Clinical and biochemical phenotypes of Lean MAFLD: machine learning-based biomarker identification and statistical...

Metabolic dysfunction-associated fatty liver disease (MAFLD) presents a significant public health issue in China. The "lean" subtype (L-MAFLD) accounts for a certain proportion of cases but is frequently underrecognized in primary care due to the absence of visible obesity. Its distinct clinical and biochemical phenotype remains unclear. This study aimed to identify routine biomarker features and statistical associations related to the metabolic variability of L-MAFLD within the general population.

This retrospective, cross-sectional study was conducted from January 2023 to January 2025 and included 12,553 adults who underwent health checkups at a tertiary hospital in Jiangsu, China. Participants were divided into groups based on body mass index (BMI) and ultrasonography: L-MAFLD (n = 330), overweight/obese MAFLD (n = 9,207), and non-MAFLD controls (n = 3,016). An integrated machine learning framework-including least absolute shrinkage and selection operator, extreme gradient boosting, and Boruta-was used to identify candidate physiological and biochemical variables. These biomarkers were then statistically evaluated using multivariable Firth-penalized logistic regression. Sensitivity analyses included age matching and additional adjustment for the available metabolic syndrome (MetS) component count. To assess non-linear associations and potential effect modification, restricted cubic splines (RCS) and stratified analyses were performed.

The machine learning pipeline initially identified 26 variables. After evaluating stability and redundancy, 19 variables remained in the final models. Higher levels of pulse rate, fasting blood glucose (FBG), alkaline phosphatase, and prealbumin were independently associated with higher odds of L-MAFLD. Higher levels of the alanine aminotransferase to aspartate aminotransferase (ALT/AST) ratio, serum creatinine (Scr), free triiodothyronine (FT3), diastolic blood pressure, and white blood cell count were associated with lower odds of L-MAFLD. Sensitivity analyses generally supported the main findings, although FT3 lost significance after age matching, and FBG was attenuated after adjustment for the MetS component count. RCS analysis showed an L-shaped nonlinear relationship between Scr and L-MAFLD, with a key inflection point at 60 μmol/L. A significant sex-specific interaction was found for the ALT/AST ratio (P = 0.012).

L-MAFLD shows a distinct clinical and biochemical phenotype, indicating metabolic imbalance despite normal BMI. These routine biomarkers may serve as potential epidemiological indicators of metabolic vulnerability among normal-weight individuals. The findings remain association-based. Future longitudinal multicenter studies with more complete metabolic profiling and external validation are needed.