Duplicate-Aware Internal Validation of Machine-Learning Models for Classifying a Tanita BIA-Derived High-Adiposity...

Background/Objectives: Anthropometric machine-learning models may approximate body-composition classifications, but performance can be inflated by inconsistent preprocessing, non-independent validation records, and incomplete calibration reporting. This study evaluated a sex-specific bioelectrical impedance analysis (BIA)-defined high-adiposity phenotype using simple anthropometric variables in adults with and without hypertension. Methods: Of 583 prespecified records, 11 with invalid placeholder-coded values in required anthropometric or body-composition fields were excluded, leaving 572 participants (385 normotensive and 187 hypertensive). A high-adiposity phenotype was defined as BIA-derived body fat ≥ 25% in males or ≥35% in females. Predictors were sex, height, body weight, waist circumference, and hypertension status; body mass index and body-fat percentage were excluded. Eight algorithms were assessed using 10 repetitions of stratified five-fold group cross-validation, with identical predictor profiles kept within the same fold. Continuous predictors were standardized within training folds. Performance was estimated from averaged out-of-fold probabilities with 2000 stratified bootstrap confidence intervals, calibration measures, and SHAP analysis. Results: A high-adiposity phenotype was present in 441 participants (77.1%). Random forest achieved the highest discrimination (ROC AUC = 0.959, PR AUC = 0.980). Gradient boosting provided the strongest threshold-dependent performance (accuracy = 0.937, balanced accuracy = 0.895, sensitivity = 0.973, specificity = 0.817, precision = 0.947, F1 score = 0.960, MCC = 0.817) and the lowest Brier score (0.057), but its calibration slope was 0.462, indicating overconfident probabilities. SHAP analysis identified waist circumference as the largest attribution within the fitted gradient-boosting model; this result must be interpreted jointly with the ablation analysis because sex, height, and body weight are inputs to the proprietary Tanita equation. Conclusions: Simple anthropometric variables classified the prespecified Tanita BIA-derived high-adiposity threshold with strong internal performance after duplicate-aware validation. Sex, height, and body weight alone achieved a ROC AUC of 0.920; adding waist circumference produced a small and uncertain increase in discrimination (ΔROC AUC = 0.0054, 95% CI -0.0060 to 0.0153), whereas hypertension status added a negligible value. The findings represent internal validation of a device-defined outcome, not prediction of an independent biological reference, and require external validation against criterion body-composition methods before clinical application.