“From this metagenomic information, the researchers developed a 31-marker microbial panel. Using a machine learning model, they tested the panel. They found that it accurately predicted ASD diagnosis across different ages, sexes, populations, and geographical locations, much better than using a single species of microorganism, such as bacteria.”
From the abstract: “Machine learning using single-kingdom panels showed area under the curve (AUC) of 0.68 to 0.87 in differentiating children with ASD from those that are neurotypical. A panel of 31 multikingdom and functional markers showed a superior diagnostic accuracy with an AUC of 0.91, with comparable performance for males and females. Accuracy of the model was predominantly driven by the biosynthesis pathways of ubiquinol-7 or thiamine diphosphate, which were less abundant in children with ASD. Collectively, our findings highlight the potential application of multikingdom and functional gut microbiota markers as non-invasive diagnostic tools in ASD.”
They don’t say in the abstract what proportion of their sample had ASD. Without this information, the ROC AUC is meaningless. You will get very high AUC on an unbalanced sample even with a trivial predictor like “always say no.”
From the abstract: “Machine learning using single-kingdom panels showed area under the curve (AUC) of 0.68 to 0.87 in differentiating children with ASD from those that are neurotypical. A panel of 31 multikingdom and functional markers showed a superior diagnostic accuracy with an AUC of 0.91, with comparable performance for males and females. Accuracy of the model was predominantly driven by the biosynthesis pathways of ubiquinol-7 or thiamine diphosphate, which were less abundant in children with ASD. Collectively, our findings highlight the potential application of multikingdom and functional gut microbiota markers as non-invasive diagnostic tools in ASD.”
https://www.nature.com/articles/s41564-024-01739-1