Machine learning model for ACLF risk in decompensated cirrhosis

To develop a machine learning-based predictive model for acute-on-chronic liver failure (ACLF) risk in patients with decompensated cirrhosis using baseline clinical indicators at hospital admission.

This retrospective cohort included 360 decompensated cirrhosis patients with intensive monitoring needs or infection/circulatory risk factors, randomly divided into training and validation sets at a 7:3 ratio. Three models-random forest, support vector machine, and logistic regression-were constructed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) were employed for model interpretability.

In the training set, 52 patients (20.6%) experienced ACLF. Multivariate analysis identified total bilirubin (TBil) (OR = 1.051, 95% CI: 1.030-1.073, P < 0.001), arterial lactate (OR = 1.341, 95% CI: 1.058-1.700, P = 0.015), and white blood cell count (WBC) (OR = 1.111, 95% CI: 1.007-1.227, P = 0.036) as independent risk factors, whereas prothrombin time activity (PTA) (OR = 0.961, 95% CI: 0.941-0.982, P < 0.001) was protective. The random forest model showed comparable performance to logistic regression with training/validation AUCs of 0.833 (95% CI: 0.757-0.910) and 0.700 (95% CI: 0.555-0.845). All models showed acceptable discrimination and net clinical benefit on DCA. Calibration was good in the training set but moderate in the validation set (calibration slopes 0.580-0.620). SHAP importance ranking placed PTA, TBil, WBC, and lactate as the top predictors.

The random forest model showed acceptable training performance, but the wide validation AUC confidence interval suggests limited stability, likely due to small sample size and few events. External validation is needed before clinical use. Nonetheless, the model shows potential for early warning and risk stratification in decompensated cirrhosis.