Benchmarking Interpretable Machine Learning for Frailty Screening in Older Adults Using Routine Clinical Variables with and Without SARC-F: A Pilot Diagnostic Study

Background/Objectives: Frailty is common in Latin America and the Caribbean, yet validated, low-cost screening tools based only on routine clinical data remain limited. We developed benchmarked interpretable machine learning models for frailty screening in older adults and assessed the added value of SARC-F beyond basic hemodynamic and anthropometric measures. Methods: In this cross-sectional pilot diagnostic modeling study, 100 older adults (mean age 69.8 ± 6.8 years; 73% women) from a tertiary care center were classified with the FRAIL scale. The primary endpoint was frail versus non-frail. Three logistic regression models were evaluated using stratified 5-fold cross-validation: Model A (age, sex, BMI, nutritional status, systolic and diastolic blood pressure), Model B (mean arterial pressure and pulse pressure), and Model C (Model A plus SARC-F). Decision tree and random forest were secondary comparators. Results: SARC-F was the only variable significantly associated with frailty category (p < 0.001). Models A and B showed near-chance discrimination (both AUC = 0.546), whereas Model C achieved an AUC of 0.921 (95% CI: 0.864–0.966), sensitivity of 0.812, specificity of 0.846, and F1-score of 0.821. SARC-F alone yielded an AUC of 0.942. In Model C, SARC-F was the dominant predictor (standardized OR 12.52, 95% CI: 8.90–23.91). Conclusions: Routine hemodynamic and anthropometric variables alone were inadequate for frailty discrimination. SARC-F captured most of the discriminative signal, supporting its use for frailty screening in resource-limited geriatric settings. Although the logistic regression pipeline was transparent and computationally inexpensive, external prospective validation is required.