Enhancing the accuracy of bioimpedance-derived appendicular skeletal muscle mass in aged adults through machine learning.

PURPOSE: Appendicular Skeletal Muscle Mass (ASMM) estimation via Bioelectrical Impedance Analysis (BIA) is a high-quality and bedside-accessible method. However, its accuracy is limited in comorbid populations, and determining the degree of error at the bedside remains a challenge. This study evaluates the application of Machine Learning (ML) algorithms as a decision-support layer to identify patients for whom the BIA-derived ASMM value is accurate. METHODS: This cross-sectional study included 701 participants aged ≥65 years (550 healthy subjects, 151 outpatients). ASMM measured by Dual-Energy X-ray Absorptiometry served as the reference. An absolute error ≤1.14 kg (the original equation's standard error) defined accurate estimation. Five algorithms-Extreme Gradient Boosting, Support Vector Machine (SVM), Random Forest, Logistic regression, and Neural Networks-were trained using three hierarchical sets of predictors: (1) anthropometric and bioimpedance variables, (2) model 1 plus handgrip strength, and (3) model 2 plus anthropometric circumferences (arm, waist, and calf) and knee height. RESULTS: Estimation error was minimal in healthy subjects but markedly higher in outpatients (median absolute difference 0.81 vs. 2.37 kg, p<0.001). Accurate estimates dropped from 62.4% in healthy individuals to 25.2% in outpatients. Discriminative performance improved progressively with each predictor set. In Set 3, SVM achieved the highest cross-validation Area Under the Curve (0.813) and a test AUC of 0.867, with an accuracy of 0.70. CONCLUSIONS: Integrating ML into BIA-based muscle assessment enables clinicians to quantify the reliability of individual ASMM estimates, even in patients with comorbidities. This approach provides a standardized framework for accepting or rejecting bedside estimations, enhancing clinical decision-making.