Abstract
Background
Prediabetes and diabetes are global health risks, yet reliable non-invasive self-assessment screening systems for early detection are lacking, leaving many undiagnosed until interventions become less effective. We developed and validated an artificial intelligence (AI)-derived self-assessment system to predict prediabetes or diabetes risk using accessible health parameters.
Methods
30-years of National Health and Nutrition Examination Survey (NHANES) data (1988-2018) were used to develop AI models, with external validation in NHANES 2021-2023 and Korea NHANES 2023. The Boruta algorithm identified key predictive parameters. Six AI models were aggregated into an ensemble model, and the best-performing model was selected as the basis for the Machineborne Early Diabetic Warning And Control System (MEDWACS). Performance was assessed using the area under the receiver operating characteristic curve (ROCAUC), calibration plots, and supplementary metrics.
Results
Among 25,993 adults, 17,458 were used for derivation, while 3,043 US and 5,492 Korean for external validation. Seven key parameters were used to build the model: age, waist circumference, systolic blood pressure, gender, upper leg length, arm circumference, and body mass index. MEDWACS, a neural network-based system, achieved an ROCAUC of 0.804 (95% CI, 0.792-0.816) in internal validation and robust performance across subpopulations. External validation in the US and South Korea confirmed strong generalizability, with ROCAUCs of 0.773 (0.756-0.790) and 0.780 (0.768-0.792), respectively. The system showed robust calibration across validations. An online interface was developed to facilitate home and clinical use.
Conclusions
The AI-derived self-assessment system provides a non-invasive early warning tool for identifying prediabetes or diabetes risk, aiding individuals and physicians in timely interventions to reduce the public health burden.
Key messages
• MEDWACS, an AI-driven tool, enables non-invasive self-assessment of prediabetes or diabetes risk, supporting early detection and timely interventions.
• Our model predicts diabetes risk using simple health metrics, achieving strong validation across US and Korean populations.