Fig. 1

Applying the 3 machine-learning analyses to identify the predictive metabolic indicators associated with Hepatitis B-related acute-on-chronic liver failure (ACLF) prognoses. (A) Least absolute shrinkage and selection operator (LASSO) regression with 13-fold cross-validation, corresponding to the 13 statistically significant ACLF patient indicators found from univariate logistic regression analysis, was used to reduce the dimension of the grouping characteristics. (B) 6 indicators corresponded to the minimum error: age, glycosylated serum protein (GSP), high-density lipoprotein cholesterol (HDL-c), Na⁺, prothrombin time (PT), and international normalized ratio (INR). (C) 13-fold cross-validation error plot from random forest (RF) analysis. (D) 9 indicators were identified from RF, based on a cut-off threshold of 20: prothrombin activity (PTA), PT, INR, Na⁺, Model for End-Stage Liver Disease (MELD), HDL-c, age, white blood cell (WBC), and GSP. (E) Support vector machine (SVM) analysis identified 12 indicators: age, creatinine, GSP, HDL-c, Na⁺, PT, PTA, INR, WBC, MELD, infection, and artificial liver treatment. (F) Venn diagram showing the number of shared indicators between multivariate logistic regression, LASSO, RF, and SVM.