Calibrated Risk Stratification Model for Impaired Fasting Glycemia in the Ugandan Population: A Nested Cross-Validation and Threshold Optimization Approach

Abstract Background The burden of type 2 diabetes is rapidly increasing in sub-Saharan Africa, and impaired fasting glycemia (IFG) is a reversible pre-disease state. In Uganda, the prevalence of hyperglycemia has doubled in the past decade, with more than 60% of cases undiagnosed, which points to an urgent need for cost-effective risk stratification tools. Methods We trained machine learning algorithms using secondary data from the nationally representative STEPS survey. We used case-cohort sampling: all IFG cases (n = 287) plus a 1:4 sub-cohort of controls (n = 1,053). Preprocessing included exclusion of participants with missing outcomes, multiple imputation (MICE, five imputations), and standardization of continuous features. Nine candidate models, including LASSO, random forest, and gradient boosting, were trained and evaluated under 5-fold nested cross-validation. Discrimination (ROC AUC), calibration, precision-recall AUC, and F1-score guided model selection. LASSO was selected as the final model for its balance of discrimination, calibration, and interpretability. Threshold moving used the G-mean to balance sensitivity and specificity. Out-of-fold predictions defined cutoffs for low-risk (NPV > = 0.90) and high-risk (PPV > = 0.40, Sn > = 0.70) strata, with an intermediate moderate-risk band. Results A total of 1,340 participants were included. The mean age was 39 years, with balanced representation across sex, regions, and residence types. The final calibrated LASSO model achieved an out-of-fold ROC AUC of 0.68 [95%CI: 0.64–0.71] and PR AUC of 0.40 [95%CI: 0.34–0.46], with Sn = 60%, Sp = 69%, PPV = 34% and NPV = 86%. Selected features included age, BMI, SBP, DBP, cholesterol, waist/hip circumference, heart rate, sedentary time, and categorical covariates (education, region, tobacco use, physical activity). Using the calibrated thresholds, 1% of participants were classified as low risk, 76% as moderate risk, and 23% as high risk. The high-risk band achieved a PPV of 0.41 and an IFG rate of 41%. Conclusions This calibrated, threshold-based risk stratification model provides a contextually relevant, non-invasive tool for identifying Ugandans at high risk of IFG. Integration into primary care could enable efficient allocation of confirmatory testing and preventive interventions. Further validation in prospective cohorts and implementation research are recommended to support scale-up.