Identifying factors associated with metabolic associated fatty liver disease based on multi-dimensional clinical data: an interpretable machine learning research

Metabolic-associated fatty liver disease (MAFLD) remains a major global public health concern, yet comprehensive identification tools integrating metabolic biomarkers with dietary and lifestyle factors remain limited. This study included 11,934 participants (22.07% with MAFLD). After multicollinearity diagnosis using VIF, we applied LASSO regression combined with stepwise logistic regression (LR) to screen key variables. Based on the identified variables, eight machine learning models were developed and compared. LR demonstrated optimal performance in the held‑out internal test set, achieving an AUC of 0.904 (95% CI: 0.894–0.915) with well‑balanced sensitivity (0.891) and specificity (0.781), and showed preferred generalizability with the smallest training‑test performance gap. Sensitivity analyses were performed to evaluate the variable selection strategy and address concerns about the homology between FLI/HSI and MAFLD. Four LR models were compared: a full‑variable model (excluding FLI and HSI), the screened‑variable model, a FLI + HSI‑only model, and a metabolic‑excluded model. The screened‑variable model achieved comparable performance to the full‑variable model while using fewer variables, and substantially outperformed the FLI + HSI‑only model in detecting MAFLD cases. The metabolic‑excluded model yielded a substantially lower AUC (0.792), confirming that metabolic variables are key drivers of performance. SHAP analysis identified BMI, TyG index, LAP, and HDL (negative contribution) as the most influential predictors. This reduced‑variable, interpretable LR model may support MAFLD risk stratification and individualized prevention, though external validation is warranted before clinical implementation.