Prognostic analysis of sepsis-related liver injury and development of a prediction model based on machine learning method

Objective To analyze the prognosis of patients with sepsis-related liver injury (SRLI) and establish a prediction model for the occurrence of SRLI after ICU admission in sepsis patients using eight machine learning methods.

Methods Patients who met the sepsis diagnostic criteria and had no underlying liver or biliary diseases were included from the MIMIC-IV database, and were classified into SRLI and non-SRLI groups based on liver enzymes ≥5 times the upper limit of normal (ULN) or bilirubin ≥2.0mg/dL. Chi-square test, multivariate Logistic regression analysis, and propensity score matching were used to analyze the mortality risk between the two groups. Eight machine learning algorithmsLogistic regression, classification and regression tree (CART), random forest (RF), support vector machine (SVM), K-nearest neighbors (K-NN), naive Bayes method (NBM), extreme gradient boosting (XGBoost), and gradient boosting decision tree (GBDT)were employed to construct and validate the SRLI prediction model.

Results The chi-square test (P < 0.001), multivariate Logistic regression analysis (P < 0.05), and log-rank test after propensity score matching (P < 0.05) all indicated that SRLI increased the mortality risk of patients. Among the SRLI prediction models, RF algorithm had the highest area under the curve (AUC), with its value of 0.866, followed by GBDT (AUC=0.862), Logistic regression (AUC=0.859), SVM (AUC=0.837), NBM (AUC=0.830), CART (AUC=0.771), XGBoost (AUC=0.764), and K-NN (AUC=0.722).

Conclusion SRLI increases the mortality risk of patients. The prediction model constructed using the RF algorithm has high diagnostic value.