Abstract / Summary
Abstract Background Conventional prediabetes screening depends on glycaemic biomarkers such as fasting plasma glucose and HbA1c, but these tests recognize overlapping but non-identical high-risk groups, no single biochemical definition has achieved international consensus, and large-scale biochemical screening carries substantial testing burden. Machine-learning risk models using multidomain baseline clinical information could triage individuals for confirmatory glycaemic testing if they are constructed to avoid target leakage, calibrated for probability-based decisions, evaluated at operationally relevant thresholds, and audited for subgroup fairness. Methods Using UK Biobank baseline data, we analysed 118,660 participants without prevalent diabetes. Prediabetes was defined according to the ADA criteria as HbA1c 5.7–6.4%. HbA1c, glucose, and glycaemic-derived features were excluded from an 82-feature glycaemic-blind predictor set to prevent target leakage. Six candidate classifiers were benchmarked; LightGBM was selected as the primary model. Bayesian hyperparameter optimisation used the 2006–2008 training cohort with a stratified 2009 validation subset; the 2010 cohort (n = 19,777) was reserved for held-out evaluation. Performance was assessed using discrimination, calibration, threshold-specific classification, decision-curve analysis, referral burden, and post-hoc SHAP, permutation-importance, and subgroup fairness analyses. Results On the held-out 2010 test cohort, the calibrated model achieved AUROC 0.702, PR-AUC 0.332, and Brier score 0.137. At the sensitivity ≥ 90% threshold, sensitivity was 0.922 and specificity 0.276, with a 76.0% referral rate (approximately six individuals screened per detected case). Decision-curve analysis showed positive net benefit across a clinically relevant threshold range. GGT and urate emerged as leading non-glycaemic biochemical candidate signals. Subgroup evaluation showed similar sensitivity by sex, with modest differences in false-positive and referral rates, but substantially greater age- and ethnicity-specific heterogeneity. Conclusion A calibrated glycaemic-blind LightGBM framework delivers stable discrimination and good calibration for prediabetes prescreening. External subgroup validation, prospective equity assessment, and implementation evaluation are required before clinical use.