's disease, and standard statistical fixes did not close the gap.

A classifier that reads gut bacteria to flag Alzheimer's disease performed well when tested on the same population it was trained on, but its accuracy fell apart when applied to a different cohort, according to a reanalysis published October 22, 2026 in Frontiers in Aging Neuroscience. The study, by a single researcher, reprocessed five public 16S rRNA gene sequencing datasets, 509 fecal samples in total from China, Kazakhstan and South Korea, using one standardized pipeline so the cohorts could be compared on equal footing.

Within a single cohort, machine learning classifiers (logistic regression and LightGBM) separated Alzheimer's cases from cognitively normal controls with areas under the curve (AUC) ranging from 0.611 to 0.996, in line with the 0.70 to 0.90 range reported in several earlier single-cohort papers the author cites. But when a classifier trained on three cohorts was tested on a fourth it had never seen, AUC dropped to between 0.506 and 0.812, a mean fall of 0.13 to 0.17 points depending on the algorithm. A statistical test (PERMANOVA) found that cohort identity accounted for more of the overall bacterial variation than diagnosis did, by a factor of roughly twelve to one across the four labeled cohorts.

The author then applied two widely used batch-correction methods, ComBat-seq and MMUPHin, designed to strip out exactly this kind of site-to-site technical variation. Neither moved the needle: mean cross-cohort AUC changed by at most 0.007. The specific bacterial genera that each cohort's classifier relied on most also differed sharply from cohort to cohort, and for some genera the direction of the association reversed entirely, flagged as linked to disease in one population and to health in another.

The paper is a computational reanalysis of previously collected, publicly deposited data, not a new clinical study, and it was conducted by one author rather than a multi-site team. It does not claim gut bacteria play no role in Alzheimer's disease; it argues that current classifiers built on small, geographically narrow 16S datasets do not transfer to new populations, and that the batch-correction tools commonly used to paper over such differences did not work here. The author concludes that harmonized, multi-cohort, prospective validation is needed before any microbiome-based Alzheimer's test could be considered for clinical use.

The author recommends that future microbiome-biomarker studies be validated across independently collected cohorts from the outset, rather than reporting accuracy from a single population alone.