Development and validation of a nomogram-based risk prediction model for non-alcoholic fatty liver disease (NAFLD): a logistic regression analysis in a physical examination population.

Background: The global prevalence of non-alcoholic fatty liver disease (NAFLD) has reached an alarming 25%, based on recent population-based studies. NAFLD diagnosis primarily relies on imaging methods, which may be invasive, costly, or limited. Existing prediction models, such as the Fatty Liver Index (FLI) and NAFLD Liver Fat Score (NLFS), mainly incorporate metabolic parameters from Western populations and neglect non-metabolic factors, limiting their generalizability. To address these limitations, this study developed a screening model integrating both metabolic and non-metabolic indicators, tailored specifically to diverse populations undergoing routine physical examinations.

Methods: A retrospective cohort (N = 6,461, 2024) served to construct and validate the NAFLD prediction model, while an external cohort (N = 2,687, 2023) provided additional validation. NAFLD was confirmed by ultrasound. The primary cohort was randomly divided into training (70%) and internal validation (30%) sets. Initially, 20 candidate variables underwent univariate analysis (p < 0.1), and statistically significant variables (p < 0.05) were selected for the final multivariate logistic regression model. Model performance was assessed using area under the curve (AUC), calibration plots, and decision curve analysis (DCA). Robustness was verified through 10 × 10-fold cross-validation and external validation.

Results: The final predictive model incorporated 12 predictors (age, sex, body mass index (BMI), uric acid (UA), high-density lipoprotein cholesterol (HDL-C), red blood cell count (RBC), creatinine (Cr), free thyroxine (FT4), aspartate aminotransferase (AST), alanine aminotransferase (ALT), alpha-fetoprotein (AFP), triglyceride-glucose index (TyG)). The model demonstrated excellent discrimination, with a training AUC of 0.909 (95% CI: 0.900-0.919) and internal validation AUC of 0.905 (95% CI: 0.891-0.919). Optimal thresholds were 0.307 in training (sensitivity 78.8%, specificity 86.3%) and 0.205 in validation (sensitivity 88.3%, specificity 77.8%). Calibration was excellent (p = 0.989 training, p = 0.263 validation). DCA indicated substantial net benefits within thresholds from 1 to 98%. External validation confirmed strong model performance regarding discrimination, net benefit, and calibration.

Conclusion: This study identified crucial risk factors and constructed a robust NAFLD prediction model suitable for routine physical examination populations, offering clinicians a valuable tool for early disease detection and intervention.