Interpretable bioimpedance-based modeling of body composition and metabolic indices related to physical function in maintenance hemodialysis patients

Objective In maintenance hemodialysis (MHD) patients, muscle wasting and altered body composition are major determinants of reduced physical function. This study aimed to identify the key body composition features associated with four physical-function-related indices—Basal Metabolic Rate (BMR), Fat-Free Mass (FFM), Skeletal Muscle Mass (SMM), and Percentage of Body Fat (PBF)—and to evaluate whether ensemble regression models can estimate these indices from routine bioelectrical impedance analysis (BIA) data. Methods Ensemble regression models were constructed using BIA-derived indicators to estimate BMR, FFM, SMM, and PBF. The SHAP algorithm was employed to assess feature importance. Individual regression models (LR, DT, SVR, GBR, Adaboost, and KNN) were compared with ensemble models (RF, ET, and LGBM) using R 2 , MAE, MSE, and RMSE. Results SHAP analysis identified Total Body Water (TBW) and height as the most significant predictors across all four indices, followed by Intracellular Water (ICW), Extracellular Water (ECW), minerals, and body weight. Ensemble models consistently outperformed individual models. LGBM achieved the highest performance (e.g., R 2 = 0.99, RMSE = 0.11 for BMR), markedly surpassing individual models such as DT ( R 2 = 0.89, RMSE = 0.39). Similar trends were observed for SMM, FFM, and PBF. Conclusion Ensemble learning models, particularly LGBM, demonstrated high apparent accuracy in estimating the four indices. SHAP analysis consistently identified TBW and height as dominant contributors. We caution that this predictive strength partly reflects expected physiological relationships between body water, body size, and these indices, rather than novel clinical insights. The practical value lies in providing an interpretable body composition profile that, if externally validated, may help identify MHD patients at risk of muscle wasting and poor functional status. Because this study used only internal validation and had no independent external validation cohort, clinical application requires validation against directly measured functional outcomes in larger, multicenter, longitudinal cohorts, with body composition standardized to a fixed point in the dialysis cycle.