Stool metagenomics added real discrimination to predicting admission in cirrhosis

Design

multinational prospective cohort with stool metagenomics and machine learning prediction modelling

Population

679 outpatients with cirrhosis across seven countries, 90-day follow-up

Primary outcome

non-elective hospitalisation within 90 days

Effect

25% admitted; combined clinical-microbiome model AUC 0.84 vs 0.79 clinical alone and 0.74 microbiome alone

Six hundred and seventy-nine outpatients with cirrhosis across seven countries provided stool for metagenomic profiling and were followed for 90 days. A quarter were admitted non-electively in that window.

Patients who were admitted had lower microbial diversity, and that held across countries despite the underlying compositions differing markedly between them. After adjustment for country, disease severity, aetiology and treatment, 19 species remained independently associated with admission - Enterococcus faecium and Veillonella rogosae enriched, multiple commensals depleted. Functional profiling showed coordinated shifts in carbohydrate degradation, glycan biosynthesis and lipid and nucleotide salvage pathways, alongside antimicrobial resistance genes. A model combining clinical and microbiome features reached an area under the curve of 0.84, against 0.79 for clinical features alone and 0.74 for microbiome alone.

The cross-country consistency of the diversity signal, despite composition varying, is the finding with legs - it suggests the relevant thing is loss of function rather than any particular organism, which is more likely to transfer to an Indian cohort than a species list would. Stool metagenomics is not available as a clinical test, and an increment from 0.79 to 0.84 would need to be substantial and cheap to change a decision. What is available today is the inference: enrichment of Enterococcus and loss of short-chain fatty acid producers is what antibiotic exposure does, and antibiotic stewardship in outpatient cirrhosis is actionable now.

• Loss of microbial diversity preceded admission consistently across seven countries.
• Enterococcus faecium enrichment and commensal depletion is the pattern antibiotic exposure produces.
• Review antibiotic courses in outpatients with cirrhosis, including those given elsewhere.
• Stool metagenomics is not a clinical test; the added discrimination was 0.05 on the curve.
• Clinical variables alone already reached an area under the curve of 0.79.

Why it matters

It identifies a modifiable exposure, not just a predictive marker, because the signature that preceded admission is the one antibiotics create.

Don't overread it

Associations from a model built and tested in the same cohort - it does not show that altering the microbiome prevents admission.

The statistics, in plain English

An increase from 0.79 to 0.84 in area under the curve is a genuine improvement in ranking, but small in the context of a test that does not exist clinically and would be expensive if it did. Note also that microbiome features were selected from a very large candidate set in 679 patients, which inflates apparent performance unless validated externally - and this model was not tested in an independent cohort.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.