GUT MICROBIOME PROFILE THROUGH MACHINE LEARNING ALGORITHMS CLARIFIES HEPATIC ENCEPHALOPATHY DIAGNOSES IN A MULTI-CENTER CIRRHOSIS COHORT

While it is assumed that cognitive complaints in patients with cirrhosis are from hepatic encephalopathy (HE), almost 50% of cognitive impairment is from another cause. Incorrect HE diagnosis worsens QOL and ignores the underlying cause. Diagnosing HE-specific cognitive impairment is challenging in clinic & new approaches are needed. Fortunately, microbiota changes in minimal HE (MHE) are different from those in PTSD & dementia, but whether the microbiota profile is useful for MHE diagnosis in practice remains unclear. Aim: Define and prospectively validate the utility of machine learning (ML) algorithms to predict MHE-related cognitive impairment using gut microbial profiles in a multi-center study of cirrhosis.
Methods: Training cohort: Outpatients with cirrhosis tested for MHE using a cirrhosis-specific test (Psychometric hepatic encephalopathy score, PHES) using US-norms underwent stool collection and metagenomics. ML algorithms to ensure >80% NPV for MHE was used to classify subjects into MHE/not (Fig A/B).Validation cohort: After informed consent, stool collection kits were mailed to patients being treated for HE at 3 centers. Samples were compared to the training cohort to diagnose MHE/not by gut microbial testing. Reasons for HE Rx initiation were re-evaluated by investigators blinded to the gut microbial test results. Gut microbial MHE diagnosis vs clinical diagnoses were analyzed.
Results: Training cohort: 340 cirrhosis pts (152 had MHE on PHES, age 56 years) were included(Fig A/B). Neural networks using patterns of 20 species had the greatest NPV for MHE and was used to define MHE on gut microbiota.
Validation cohort: 28 men with cirrhosis and a clinical diagnosis of HE from 3 centers (Richmond, West Haven & Dallas VAMCs) underwent prospective stool metagenomics & comparison (Fig C). Probability ≤0.50 on ML was considered no-MHE & >0.51 was considered MHE. Of the 9 pts with no-MHE on gut microbial ML algorithm, 8 had unclear reasons (4 sleep issues, 2 stopped HE Rx >1 yr ago) & 2 no reason for their HE diagnosis in the past while 1 had true HE. Sites were encouraged to re-evaluate the HE diagnosis in these 8 patients (Fig E). In the 19 MHE patients on gut microbial ML, 17 had true prior HE (hospitalization, and >grade 2) while 2 had unclear reasons. This was associated with excellent accuracy, specificity, and sensitivity (Fig D/ Table 1).
Conclusions: Using a training set of 340 patients with cirrhosis, ML gut microbial neural networks discriminated between those with and without HE. The ML gut microbial algorithm was able to determine HE vs no-HE with >89% accuracy in a multi-center cohort of patients with cirrhosis. This led to re-evaluation of the HE diagnosis in 32% of patients with previously diagnosed HE, which was corroborated with interview and chart review.