Ablation-Based Machine Learning Framework for Body Mass Index Classification

Abstract

Body Mass Index (BMI) status classification can support population screening, but the relative predictive value of survey-based behavioral data and laboratory biomarkers remains unclear. This study develops an ablation-based health analytics framework to quantify the contribution of these feature domains and a derived lifestyle profile. Nineteen NHANES 2021–2023 component datasets were integrated, producing an analytical sample of 3,631 adults aged 18–80 years. The four-class target comprised Underweight, Normal, Overweight, and Obese categories; predictors included 28 behavioral and 11 biomarker variables. A two-stage framework applied K-Means lifestyle profiling to training data and Gradient Boosting classification across five controlled ablation configurations. Missing data processing, scaling, and Gaussian noise-augmented oversampling were fitted or applied only to the training set to minimize leakage. The primary two-stage full model achieved an AUC of 0.700 (95% CI: 0.643–0.751), accuracy of 0.556, macro F1 of 0.407, and Cohen’s kappa of 0.329. The biomarker-only configuration obtained the highest AUC (0.713), while the behavioral-only model retained moderate discrimination (AUC=0.635). The lifestyle cluster added only 0.002 AUC. Permutation importance identified diastolic blood pressure, uric acid, HDL cholesterol, alanine aminotransferase, and albumin as leading predictors, and a three-class formulation increased macro F1 to 0.545. The results indicate that biomarkers should be prioritized when laboratory resources are available, whereas behavioral variables remain useful for preliminary screening in lower-resource settings.