Development and internal validation of an interpretable model for MASLD identification using metabolic and body composition indicators in middle-aged and older adults

Metabolic dysfunction-associated steatotic liver disease (MASLD) is closely linked to metabolic dysfunction, insulin resistance, dyslipidemia, obesity, and altered body composition. In middle-aged and older adults, age-related changes in fat distribution and skeletal muscle status may further modify the metabolic phenotype of MASLD. This single-center retrospective cross-sectional study enrolled 1,267 subjects aged ≥ 45 years (488 MASLD, 779 non-MASLD) who underwent health examinations. Demographic, metabolic, biochemical, anthropometric, and body composition data were collected. Four incremental models were constructed, followed by variable selection using correlation, collinearity, Elastic Net regression, and Bootstrap stability screening. Eight algorithms were compared, with model performance evaluated by AUC, Brier score, calibration slope, decision curve analysis, and SHAP interpretation. Incremental modeling showed that adding biochemical and anthropometric parameters increased AUC from 0.648 to 0.839 and 0.873, respectively. Further adding BIA metrics slightly raised AUC to 0.877 ( p = 0.604), though sensitivity, specificity, and calibration modestly improved. Twelve variables identified through Elastic Net stability screening were used to construct multiple screening-oriented identification models, among which the logistic regression model exhibited the best overall performance(validation AUC = 0.869, sensitivity = 0.836, specificity = 0.755), SHAP analysis revealed that TyG-BMI, ALT, AST, dASMI, ALB, LDL-C, UA, history of metabolic diseases, dPhaseAngle, HDL-C, Age, and WHtR were the primary contributing variables. Sensitivity analyses indicated that the non-diagnostic overlapping variable model still retained substantial diagnostic capacity, suggesting that the final model is not entirely dependent on variables associated with MASLD diagnostic criteria.In an additional criteria-based analysis, adding eight non-overlapping predictors increased the AUC from 0.749 to 0.809 (ΔAUC = 0.060, p = 0.0003), indicating incremental discrimination beyond the metabolic criteria, although potential incorporation bias could not be eliminated. Conventional biochemical, metabolic, and anthropometric indicators provided most of the discriminative information for current MASLD identification. Adding BIA-derived body composition variables did not materially improve model discrimination beyond conventional variables. Because several predictors overlapped with or approximated components of the MASLD definition, part of the full model’s discrimination was definition-dependent. The final logistic regression model should be regarded as an internally validated auxiliary classification model, and external validation is required before clinical implementation. Not applicable.