Sex-specific predictors of lower-limb strength: an interpretable machine learning analysis of anthropometric and body...

Muscular strength is a key indicator of physical function. However, its direct laboratory assessment often requires specialized, high-cost equipment and can impose significant physical fatigue or safety risks for certain populations, limiting its routine implementation in large-scale screenings. Using anthropometric and body-composition data from 200 healthy Korean adults aged 19-68 years, this study initially evaluated predictive models for leg extensor strength (LES), leg flexor strength (LFS), and handgrip strength (HGS). The interpretable machine-learning framework was subsequently focused on lower-limb strength outcomes showing stronger predictive performance to characterize the body-composition predictors of muscular strength. Correlation and linear regression analyses were combined with seven regression models, and feature contributions were quantified using SHAP values aggregated across all models; a linear-family ensemble (four homogeneous linear and regularized models) served as the primary analysis and an all-model ensemble as a sensitivity analysis. LES and LFS models consistently outperformed HGS models, and linear-based models showed the most stable performance, indicating that the strength-body-composition relationships were predominantly linear. SHAP analysis identified basal metabolic rate, age, and visceral fat area as the dominant predictors of both LES and LFS, jointly accounting for more than 70% of the total feature importance. The predictive profile were markedly sex-specific: in males, strength was most strongly predicted by skeletal muscle mass and body cell mass, whereas in females, age and bone mineral content were more informative for lower-limb strength. Notably, age contributed more strongly to extensor strength in males but to flexor strength in females, and visceral fat contributed approximately four-fold more to female than to male extensor strength. These findings demonstrate that practical, clinically relevant, and sex-specific prediction of lower-limb strength is achievable from simple body-composition data.