Survival machine learning models for predicting all-cause and case-specific mortality risk in metabolic dysfunction-associated fatty liver disease patients

Emerging evidence links metabolic dysfunction-associated fatty liver disease (MAFLD) with increased all-cause and circulatory system disease (CSD) mortality in adults, yet survival machine learning studies are limited. This study analyzed 4415 NHANES participants with MAFLD to predict mortality using five survival models, and further, the optimal models were selected to identify the most significant predictors of mortality. Machine learning models proved highly effective in prediction. The Gradient Boosted Survival (GBS) model performed best for all-cause mortality, while Extra Survival Trees (EST) excelled for CSD mortality. The Shapley Additive Explanations (SHAP) analyses revealed that the five clinical factors most strongly associated with all-cause mortality were age, gender, platelet count, high-density lipoprotein cholesterol, and smoking status. For CSD mortality, the key factors associated with increased risk were age, blood urea nitrogen, systolic blood pressure, history of heart attack, and gender. Subgroup analyses confirmed GBS and Cox proportional hazard (CoxPH) were optimal for middle-aged and older all-cause mortality, whereas Elastic Net-regularized Cox proportional hazard (CoxNet) was best for older CSD mortality. The findings demonstrate that survival machine learning models effectively predict mortality risk in MAFLD patients. Integrating these models with permutation importance and SHAP provides transparent insights into individual risk profiles, enabling clinicians to clearly interpret how key variables contribute to predictions and improve risk stratification.