Navigating machine learning and classical statistics in sarcopenia screening for community-dwelling older women

Machine learning (ML) is increasingly used for sarcopenia screening, but its added value over conventional models remains uncertain when datasets are moderately sized and predictors are few. We compared logistic regression (LR) with five ML algorithms in older women.

We analyzed 433 community-dwelling women aged ≥65 years. Sarcopenia was defined as handgrip strength <18 kg and bioelectrical impedance-derived appendicular skeletal muscle mass index <5.7 kg/m². LR, XGBoost, random forest, a CART decision tree, support vector machine (SVM), and multilayer perceptron (MLP) were evaluated using stratified five-fold cross-validation and pooled out-of-fold predictions. Analyses included a full nine-predictor set and a previously identified set comprising age, height, and calf circumference. A sensitivity analysis used nested cross-validation for hyperparameter selection.

Mean age was 76.2 ± 7.0 years, and 105 women (24.2%) had sarcopenia. In the primary nine-predictor analysis, LR and XGBoost achieved areas under the receiver operating characteristic curve (AUROCs) of 0.845 and 0.838, respectively, followed by random forest at 0.832. SVM, CART, and MLP showed lower discrimination. The three-predictor LR model achieved an AUROC of 0.852. With nested tuning, full-model AUROCs were numerically similar for LR (0.835), SVM (0.843), and MLP (0.840); corresponding three-predictor AUROCs were 0.865, 0.850, and 0.852.

LR remained competitive with the evaluated ML models, and the previously identified three-predictor model showed good discrimination. Relative performance depended on model configuration and training procedures. Independent external validation is required before clinical implementation.