Body Mass Index underestimates excess adiposity: diagnostic discrepancy with bioelectrical impedance analysis and misclassification of nutritional status.

Background: Body mass index (BMI) remains the most widely used tool for obesity screening in clinical and public health settings due to its simplicity and accessibility. However, as BMI does not directly quantify adiposity, it may fail to detect excess body fat and obscure clinically relevant metabolic phenotypes. This limitation has not been adequately characterized in large Latin American cohorts. We quantified diagnostic discordance between BMI- and bioelectrical impedance analysis (BIA)-based obesity classification in Chilean and Mexican adults using age- and sex-specific body fat percentage thresholds. Methods: In this cross-sectional study, we pooled data from three adult body-composition databases, including 12,604 participants from Chile and Mexico. Adiposity-defined obesity was classified using age- and sex-specific body fat percentage cut-offs proposed by Gallagher et al., whereas BMI-defined obesity followed World Health Organization criteria (BMI ≥30 kg/m²). Participants were allocated to four phenotypes: concordant non-obese, normal-weight obesity, high-BMI/normal-adiposity, and concordant obese. Agreement between BMI and BIA classifications was evaluated using Cohen's kappa, sensitivity, specificity, and Bland-Altman analysis. Ten supervised machine learning algorithms were trained and internally validated by 10-fold cross-validation to predict BIA-defined obesity using anthropometric variables. Results: Agreement between BMI and BIA was moderate (κ = 0.443). BMI showed high specificity (96.9%) but low sensitivity (46.3%), failing to identify 53.7% of BIA-defined obesity. Misclassification was more pronounced in women (κ = 0.386). Normal-weight obesity was identified in 16.2% of women and 4.6% of men. Among men with elevated BMI (≥ 25 kg/m²), 42.8% showed the high-BMI/normal-adiposity phenotype. Ordinary least squares models explained 68.6%-70.6% of body fat variance, with systematic error at extremes. Machine learning models showed high discrimination, with multilayer perceptron achieving AUC = 0.999. Conclusion: BMI underestimates excess adiposity in this Latin American adult sample and misclassifies relevant body composition phenotypes, particularly in women. Reliance on BMI alone may obscure a large proportion of individuals with excess adiposity, potentially limiting early detection and preventive strategies. Predictive models using simple anthropometric inputs may improve adiposity screening where direct body composition assessment is unavailable.