Psychological and Cardiometabolic Factors for Screening Current Type 2 Diabetes Status Using Machine Learning in the Kerman Cohort Study (KERCADRS)

Abstract

Background: The incremental contribution of psychological measures to population-based diabetes screening remains uncertain. Objectives: To classify current glycemic status as diabetes versus normal/prediabetes and to evaluate whether psychological measures provide incremental screening information beyond demographic and cardiometabolic predictors. Methods: This pooled, record-level cross-sectional analysis used data from two phases of the Kerman Coronary Artery Disease Risk Factors Study (KERCADRS). The analytic dataset comprised 15,839 records. Diabetes was defined as fasting blood glucose (FBS) ≥ 126 mg/dL or current use of antidiabetic medication; FBS and medication variables were excluded from the predictor set. Twenty clinically interpretable candidate predictors were considered, and Random Forest permutation importance was used within the fitting data to select 12 predictors. Logistic Regression, Random Forest, and HistGradientBoosting models were trained using a stratified, record-level development/test split. Within the development data, separate fitting, calibration, and threshold-selection subsets were used. Probabilities were calibrated using isotonic regression, screening thresholds were selected to prioritize sensitivity, and 95% confidence intervals were estimated using 1,000 bootstrap resamples of the hold-out, record-level test set. HistGradientBoosting was selected as the primary model because it achieved the highest hold-out test-set AUC and balanced accuracy. Its incremental performance with and without anxiety and depression was evaluated in an ablation analysis using the same development-data fitting, calibration, and threshold-selection procedure. Results: HistGradientBoosting demonstrated the highest overall discrimination (AUC, 0.800; 95% CI, 0.777 - 0.821) and balanced accuracy (0.736; 95% CI, 0.714 - 0.758). In the ablation analysis, the cardiometabolic model excluding anxiety and depression had an AUC of 0.801 and a balanced accuracy of 0.734. Paired bootstrap differences after adding psychological variables were −0.001 for AUC (95% CI, −0.008 to 0.006), 0.002 for balanced accuracy (95% CI, −0.010 to 0.013), and 0.000 for Brier score (95% CI, −0.001 to 0.001). Secondary coefficient-based associations are reported separately and were not interpreted as measures of feature importance for the selected model. Conclusions: HistGradientBoosting demonstrated the strongest overall performance in the hold-out, record-level test set for screening current diabetes status. Although anxiety was statistically associated with the screening outcome in the secondary Logistic Regression analysis, anxiety and depression together did not provide material incremental screening value beyond the selected cardiometabolic predictors. These models are intended solely as screening aids and do not replace confirmatory diagnostic testing.