This resarch aimed to investigate the relationship of multiple adipose-muscle indices with secondary osteoporosis (OP) and sarcopenia in rheumatoid arthritis (RA) patients.

14 Machine learning (ML) algorithms were used to select characteristics  from 33 candidate variables and incorporated to construct clinical predictive models for OP, sarcopenia and osteo-sarcopenia (OS). Restricted cubic spline (RCS) explored dose-effect relationships between body composition and musculoskeletal loss. 

Six key features were identified, with the fat-to-muscle ratio (FMR) and skeletal muscle mass to visceral fat area ratio (SVR) being the most prominent. FMR was a significant risk factor for OP (OR=3.31, 95%CI:1.33-8.27), whereas SVR was a protective factor against sarcopenia (OR=0.23, 95%CI:0.04-0.76). The Light Gradient Boosting Machine (LightGBM) model demonstrated superior performance for predicting OS, achieving an AUC of 0.988 (95%CI:0.958-1.000), significantly outperforming traditional logistic regression (AUC=0.888). RCS analysis revealed a nonlinear, threshold-dependent relationship, for instance, the predictive power for OP increased markedly when SVR exceeded 0.449. MR analysis indicated that body fat mass fully mediated the causal pathway from RA to OP/sarcopenia (RA-Fat:OR=9.53,P=0.009; Fat-OP:OR=1.00,P<0.001; Fat-Muscle:OR=1.10,P<0.001). Mediation analysis demonstrates that albumin negatively mediates the protective effect of SVR on RA-OS (proportion = -17.65%). Weighted quantile sum(WQS) analysis validated skeletal Muscle Mass Index (SMI) (38.3%) and SVR (14.2%) are core inhibitory factors of RA-OS.

In summary, FMR and SVR are important risk/protective factors for RA patients with concurrent musculoskeletal metabolic diseases, respectively, and the construction of a comprehensive prediction model is more superior than a single predictor, and the machine learning algorithm has a broader clinical application prospect than the traditional regression model. Meanwhile, clinicians should pay attention to the dose-response relationship between the factors and the outcomes and the causality covered to find the optimal range of application of the model, so that the accuracy and specificity of clinical diagnosis can be improved and the diagnosis and treatment can be more convenient for more patients.