Performance of diabetes risk prediction models: a systematic review and meta-analysis

Type 2 diabetes mellitus (T2DM) poses a significant global public health burden, where early detection of at-risk populations is imperative for implementing targeted preventive strategies. This systematic review and meta-analysis aimed to evaluate the methodological quality and predictive performance of existing T2DM risk prediction models in screening contexts.

Following the TRIPOD-SRMA statement, eligible studies were selected through searching seven databases (CNKI, WanFang Database, VIP, PubMed, Embase, Web of Science, and the Cochrane Library) from database inception through December 2024. Methodological quality was assessed using the PROBAST tool. Random-effects models synthesized discrimination (AUC). Subgroup analyses explored geographic, modeling, and validation-related heterogeneity. Funnel plots and Egger's regression test assessed small-study effects.

A total of 65 studies (encompassing 97 distinct prediction models) were included in the analysis. Among 97 models, logistic regression dominated (97.9% of models), achieving moderate discrimination (AUC: 0.628-0.916), while machine learning (ML) models showed marginally higher AUCs (up to 0.998). Geographic and cohort disparities emerged, with USA-based models outperforming others (USA AUC: 0.97 vs China AUC: 0.79) and poor performance in prediabetic cohorts (AUC: 0.72 vs 0.80 in normoglycemic). External validation remained limited (21 models), though spatial/temporal validation cohorts demonstrated stable performance. High risk of bias and application (>80% of models) stemmed from inadequate statistical reporting and external verification definitions.

ML has favorable diagnostic accuracy for the progression of T2DM. This provides evidence for the development of predictive tools with broader applicability. Future research should prioritize external validation to enhance precision.