Answers are grounded in this paper and cite their sources.
Sarcopenia, particularly low skeletal muscle index (SMI), is an increasing concern in ageing populations and is associated with frailty, falls, hospitalization, and premature mortality. Early identification of individuals at risk requires consideration of a range of demographic, lifestyle, anthropometric, and metabolic factors. Because these factors may interact in complex ways, conventional regression methods may not fully capture their combined effects. Interpretable machine-learning methods provide a flexible alternative that can account for such complexity while maintaining a level of transparency that is important for clinical and epidemiological applications. We analysed baseline data from 660 adults aged 60 years or older who participated in the Fasa Adults Cohort Study. Low SMI was defined according to the Asian Working Group for Sarcopenia criteria and assessed using bioelectrical impedance analysis. We systematically evaluated 154 combinations of 11 classifiers and 14 oversampling methods using repeated stratified cross-validation (5-fold, 10 repeats). All preprocessing steps were performed within the training folds to minimize the risk of data leakage. Models were ranked according to PR-AUC, which is particularly informative for imbalanced outcomes. Analyses were conducted in the overall cohort and separately among men and women. Model interpretation was based on Random Forest importance, SHAP values, and depth-limited decision trees. Low SMI was present in 19.84% of participants (131 of 660). Among the 154 classifier–oversampling combinations evaluated, logistic regression combined with appropriate oversampling methods achieved the highest PR-AUC across all three analytic groups, with values of 0.836 in the overall cohort, 0.850 in men, and 0.861 in women. The corresponding AUC values were above 0.95 in all three groups. Feature-importance analysis based on Random Forest and SHAP results identified body mass index as the most influential predictor, followed by hip circumference and waist circumference, with total body water ranking fourth in the overall cohort. This pattern was broadly similar in men and women. Model performance was somewhat more variable in the sex-specific analyses, which may be related to the smaller number of outcome events in these subgroups. Using routinely available clinical and anthropometric measurements, interpretable machine-learning models demonstrated good internal discrimination for low SMI in this cohort of older Iranian adults. The prominence of body-composition measures, many of which are already routinely collected, highlights their potential practical relevance. Differences in predictor importance between men and women may provide useful hypotheses for future research but should be interpreted cautiously. External validation in independent populations is needed before these models can be considered for screening or clinical applications.
No takes yet. Share an insight, caveat, or question.
Honarvar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: