Bayesian-optimized convex neural networks with embedded feature selection for high-dimensional anthropometric health prediction

Background and Objectives Anthropometric prediction and metabolic disorder screening require cost-effective solutions in resource-limited settings. We develop a convex-optimized neural network with embedded feature selection to establish quantitative relationships between body morphology and metabolic health. Methods Our framework combines Bayesian hyperparameter tuning with 1–2 regularized feature selection (λ1 = 0.020, λ2 = 0.002), processing 4,464 3D body scans, 1,000 Iraqi clinical records, and 768 Pima Indian diabetes cases. Spatial fully connected networks preserve anatomical relationships during feature selection. Results The model achieved 99.000% accuracy (95% CI [98.750–99.250]) for anthropometric gender classification, 94.700% accuracy (95% CI [93.200–96.200]) for Iraqi diabetes triage, and 90.040% accuracy (95% CI [86.800–91.280], Area Under the Receiver Operating Characteristic Curve (AUC) = 0.941) for Pima diabetes prediction, outperforming conventional methods (paired t = 27.340, p < 0.001). Key findings include identification of spine-to-scye length (weight = 0.990, Pearson r = 0.920) and Very Low-Density Lipoprotein (VLDL) (weight = 0.990, r = 0.982) as previously unreported predictors requiring further investigation, alongside traditional indicators. Conclusion The framework demonstrates anthropometry’s viability for security and diagnostic screening, with 57% parameter reduction vs . ensemble baselines while maintaining application-grade performance.