Early detection of low lean mass in adults using machine learning: a primary care-oriented approach.

Abstract

Low lean mass (LLM) is a key determinant of morbidity and mortality, particularly in older adults. Although dual-energy X-ray absorptiometry (DXA) is the gold standard for assessment, its high cost and limited availability hinder widespread screening. To develop and evaluate machine learning (ML) models for predicting LLM using easily obtainable demographic and clinical features, based on a cross-sectional analysis of NHANES 2011-2014, focusing on young and middle-aged adults (18-59 years). LLM was defined using DXA-derived appendicular skeletal muscle mass adjusted for BMI (ASM/BMI) based on the Foundation for the National Institutes of Health (FNIH) criteria. Six ML algorithms-Logistic Regression, Decision Tree, Random Forest, LightGBM, XGBoost, and Support Vector Machine-were trained using age, gender, height, weight, and handgrip strength. Among 6045 participants, 7% had LLM. The XGBoost model achieved an AUC of 0.94 (95% CI: 0.92-0.96), an F1-score of 0.54, sensitivity of 46%, and specificity of 98%. Although XGBoost showed the highest numerical F1-score, bootstrap comparisons indicated no statistically significant differences in F1-score across models. In contrast, XGBoost demonstrated significantly higher discrimination than the Decision Tree model according to DeLong's test. All algorithms showed high specificity (>98%) and accuracy (>94%), with markedly variable sensitivity (35%-48%). Machine learning models using simple and low-cost predictors can estimate the probability of LLM with acceptable discrimination. The proposed model may assist primary care clinicians in risk stratification once externally validated. Further work is needed to establish operational thresholds and evaluate real-world clinical utility.