International Journal of Clinical Case Reports, 2025, Vol.15, No.6, 293-302 http://medscipublisher.com/index.php/ijccr 294 This study will explore the application of artificial intelligence (AI) and machine learning (ML) in the prediction and early screening of diabetes. Ai-based models can utilize multi-source data such as electronic health records, clinical indicators, images, and wearable devices to construct screening tools with high sensitivity and specificity. It is used to identify complications such as type 1 and type 2 diabetes and diabetic retinopathy, and has shown good feasibility and acceptability in some clinical practices. At the same time, there are still challenges in terms of model universality, consistency among different data sources, and deep integration with clinical workflows. In view of this, this study will systematically analyze the accuracy of artificial intelligence prediction models in the early screening of diabetes, evaluate their advantages and disadvantages, and put forward suggestions for future research directions and clinical application prospects. 2 Early Screening of Diabetes and AI Prediction Model 2.1 Definitions of diabetes, prediabetes and target screening population Diabetes is a chronic metabolic disease caused by abnormal insulin secretion or action, resulting in persistent hyperglycemia. Among them, type 1 diabetes (T1D) is mainly an autoimmune disease, while type 2 diabetes (T2D) is mostly related to insulin resistance. Prediabetes refers to an intermediate state where blood glucose is higher than normal but has not yet reached the diagnostic criteria for diabetes. The risk of T2D and cardiovascular complications in this group of people is significantly increased (Davidson et al., 2021; Rodacki et al., 2025). The diagnosis of diabetes and prediabetes usually relies on indicators such as fasting plasma glucose (FPG), oral glucose tolerance test (OGTT), and glycated hemoglobin (HbA1c). Different international guidelines have slight differences in specific thresholds (Davidson et al., 2021; Rodacki et al., 2025). For whom diabetes screening should be targeted mainly depends on two types of situations: one is whether there is a risk of developing the disease, and the other is the significance of early intervention. For type 2 diabetes, it is generally recommended that adults aged 35-70 who are overweight or obese, as well as those who are younger but have a family history of diabetes, hypertension or dyslipidemia, etc. undergo screening (Davidson et al., 2021). Among children and adolescents, those who are overweight or obese, have a family history, or show insulin resistance are more recommended to be examined (Rodacki et al., 2025). Early detection of prediabetes or diabetes among these people can help prevent and reduce complications in a timely manner. However, traditional screening methods may still miss some high-risk groups, so more sensitive and easily accessible screening tools are needed. 2.2 Common types of AI prediction models The artificial intelligence (AI) prediction models used for diabetes screening mainly include various machine learning (ML) and deep learning (DL) algorithms. Commonly used ML models include logistic regression, decision tree, random forest, Support vector machine (SVM), and K-nearest Neighbor (KNN), which have good stability and interpretability when dealing with structured clinical data (Fregoso-Aparicio et al., 2021; Khanam and Foo, 2021; Mohsen et al., 2023; Dritsas and Trigka, 2022). Integrated methods such as random forest, AdaBoost and extreme gradient boosting improve prediction accuracy by combining multiple base models (Okwudili et al., 2025; Bontha et al., 2025). DL models such as deep neural networks (DNN) and convolutional neural networks (CNN) are more suitable for processing large-scale and complex data such as electronic health records, genetic information, and medical images, and can better capture nonlinear and high-dimensional features (Huang et al., 2023; Gong et al., 2025; Khokhar et al., 2025; Nie et al., 2025). Understandable artificial intelligence (XAI) technologies, such as SHapley Additive Interpretation (SHAP) and LIME, have been used to make models more transparent and doctors trust them more (Khokhar et al., 2025; Zhang et al., 2025). In terms of prediction accuracy and risk classification, multi-data models that combine clinical, genetic and imaging information usually perform better than those that use only one type of data (Mohsen et al., 2023; Huang et al., 2023). Such AI models are generally evaluated by criteria such as accuracy, area under the ROC curve (AUC, an indicator for assessing the model's discriminatory ability), sensitivity, specificity, and F1 value. Many studies have mentioned that the AUC of the model can exceed 0.80, indicating its strong ability to distinguish between illness and non-illness (Fregoso-Aparicio et al., 2021; Nie et al., 2025).
RkJQdWJsaXNoZXIy MjQ4ODYzNA==