ARTIFICIAL INTELLIGENCE IN PRE-DIABETES: CGM-DERIVED PHENOTYPES AND NON-INVASIVE BIOMARKERS FOR EARLY DETECTION AND RISK STRATIFICATION

Abstract

Prediabetes represents a critical window for preventing progression to type 2 diabetes mellitus (T2DM). Artificial intelligence (AI) offers novel opportunities for precision prevention through early detection, risk stratification, and individualized lifestyle interventions. We reviewed twenty-seven recent studies that explored applications of AI in prediabetes, including machine learning models leveraging electronic health records, deep learning approaches using wearable sensors, and non-invasive digital biomarkers. Emerging evidence strengthens the case for AI-enabled prediabetes management, with area under the receiver operating characteristic curve (AUROC) values ranging from 0.79 to 0.91 and root mean squared error (RMSE) below 15 mg/dL, surpassing conventional methods. AI-driven clustering reveals distinct metabolic phenotypes, highlighting the potential for tailored preventive strategies. Despite promising performances, challenges such as external validation, model generalizability, fairness, and clinical implementation remain. Current literature is also limited by homogeneous study designs and poor external validation, restricting opportunities for clinical translation. Our findings underscore the transformative potential of AI in prediabetes management by uniquely integrating available evidence, while emphasizing the need for rigorous translational studies to achieve scalable, data-driven, and personalized prevention at the population level.