Beyond Algorithmic Novelty: Deployment-Centred Diabetes Detection and Disease Progression Prevention Using Population-Representative Data

Diabetes prediction research faces methodological challenges. Many published models lack adequate validation and overlook gestational diabetes mellitus (GDM) history despite its 7.4-fold risk elevation for Type 2 diabetes. Here, cross-sectional detection is examined across three realistic contexts: mobile screening, GP surgeries, and specialist clinics using population representative NHANES data (2013-2016, n=8,259). Survey-weighted, cost-sensitive methods maintained representativeness, with neglect of design effects shown to almost halve statistical power. Results challenge expectations: traditional machine learning (ML) consistently outperformed deep learning (DL), with gradient boosting identifying female-specific risk patterns missed by linear models and neural networks. AUC values were reasonably high (83.2% mobile, 86.2% GP, 86.6% specialist), yet class imbalance exposed modest precision rates where mobile screening produced 64.8% false positives versus 59.1% in GP surgeries and 52.3% in specialist clinics. GDM features were not retained in optimal models, with metabolic measurements capturing risk information more efficiently. Advanced laboratory panels offered marginal gains over standard tests, suggesting their complexity may be hard to justify outside specialist contexts. The findings support a tiered deployment strategy: matching model and assessment depth to setting rather than chasing algorithmic novelty. Differential risk weighting between females and males suggests unintentional bias, evidence of gender disparities, alongside varying performance across ethnic groups, which warrants further investigation into stratified models.