Metabolic syndrome and liver fibrosis in nondiabetic MASLD: A cross-sectional double machine learning study

Metabolic syndrome (MetS) is associated with hepatic fibrosis in metabolic dysfunction-associated steatotic liver disease (MASLD), but whether this association is causal in nondiabetic MASLD has not been assessed with semiparametric causal methods. We analyzed 1306 nondiabetic adults with MASLD from the National Health and Nutrition Examination Survey 2017 to 2020; 959 (73.4%) had prediabetes. The contrast was the supra-threshold MetS phenotype (≥3/5 harmonized components) versus the subthreshold phenotype (0-2/5), estimated as a covariate-standardized stratum contrast. Three co-primary outcomes were prespecified: fibrosis-4 index (FIB-4) ≥ 1.3, age-residualized FIB-4, and age-independent FibroScan ≥ 7.9 kPa. We applied double machine learning, augmented inverse probability weighting, causal forests, and 4 meta-learners, with 20 robustness analyses prespecified and 9 added at peer review. Estimates diverged with the age dependence of the outcome. For FIB-4 the adjusted risk difference was 0.153 (95% confidence interval [CI], 0.057-0.250; P = .002; E value = 3.75), or 0.095 (0.050-0.140) under a constant-effect specification; across 10 specifications estimates ranged from 0.078 to 0.163, attenuating to 0.091 (95% CI, -0.018 to 0.201; P = .10) after age residualization and to 0.001 (95% CI, -0.094 to 0.097; P = .98) for FibroScan, whose interval excludes the FIB-4 estimate at equal precision. Age stratification showed no effect below 40, with heterogeneity for FIB-4 (I2 = 72%) and age-residualized FIB-4 (I2 = 73%) but none for FibroScan (I2 = 0%), so the gradient is specific to FIB-4. Shapley decomposition ranked fasting glucose and central obesity highest, but every component's CI included zero, and no component pair showed additive-scale interaction after adjustment. Selection into the MASLD sample was exposure-dependent (40.7% vs 30.7%; P < .001); weighting gave 0.126 (0.003-0.250). The supra-threshold MetS phenotype was associated with a higher prevalence of FIB-4-defined fibrosis risk but not with age-independent elastography. The divergence tracks the outcome measure rather than sample composition: perturbations addressing what FIB-4 measures converge on 0.08 to 0.10, while those addressing sample composition leave it near 0.13 to 0.15. These findings generate hypotheses for confirmatory longitudinal testing rather than definitive causal evidence. They apply to the predominantly prediabetic MASLD population studied here and cannot be extrapolated to metabolically healthy or lean MASLD.