Machine-learning–based plasma metabolomic profiles for predicting long-term complications of cirrhosis

Background and Aims: The complications of liver cirrhosis occur after long asymptomatic stages of progressive fibrosis and are generally diagnosed late. We aimed to develop a plasma metabolomic–based score tool to predict these events. Approach and Results: We enrolled 64,005 UK biobank participants with metabolomic profiles. Participants were randomly divided into the training (n=43,734) and validation cohorts (n=20,271). Liver cirrhosis complications were defined as hospitalization for liver cirrhosis or presentation with HCC. An interpretable machine-learning framework was applied to learn the metabolomic states extracted from 168 circulating metabolites in the training cohort. An integrated nomogram was developed and compared to conventional and genetic risk scores. We created 3 groups: low-risk, middle-risk, and high-risk through selected cutoffs of the nomogram. The predictive performance was validated through the area under a time-dependent receiver operating characteristic curve (time-dependent AUC), calibration curves, and decision curve analysis. The metabolomic state model could accurately predict the 10-year risk of liver cirrhosis complications in the training cohort (time-dependent AUC: 0.84 95% CI: 0.82–0.86), and outperform the fibrosis-4 index (time-dependent AUC difference: 0.06 0.03–0.10) and polygenic risk score (0.25 0.21–0.29). The nomogram, integrating metabolomic state, aspartate aminotransferase, platelet count, waist/hip ratio, and smoking status showed a time-dependent AUC of 0.930 at 3 years, 0.889 at 5 years, and 0.861 at 10 years in the validation cohort, respectively. The HR in the high-risk group was 43.58 (95% CI: 27.08–70.12) compared with the low-risk group. Conclusions: We developed a metabolomic state–integrated nomogram, which enables risk stratification and personalized administration of liver-related events.