UC San Diego

Non-alcoholic fatty liver disease (NAFLD) has emerged as a health crisis not only in the US but throughout the world. Current methods of identifying indicators of liver disease progression, such as advanced fibrosis (AF), are accurate but costly, slow, and impossible to perform on population-wide studies. To fill in this diagnostic gap, we look to build machine learning (ML) models using serum and fecal metabolomics data from NAFLD patients that can accurately classify AF. These patients encompass the entire spectrum of NAFLD which allows us to dig into the diversity and composition of their metabolomes to understand the changing metabolites between different liver disease states. Using these metabolites, we built baseline ML models and improved performance using a multitude of feature selection methods. These feature selection methods allowed us to develop a fecal metabolome ML model that could classify AF with an area under the receiver operating characteristic (AUROC) of 0.82. To further develop our findings, we combined our strongest fecal metabolome results with fecal microbial features from a previous study to build a multi-omic machine learning model. Building this multi-omic ML model expanded on our results and led us to a stronger performing model that could accurately classify AF patients with an AUROC of 0.86.