Association of residual cholesterol-inflammation index with MAFLD and related mortality risk: a population-based study integrating mediation and machine learning analyses

Background The residual cholesterol-inflammation index (RCII), a composite indicator integrating lipid metabolism and systemic inflammation, may serve as a novel predictor for metabolic dysfunction-associated fatty liver disease (MAFLD) and its related adverse outcomes. This study aimed to investigate the association between RCII and the risks of MAFLD and related mortality, assess its predictive value in clinical settings, and explore the mediating role of fasting plasma glucose (FPG) in these relationships. Methods A total of 13,254 participants from the NHANES 1999–2010 cycles were included. RC, CRP, and RCII were evaluated as exposures, with their distributions compared between MAFLD and non-MAFLD populations. Multivariable logistic and Cox regression models were used to assess the associations of RCII with MAFLD prevalence and three types of mortality (all-cause, cardiovascular, and premature). Nonlinear relationships were examined using restricted cubic splines (RCS). Mediation analysis was conducted to quantify the contribution of FPG to RCII-related risks, complemented by Mendelian randomization to infer causal effects of TC, HDL-C, LDL-C, and CRP on MAFLD. Multiple machine learning models were constructed to evaluate the predictive utility of RCII, with SHapley Additive exPlanations (SHAP) used for model interpretation. Results Compared to non-MAFLD individuals, participants with MAFLD exhibited pronounced metabolic dysregulation and inflammation, with significantly elevated RCII levels. RCII showed the strongest predictive power for MAFLD (Q4 vs Q1: OR = 17.79, P < 0.001). Higher RCII levels were independently associated with increased risks of MAFLD-related all-cause, cardiovascular, and premature death in both Kaplan–Meier and Cox models, with a clear dose-response pattern. These associations remained consistent across subgroups, with evidence of interaction effects. Mediation analysis revealed that FPG partially mediated the relationship between RCII and adverse outcomes, accounting for 2.02%–8.06% of the total effect. Among all models, the random forest algorithm achieved the highest predictive performance (accuracy = 89.70%, AUC = 0.960), with SHAP analysis confirming RCII as a top-ranking feature. Conclusions: RCII is independently and positively associated with both MAFLD risk and related mortality outcomes, demonstrating robust predictive capability. Its effects may be partially mediated by FPG. These findings underscore the potential of RCII as a clinically valuable biomarker for early identification and stratified management of individuals with high metabolic-inflammatory burdens.