Needle-Free Bedside and Anthropometric Measures Show Weak Discrimination of an Adverse Blood-Biomarker Profile in Middle-Aged and Older Mexican Adults: An Internally Validated Machine-Learning Study

Background/Objectives: Needle-free measures such as grip strength, gait speed, and anthropometry are attractive as potential substitutes for blood-based biomarkers in low-resource settings, yet whether they can predict an adverse blood-biomarker profile remains unknown. This study aimed to determine whether machine-learning algorithms can predict such a profile from a defined set of needle-free bedside and anthropometric measures in Mexican adults. Methods: Using the Mexican Health and Aging Study (MHAS) 2012 biomarker subsample (n = 2000; mean age 62.4 years; 59.8% women), we defined a blood-based outcome: two or more of four abnormal markers (high CRP, low HDL, high cholesterol, vitamin D deficiency). Needle-free predictors were age, sex, grip strength, gait speed, blood pressure, heart rate, pulse pressure, BMI, waist, hip, waist-to-height ratio, waist-to-hip ratio, knee height, and walking-aid use. A circularity audit confirmed zero overlap. Sixteen algorithms were compared by fivefold cross-validation. Logistic regression and LightGBM were trained on 80% of the data and evaluated on a held-out 20% test set, with bootstrap CIs, calibration, SHAP, decision-curve analysis, permutation tests of the complete pipeline, and continuous regression. Sensitivity analyses varied the outcome definition, class balancing, survey weighting, and medication-related proxies. Results: Cross-validated ROC AUC ranged 0.481 to 0.524. LightGBM achieved AUC 0.539 (95% CI 0.477–0.596), and logistic regression 0.584 (0.528–0.638) in the held-out set, but the median held-out AUC of logistic regression across 200 random partitions was 0.518. Permutation tests of the complete cross-validated pipelines did not separate performance from chance (p ≥ 0.38). Continuous regression yielded negative R2 values for all markers (−0.14 to −0.08). No marker, sex stratum (men AUC 0.457; women 0.54), or sensitivity analysis reached useful discrimination (all AUC ≤ 0.58), and no model showed net benefit. SHAP importance was diffuse, with grip and gait ranked below all anthropometric and vital-sign features. Conclusions: In this sample, the evaluated bedside and anthropometric measures carried at most weak information about the predefined blood-biomarker profile, well below the level required for clinical use. They should not replace laboratory measurement of these biomarkers when biochemical assessment is clinically indicated.