Machine Learning-Based Prediction of Functional Deficit from Anthropometric and Cardiometabolic Biomarkers in Middle-Aged and Older People

@article{urrutia2026machine,
  title = {{Machine Learning-Based Prediction of Functional Deficit from Anthropometric and Cardiometabolic Biomarkers in Middle-Aged and Older People}},
  author = {Jessica Mondaca Urrutia and Rodrigo Yáñez-Sepúlveda and Paz Pezoa-Fuentes and Yeny Concha-Cisternas and Eduardo Guzmán-Muñoz and Carlos Abraham Herrera-Amante and Lawrence W. Judge and Claudio Farías-Valenzuela and Pablo Valdés-Badilla and Bastian Parada-Flores and Jordan Hernandez-Martínez and Exal Garcia-Carrillo},
  journal = {Journal of Clinical Medicine},
  year = {2026},
  month = nov,
  doi = {10.3390/jcm15218618},
  url = {https://www.mdpi.com/2077-0383/15/21/8618},
  note = {CC BY 4.0},
}

TY  - JOUR
TI  - Machine Learning-Based Prediction of Functional Deficit from Anthropometric and Cardiometabolic Biomarkers in Middle-Aged and Older People
AU  - Jessica Mondaca Urrutia
AU  - Rodrigo Yáñez-Sepúlveda
AU  - Paz Pezoa-Fuentes
AU  - Yeny Concha-Cisternas
AU  - Eduardo Guzmán-Muñoz
AU  - Carlos Abraham Herrera-Amante
AU  - Lawrence W. Judge
AU  - Claudio Farías-Valenzuela
AU  - Pablo Valdés-Badilla
AU  - Bastian Parada-Flores
AU  - Jordan Hernandez-Martínez
AU  - Exal Garcia-Carrillo
PY  - 2026
DA  - 2026/11/05
JO  - Journal of Clinical Medicine
DO  - 10.3390/jcm15218618
UR  - https://www.mdpi.com/2077-0383/15/21/8618
AB  - Background: Slow gait, weak handgrip strength, and impaired balance are common functional deficits in aging populations that predict adverse health outcomes. Accurate, scalable prediction tools are needed for community-based screening in low-resource settings. Objectives: To benchmark nine machine learning (ML) classifiers for predicting the presence of ≥1 functional deficit using cardiometabolic and anthropometric predictors; to assess calibration, generalization, and clinical utility; to quantify the incremental value of these predictors over age and sex; and to characterize the stability of SHAP-based feature importance. Methods: We analyzed cross-sectional data from the Mexican Health and Aging Study 2012 (MHAS 2012) biomarker sub-study (N = 1961; mean age 61.9 ± 10.5 years; 59.8% women). The outcome was ≥1 of slow gait (<0.8 m/s), weak handgrip (EWGSOP2 sex-specific cut-points), or inability to sustain a 10-s one-leg stance (prevalence 78.7%). Twenty cardiometabolic and anthropometric predictors were used, excluding the functional measures used in outcome construction. Performance was estimated with nested, repeated stratified 10-fold cross-validation (5 × 10 outer folds; inner 5-fold hyperparameter tuning), with all preprocessing confined to training folds. Results: Logistic Regression (LR) achieved the highest discrimination (AUC-ROC 0.751, 95% CI 0.724–0.779), with good calibration (scaled Brier 0.125; calibration slope 0.96) and a small overfit gap (0.011); no classifier outperformed it (ET: AUC 0.745). An age–sex model alone reached AUC 0.734; all other predictors added ΔAUC = +0.018 (95% CI 0.004–0.031), almost entirely through anthropometry, whereas blood biomarkers and blood pressure added no measurable discrimination. SHAP identified age, sex, height, body weight, and BMI × Age as the most influential features; BMI ranked 12th of 20 (bootstrap 95% rank interval 8–20). Conclusions: Cardiometabolic and anthropometric predictors identify middle-aged and older Mexican people with ≥1 functional deficit with moderate accuracy (AUC ≈ 0.75), but most of the signal is carried by age, sex, and simple anthropometry. This reproducible, circularity- and leakage-free framework supports risk stratification rather than diagnosis, and requires external and temporal validation before clinical implementation.
SN  - 2077-0383
ER  -