Machine-learning-based Screening for Diabetes Using Population-level Sociodemographic, Behavioral, and Anthropometric Factors
Akmaral Baspakova, Sanimgul S. Sambayeva, Kulyash R. Zhilisbayeva, Roza Suleimenova, Gulden Yelgondina, Akmeiir E. Kaliyeva, Aigerim A. Umbetova
A
BSTRACT
Introduction:
Scalable population-based screening tools are needed to identify individuals at high risk of undiagnosed or existing diabetes. Machine-learning approaches can capture complex nonlinear relationships among routinely collected risk factors and may improve screening-oriented risk stratification.
Methods:
We developed a supervised machine-learning model to predict diabetes status using a population-based cross-sectional dataset from West Kazakhstan derived from the WHO STEPwise survey (2021–2022). The analytical dataset included 4800 adults aged 18–69 years. Diabetes was defined according to standardized clinical and self-reported criteria, yielding a prevalence of 15.4% (739/4800). Predictor variables included sociodemographic, behavioral, and anthropometric characteristics, including age, sex, education, marital status, employment status, smoking, alcohol consumption, physical inactivity, high salt intake, body mass index (BMI), and waist circumference. A class-weighted Random Forest classifier was trained using stratified data splitting to preserve the original class distribution. Model performance was primarily assessed using precision–recall analysis, together with receiver operating characteristic–area under curve, calibration analysis, comparisons with baseline models, and repeated cross-validation. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).
Results:
The model generated probability-based estimates of diabetes status for screening-oriented risk stratification. Despite the class imbalance and relatively low prevalence of diabetes, the model achieved an average precision of approximately 0.31, indicating meaningful discriminative performance. SHAP analysis identified age, BMI, and waist circumference as the strongest contributors to the predicted probability of diabetes. The model was implemented as an interactive web-based tool providing real-time probability estimates for screening and educational purposes rather than diagnostic decision-making.
Conclusion:
This explainable Random Forest model, based on routinely collected population-level risk factors, may support screening-oriented diabetes risk stratification in community settings. Its implementation as an interactive web-based tool enables transparent, real-time risk estimation for preventive and educational purposes. Further external validation and probability recalibration are warranted before clinical implementation requiring reliable absolute risk estimates.